{"title":"Graphics Cards","description":"\u003cp\u003eGraphics cards for gaming, content creation, workstation workloads and local AI, from mainstream GPUs to high-memory professional accelerators.\u003c\/p\u003e","products":[{"product_id":"tenstorrent-blackhole-p150a-32gb","title":"Tenstorrent Blackhole p150a (32GB)","description":"\u003cp\u003eThe Tenstorrent Blackhole p150a is Tenstorrent’s newest and most powerful single-card AI accelerator, an actively-cooled PCIe card built around the Blackhole architecture — the successor to Wormhole — and aimed at researchers, kernel engineers, and AI developers who want serious on-card compute and a fully open software stack in one desktop- or workstation-ready package. Where Wormhole proved the open-hardware approach works, Blackhole scales it up, packing more Tensix cores, more on-chip SRAM, and faster memory into a single-ASIC design.\u003c\/p\u003e\n\u003ch6\u003e\u003cstrong\u003eKey Features\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cstrong\u003e120 Tensix Cores\u003c\/strong\u003e — Tenstorrent’s core compute unit, each built around a cluster of small RISC-V processors rather than a traditional GPU shader pipeline, giving fine-grained programmability over how work is scheduled and executed.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e16 “big” RISC-V cores\u003c\/strong\u003e — dedicated general-purpose RISC-V processors on-die for control, orchestration, and running host-side logic close to the compute, reducing round-trips to the CPU.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e180MB of on-chip SRAM\u003c\/strong\u003e — roughly 1.5MB per Tensix core, a large local memory pool that keeps data close to the compute units and reduces pressure on off-chip DRAM bandwidth.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e32GB GDDR6 on a 256-bit bus\u003c\/strong\u003e — running at 16GT\/sec for 512GB\/s of memory bandwidth, enough headroom for meaningfully large local models and datasets.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eUp to 1.35GHz AI clock\u003c\/strong\u003e — delivering 664 TeraFLOPS at BLOCKFP8 precision, a substantial generational step up from Wormhole.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e4x QSFP-DD 800G ports\u003c\/strong\u003e — passive high-speed networking that lets multiple Blackhole cards link together to pool memory and scale compute across a multi-card system.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003ePCIe 5.0 x16 host interface\u003c\/strong\u003e — the latest-generation host interconnect for maximum data throughput between card and system.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eFully open-source software stack\u003c\/strong\u003e — the same open compiler and runtime approach (TT-Metalium, TT-NN) that defines Tenstorrent’s entire product line, with nothing hidden behind proprietary drivers.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch6\u003e\u003cstrong\u003eTenstorrent’s Most Capable Open AI Accelerator\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cp\u003eBlackhole represents a real architectural step forward from Wormhole, and the p150a is the card that showcases it best on a single-slot budget: more Tensix cores, nearly double the SRAM density per core, faster GDDR6 memory, and a move to PCIe 5.0 and 800G networking that positions it for genuinely serious multi-card scale-out. For AI research teams building or fine-tuning custom model architectures, kernel engineers who want to hand-optimize compute graphs at a level GPU vendors simply don’t expose, and RISC-V-focused developers who want general-purpose cores sitting directly alongside the AI compute fabric, the p150a is currently Tenstorrent’s flagship answer.\u003c\/p\u003e\n\u003cp\u003eThe 16 on-die “big” RISC-V cores are a distinctive feature worth calling out: they let orchestration, data preprocessing, and control-flow logic run directly on the card rather than round-tripping to the host CPU for every step, which matters for latency-sensitive or highly custom inference and training pipelines. Combined with 180MB of SRAM, the architecture is built to minimize the memory bottlenecks that constrain many accelerator designs, and the 800G QSFP-DD ports mean a multi-card Blackhole cluster can scale bandwidth well beyond what standard PCIe-only multi-GPU setups typically achieve.\u003c\/p\u003e\n\u003cp\u003eAs with every Tenstorrent product, the honest trade-off is software maturity versus openness. TT-Metalium and TT-NN are genuinely open, actively developed, and give you full visibility into how your workload maps to silicon — but the ecosystem of pre-optimized model support is smaller than CUDA’s or even ROCm’s, and getting the most out of the hardware generally means engaging directly with Tenstorrent’s programming model rather than expecting instant, zero-effort parity with a mainstream GPU framework. For teams who value that openness and are willing to invest in the stack, the p150a offers real, differentiated compute — and one of the only actively-cooled, single-card AI accelerators on the market where you can read and modify every layer of the software that runs on it.\u003c\/p\u003e\n\u003ch4\u003e\u003cstrong\u003eSpecifications\u003c\/strong\u003e\u003c\/h4\u003e\n\u003cul\u003e\n\u003cli\u003eTensix Cores: 120\u003c\/li\u003e\n\u003cli\u003eRISC-V Cores: 16 (“big” RISC-V, on-die)\u003c\/li\u003e\n\u003cli\u003eOn-Chip SRAM: 180MB (~1.5MB per Tensix core)\u003c\/li\u003e\n\u003cli\u003eMemory: 32GB GDDR6, 256-bit bus @ 16GT\/sec\u003c\/li\u003e\n\u003cli\u003eMemory Bandwidth: 512GB\/s\u003c\/li\u003e\n\u003cli\u003eAI Clock: Up to 1.35GHz\u003c\/li\u003e\n\u003cli\u003eCompute Performance: 664 TeraFLOPS (BLOCKFP8)\u003c\/li\u003e\n\u003cli\u003ePower: 300W total board power, 1x 12+4-pin (12V-2×6) connector\u003c\/li\u003e\n\u003cli\u003eInterface: PCIe 5.0 x16\u003c\/li\u003e\n\u003cli\u003eNetworking: 4x QSFP-DD 800G (passive) for multi-card scale-out\u003c\/li\u003e\n\u003cli\u003eDimensions: 42mm x 270mm x 111mm, actively cooled\u003c\/li\u003e\n\u003cli\u003eSoftware: Fully open-source stack (TT-Metalium, TT-NN)\u003c\/li\u003e\n\u003c\/ul\u003e","brand":"Tenstorrent","offers":[{"title":"Default Title","offer_id":66189487145309,"sku":"TT-BLACKHOLE-P150A","price":1056.53,"currency_code":"GBP","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/1080\/9321\/0973\/files\/ai-blackhole-p150.png?v=1791752014"},{"product_id":"tenstorrent-wormhole-n300-24gb","title":"Tenstorrent Wormhole n300 (24GB)","description":"\u003cp\u003eThe Tenstorrent Wormhole n300 is not a GPU repurposed for AI — it’s a purpose-built AI accelerator, designed from the ground up by Jim Keller’s Tenstorrent around a fully open-source software stack. Each card carries a pair of Wormhole Tensix processors on a single PCIe board, giving researchers, AI engineers, and hardware enthusiasts a genuinely open alternative to closed-driver GPU stacks — full visibility into the compiler, the kernel implementations, and the hardware itself, with nothing hidden behind a proprietary black box.\u003c\/p\u003e\n\u003ch6\u003e\u003cstrong\u003eKey Features\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cstrong\u003eDual Wormhole Tensix Processors\u003c\/strong\u003e — two full ASICs on one card, totaling 128 Tensix cores (64 per processor), each core built around Tenstorrent’s signature array of small RISC-V control processors rather than a fixed-function shader pipeline.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e24GB GDDR6 memory\u003c\/strong\u003e — running at 12GT\/sec for 576GB\/s of memory bandwidth, enough to hold meaningfully large quantized models for local inference and experimentation.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e192MB of on-chip SRAM\u003c\/strong\u003e — 96MB per ASIC, a large, fast local memory pool that Tensix cores can lean on heavily to reduce off-chip memory pressure, a distinctive architectural choice versus traditional GPU cache hierarchies.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eFully open-source software stack\u003c\/strong\u003e — the entire compiler and runtime (TT-Metalium, TT-NN) is open source on GitHub, giving developers low-level access most GPU vendors never expose, and the ability to hand-tune kernels for their exact workload.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eNative multichip mesh networking\u003c\/strong\u003e — 2x QSFP-DD 400GbE ports plus 2x Warp 100 bridge connectors let you link multiple Wormhole cards into larger mesh configurations, scaling compute and effective memory across cards.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e1GHz AI clock\u003c\/strong\u003e — delivering up to 466 TFLOPS FP8, 262 TFLOPS BFP8, and 131 TFLOPS FP16 across the card.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003ePCIe 4.0 x16 host interface\u003c\/strong\u003e — a standard, drop-in interface for any modern workstation or server motherboard.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch6\u003e\u003cstrong\u003eA Genuinely Open Platform for AI Research\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cp\u003eThe Wormhole n300’s appeal is fundamentally different from a GPU’s. Where NVIDIA and AMD cards run AI workloads through layers of proprietary or semi-proprietary driver and compiler stack, Tenstorrent publishes essentially everything — the ISA documentation, the compiler source, the low-level kernel programming model — as open source. For researchers building novel model architectures, kernel engineers optimizing for specific operator patterns, or anyone who’s hit a wall with a closed compute stack that won’t let them see what’s actually happening on-silicon, that openness is the entire value proposition. It’s the same philosophy that made RISC-V attractive at the CPU level, applied to AI accelerator hardware — and Tensix cores are themselves built around arrays of small RISC-V processors rather than a traditional fixed-function GPU pipeline.\u003c\/p\u003e\n\u003cp\u003ePractically, the n300 is well suited to local LLM inference on quantized open-weight models, to multichip AI research where the mesh-networking capability lets you scale beyond a single card’s memory, and to teams who specifically want to develop and profile custom kernels rather than rely entirely on vendor-tuned libraries. It is not a drop-in replacement for a mature CUDA or ROCm stack running mainstream, off-the-shelf inference frameworks — Tenstorrent’s ecosystem, while genuinely capable and rapidly improving, has a smaller library of pre-optimized model support and a steeper learning curve than the incumbent GPU vendors. Buyers should go in expecting to engage with the software stack directly, not simply pip-install a framework and expect instant parity with an NVIDIA card. For the audience it’s built for — open-hardware researchers, RISC-V enthusiasts, and teams who want to build their own optimized inference and training pipelines from first principles — the Wormhole n300 offers a rare combination: real silicon, real memory bandwidth, and nothing hidden.\u003c\/p\u003e\n\u003ch4\u003e\u003cstrong\u003eSpecifications\u003c\/strong\u003e\u003c\/h4\u003e\n\u003cul\u003e\n\u003cli\u003eProcessors: 2x Wormhole Tensix Processor (128 Tensix cores total, 64 per ASIC)\u003c\/li\u003e\n\u003cli\u003eOn-Chip SRAM: 192MB (96MB per ASIC)\u003c\/li\u003e\n\u003cli\u003eMemory: 24GB GDDR6 @ 12GT\/sec\u003c\/li\u003e\n\u003cli\u003eMemory Bandwidth: 576GB\/s\u003c\/li\u003e\n\u003cli\u003eAI Clock: 1GHz\u003c\/li\u003e\n\u003cli\u003eCompute Performance: 466 TFLOPS FP8 \/ 262 TFLOPS BFP8 \/ 131 TFLOPS FP16\u003c\/li\u003e\n\u003cli\u003ePower: Up to 300W total board power\u003c\/li\u003e\n\u003cli\u003eInterface: PCIe 4.0 x16\u003c\/li\u003e\n\u003cli\u003eNetworking: 2x QSFP-DD 400GbE + 2x Warp 100 bridge for multichip mesh scaling\u003c\/li\u003e\n\u003cli\u003eForm Factor: Actively cooled PCIe card\u003c\/li\u003e\n\u003cli\u003eSoftware: Fully open-source stack (TT-Metalium, TT-NN)\u003c\/li\u003e\n\u003c\/ul\u003e","brand":"Tenstorrent","offers":[{"title":"Default Title","offer_id":66189487178077,"sku":"TT-WORMHOLE-N300","price":1056.53,"currency_code":"GBP","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/1080\/9321\/0973\/files\/ai-wormhole-n300-scaled.png?v=1791752014"},{"product_id":"intel-arc-pro-b60-24gb","title":"Intel Arc Pro B60 (24GB)","description":"\u003cp\u003eThe Intel Arc Pro B60 steps up from the B50 to 24GB of memory on the same Xe2 “Battlemage” architecture, aimed squarely at AI inference, generative design, and 3D workloads where memory capacity matters as much as raw compute. Intel does not manufacture a reference-design B60 — every card ships from a board partner (ASRock, Sparkle, and others), built to Intel’s reference specification, so this listing is a partner card faithful to that spec rather than an Intel-branded reference board.\u003c\/p\u003e\n\u003ch6\u003e\u003cstrong\u003eKey Features\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cstrong\u003e24GB GDDR6 on a 192-bit bus\u003c\/strong\u003e — at up to 456GB\/s of bandwidth, well ahead of most cards in its price class and enough to hold meaningfully large quantized AI models or complex 3D scenes entirely in VRAM.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e20 Xe2 Cores, 160 XMX Engines\u003c\/strong\u003e — a step up from the B50’s core count, with 20 dedicated ray tracing units for professional visualization work.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e197 TOPS INT8 AI performance\u003c\/strong\u003e — strong dedicated matrix acceleration for AI inference workloads, purpose-built for the low-precision math that dominates modern LLM and vision-model inference.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e12.28 TFLOPS FP32 compute\u003c\/strong\u003e — solid general shader performance for CAD, rendering, and design applications.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eMulti-GPU ready\u003c\/strong\u003e — pairs of B60 cards can combine (Intel’s “Battlemage” dual-GPU and Project Battlematrix configurations) to pool up to 48GB of addressable memory for larger models.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003ePCIe 5.0 x8 interface\u003c\/strong\u003e — modern interface bandwidth with efficient, scalable power delivery.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e120-200W total board power\u003c\/strong\u003e — a wide, partner-dependent power envelope that lets board vendors tune for either efficiency or maximum clocks.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch6\u003e\u003cstrong\u003eSerious Memory Capacity for AI Inference and 3D Work\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cp\u003eThe B60’s headline feature is memory: 24GB is a genuinely large pool for a card in this price bracket, and it changes what’s realistically possible locally. Quantized LLMs in the 13B–34B range, larger diffusion and vision models, and dense 3D scenes with high-resolution textures all fit comfortably, where a 12GB or 16GB card would force compromises on batch size, context length, or scene complexity. For AI developers and small studios evaluating local inference hardware, that memory headroom — paired with 197 TOPS of dedicated INT8 throughput — makes the B60 a genuinely competitive option against similarly priced cards from AMD and NVIDIA, and the multi-GPU pooling option gives a clear, affordable upgrade path to 48GB for teams that outgrow a single card.\u003c\/p\u003e\n\u003cp\u003eFor 3D and generative design workloads, the combination of 24GB, 20 dedicated ray tracing units, and Intel’s Xe2 rendering pipeline supports professional visualization and CAD applications that need to hold large assemblies or high-poly scenes in memory without swapping. Intel has been actively expanding ISV certifications and driver validation for the Arc Pro line, closing the gap with the more established professional driver programs from AMD and NVIDIA.\u003c\/p\u003e\n\u003cp\u003eThe honest caveat is the same one that applies across Intel’s discrete GPU push: this is a newer software ecosystem, and while Intel’s oneAPI, OpenVINO, and driver stack have matured quickly and now support the mainstream AI and CAD frameworks most teams rely on, it doesn’t yet have the sheer breadth of tooling and community support that CUDA has built over more than a decade. For teams using standard, well-supported frameworks — llama.cpp, PyTorch with Intel extensions, OpenVINO-based inference, mainstream CAD applications — the B60 delivers strong, genuinely competitive performance per pound. For bleeding-edge research code that assumes NVIDIA hardware, expect more setup work. Given that every B60 ships from a board partner rather than an Intel reference design, exact clocks, cooling, and connector layout vary slightly by vendor, but all cards adhere to Intel’s core specification below.\u003c\/p\u003e\n\u003ch4\u003e\u003cstrong\u003eSpecifications\u003c\/strong\u003e\u003c\/h4\u003e\n\u003cul\u003e\n\u003cli\u003eArchitecture: Intel Xe2 “Battlemage,” TSMC N5\u003c\/li\u003e\n\u003cli\u003eXe2 Cores: 20 (160 XMX engines, 20 ray tracing units)\u003c\/li\u003e\n\u003cli\u003eMemory: 24GB GDDR6, 192-bit bus\u003c\/li\u003e\n\u003cli\u003eMemory Bandwidth: Up to 456GB\/s\u003c\/li\u003e\n\u003cli\u003eCompute Performance: 12.28 TFLOPS FP32 \/ up to 197 TOPS INT8\u003c\/li\u003e\n\u003cli\u003eTotal Board Power: 120-200W (partner-dependent)\u003c\/li\u003e\n\u003cli\u003eInterface: PCIe 5.0 x8\u003c\/li\u003e\n\u003cli\u003eMulti-GPU: Pairable for up to 48GB pooled memory\u003c\/li\u003e\n\u003cli\u003eAvailability: Board-partner card built to Intel reference specification (no Intel-branded reference card exists)\u003c\/li\u003e\n\u003c\/ul\u003e","brand":"Intel","offers":[{"title":"Default Title","offer_id":66189487276381,"sku":"INTEL-ARCB60","price":453.11,"currency_code":"GBP","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/1080\/9321\/0973\/files\/gpu-arc-b60.png?v=1791752015"},{"product_id":"intel-arc-pro-b50-16gb","title":"Intel Arc Pro B50 (16GB)","description":"\u003cp\u003eThe Intel Arc Pro B50 is proof that a workstation GPU doesn’t need a huge power budget or a huge price tag to be genuinely useful. Built on Intel’s Xe2 “Battlemage” architecture, this compact, low-profile card packs 16GB of memory and dedicated AI acceleration into a 70W design that needs no external power connector at all — making it an easy, low-risk upgrade for small-form-factor workstations, entry-level AI inference boxes, and multi-display CAD or design setups where space and power budget are tight.\u003c\/p\u003e\n\u003ch6\u003e\u003cstrong\u003eKey Features\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cstrong\u003eIntel Xe2 “Battlemage” architecture\u003c\/strong\u003e — built on TSMC’s N5 process, Intel’s second-generation discrete GPU architecture with dedicated ray tracing and matrix (XMX) engines for AI workloads.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e16GB GDDR6\u003c\/strong\u003e — more memory than many competing entry workstation cards at this price and power point, useful headroom for CAD models, multi-layer design work, and small-to-mid AI models.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e16 Xe2 Cores, 128 XMX Engines\u003c\/strong\u003e — dedicated AI acceleration matrix hardware alongside 16 ray tracing units, delivering up to 170 TOPS of INT8 performance for AI inference workloads.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e10.65 TFLOPS FP32 compute\u003c\/strong\u003e — solid general-purpose shader performance for CAD, design, and light rendering work.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eJust 70W total board power\u003c\/strong\u003e — no external power connector needed on most builds, drawing power entirely from the PCIe slot.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eLow-profile, dual-slot, single-fan design\u003c\/strong\u003e — fits compact and small-form-factor workstations where larger workstation cards simply won’t physically fit.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003ePCIe 5.0 x8 interface\u003c\/strong\u003e — a modern interface standard at a fraction of the power draw of higher-end workstation cards.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e224GB\/s memory bandwidth\u003c\/strong\u003e — solid throughput for a card in this power and price class.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch6\u003e\u003cstrong\u003eThe Easy Entry Point Into Professional GPU Compute\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cp\u003eThe Arc Pro B50 targets a segment that’s been underserved for years: professionals who need a certified, reliable workstation GPU but don’t need — or can’t power and cool — a 200W-plus card. Small-form-factor workstations, dense multi-seat deployments, and thin-and-light mobile workstation docks all benefit from a card that draws its entire power budget from the PCIe slot and still fits a low-profile, dual-slot bracket. For CAD, 2D\/3D design, video editing, and light content creation, 16GB of memory and Intel’s Xe2 ray tracing hardware provide a genuinely capable professional experience, not a stripped-down compromise.\u003c\/p\u003e\n\u003cp\u003eFor AI, the 128 XMX engines put the B50 in an interesting position: it’s not built for training or heavy fine-tuning, but for lightweight local inference — running smaller open-weight LLMs, image classification models, or vision pipelines at the edge — it offers real dedicated matrix acceleration in a card that costs and draws a fraction of what a flagship AI GPU does. For businesses deploying entry-level AI inference across many desks or edge devices, the B50’s low power draw and no-external-power-connector design also simplify procurement and system design considerably: virtually any modern PSU can power one, and OEM systems can spec it without redesigning power delivery.\u003c\/p\u003e\n\u003cp\u003eIt’s worth being clear-eyed about where the B50 sits in the market: this is an entry-level professional card, not a high-end rendering or AI training GPU, and Intel’s software ecosystem (drivers, ISV certifications, oneAPI tooling) is younger than NVIDIA’s or AMD’s — though it has matured quickly and Intel actively certifies against major professional applications. For buyers whose workloads fit within its power envelope — compact CAD workstations, multi-display professional setups, and light local AI inference — the B50 offers genuine professional-grade capability at a price and power draw nothing else in Intel’s, AMD’s, or NVIDIA’s current lineups can match.\u003c\/p\u003e\n\u003cp\u003eIt’s also a sensible fleet-upgrade card: because it draws all its power from the PCIe slot, IT teams refreshing rows of compact workstations don’t need to audit or upgrade power supplies system by system, which meaningfully lowers the total cost and time of a rollout compared to swapping in higher-wattage cards that demand new PSUs or additional power connectors.\u003c\/p\u003e\n\u003ch4\u003e\u003cstrong\u003eSpecifications\u003c\/strong\u003e\u003c\/h4\u003e\n\u003cul\u003e\n\u003cli\u003eArchitecture: Intel Xe2 “Battlemage,” TSMC N5\u003c\/li\u003e\n\u003cli\u003eXe2 Cores: 16 (128 XMX engines, 16 ray tracing units)\u003c\/li\u003e\n\u003cli\u003eMemory: 16GB GDDR6, 128-bit bus\u003c\/li\u003e\n\u003cli\u003eMemory Bandwidth: 224GB\/s\u003c\/li\u003e\n\u003cli\u003eCompute Performance: 10.65 TFLOPS FP32 \/ up to 170 TOPS INT8\u003c\/li\u003e\n\u003cli\u003eClocks: 1,700MHz base \/ 2,600MHz boost\u003c\/li\u003e\n\u003cli\u003eTotal Board Power: 70W (no external power connector)\u003c\/li\u003e\n\u003cli\u003eInterface: PCIe 5.0 x8\u003c\/li\u003e\n\u003cli\u003eForm Factor: Low-profile, dual-slot, single-fan\u003c\/li\u003e\n\u003c\/ul\u003e","brand":"Intel","offers":[{"title":"Default Title","offer_id":66189487341917,"sku":"INTEL-ARCB50","price":225.81,"currency_code":"GBP","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/1080\/9321\/0973\/files\/gpu-arc-b50.png?v=1791752016"},{"product_id":"amd-radeon-ai-pro-r9700-32gb","title":"AMD Radeon AI PRO R9700 (32GB)","description":"\u003cp\u003eThe AMD Radeon AI PRO R9700 is AMD’s first RDNA 4 workstation card, purpose-built for the wave of teams who want serious local AI capability without stepping up to a 48GB flagship card or a datacenter GPU. With 32GB of memory, a 2-slot blower design, and a genuinely competitive price for its class, the R9700 is aimed squarely at AI developers, data scientists, and technical creators who need local inference and fine-tuning headroom on a workstation budget.\u003c\/p\u003e\n\u003ch6\u003e\u003cstrong\u003eKey Features\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cstrong\u003eRDNA 4 architecture (Navi 48)\u003c\/strong\u003e — AMD’s newest GPU architecture, built on TSMC’s 4nm process, bringing a generational leap in AI-specific hardware over the previous RDNA 3 workstation cards.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e32GB GDDR6 on a 256-bit bus\u003c\/strong\u003e — running at 20Gbps for 640GB\/s of bandwidth, enough to comfortably run quantized LLMs in the 30B+ parameter range or fine-tune smaller models entirely on-card.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e64 Compute Units, 4,096 Stream Processors\u003c\/strong\u003e — plus 64 second-generation Ray Accelerators for professional rendering and visualization work.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e128 second-generation AI Accelerators\u003c\/strong\u003e — up to 95.7 TFLOPS of dense FP16 compute and up to 1,531 TOPS of INT4 performance with sparsity, purpose-built for modern quantized inference workloads.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eMulti-GPU scaling\u003c\/strong\u003e — pool memory and compute across multiple R9700 cards for larger models than a single card can hold, at a fraction of the cost of scaling with flagship workstation GPUs.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e2-slot blower cooling\u003c\/strong\u003e — a genuinely compact, workstation-friendly form factor that supports dense multi-GPU builds without the space and airflow demands of larger cards.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eQuad DisplayPort 2.1a outputs\u003c\/strong\u003e — full multi-monitor support for engineering and creative workstation use alongside AI workloads.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eROCm on Linux and Windows\u003c\/strong\u003e — an open software stack for AI development, with official support on both major desktop operating systems.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch6\u003e\u003cstrong\u003ePurpose-Built for Local AI Development\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cp\u003eThe R9700 occupies a genuinely useful niche: more memory and AI-specific hardware than any consumer gaming GPU at a similar price, without the cost jump to a 48GB flagship workstation card. For AI developers running local inference on open-weight models, fine-tuning smaller architectures, or building and testing RAG pipelines before deploying to cloud infrastructure, 32GB is a sweet spot — enough for meaningfully large quantized models, without the premium of the very largest memory pools. The 128 second-generation AI Accelerators are specifically tuned for the low-precision matrix math (INT4, INT8, FP8) that dominates modern LLM inference, which is where AMD has focused RDNA 4’s biggest architectural gains over RDNA 3.\u003c\/p\u003e\n\u003cp\u003eFor teams building multi-GPU AI workstations, the R9700’s compact 2-slot blower design and multi-GPU scaling support mean you can realistically fit several cards in one chassis to pool memory for larger models — a build strategy that’s considerably more affordable per gigabyte of usable VRAM than a single flagship card, provided your framework and workload can take advantage of multi-GPU sharding. Creative and technical professionals also benefit from the same hardware for GPU-accelerated rendering and data-heavy visualization work, not just pure AI inference.\u003c\/p\u003e\n\u003cp\u003eAs with any AMD professional card, it’s worth setting expectations on software: ROCm has matured significantly and is officially supported on this card across both Linux and Windows, but framework compatibility and out-of-the-box performance still generally trail CUDA on the newest, most experimental AI research code. For teams running mainstream inference stacks (llama.cpp, vLLM, PyTorch with ROCm backends) rather than bleeding-edge research repos, the R9700 delivers strong practical performance at a price point that undercuts comparable-memory NVIDIA workstation cards substantially.\u003c\/p\u003e\n\u003cp\u003eFor a workstation builder evaluating options, the R9700 is best thought of as the entry point into serious local AI hardware: it sits above consumer gaming GPUs in both memory and dedicated AI throughput, while staying meaningfully cheaper per gigabyte than the 48GB-and-up professional tier. That makes it a sensible first purchase for a team just standing up local inference infrastructure, and an easy card to scale out once the workload and framework choices are proven.\u003c\/p\u003e\n\u003ch4\u003e\u003cstrong\u003eSpecifications\u003c\/strong\u003e\u003c\/h4\u003e\n\u003cul\u003e\n\u003cli\u003eArchitecture: AMD RDNA 4 (Navi 48), TSMC 4nm\u003c\/li\u003e\n\u003cli\u003eCompute Units: 64 (4,096 Stream Processors)\u003c\/li\u003e\n\u003cli\u003eRay Accelerators: 64 (2nd-gen) \/ AI Accelerators: 128 (2nd-gen)\u003c\/li\u003e\n\u003cli\u003eMemory: 32GB GDDR6, 256-bit bus, 20Gbps\u003c\/li\u003e\n\u003cli\u003eMemory Bandwidth: 640GB\/s\u003c\/li\u003e\n\u003cli\u003eCompute Performance: 47.8 TFLOPS FP32 \/ 95.7 TFLOPS FP16 \/ up to 1,531 TOPS INT4 (sparse)\u003c\/li\u003e\n\u003cli\u003eClocks: 1,620MHz base \/ 2,920MHz boost\u003c\/li\u003e\n\u003cli\u003eTotal Board Power: 300W\u003c\/li\u003e\n\u003cli\u003ePower Connector: 1x 12V-2×6\u003c\/li\u003e\n\u003cli\u003eInterface: PCIe 5.0\u003c\/li\u003e\n\u003cli\u003eForm Factor: 2-slot blower\u003c\/li\u003e\n\u003cli\u003eDisplay Outputs: 4x DisplayPort 2.1a\u003c\/li\u003e\n\u003cli\u003eLaunch: July 2025, MSRP $1,299\u003c\/li\u003e\n\u003c\/ul\u003e","brand":"AMD","offers":[{"title":"Default Title","offer_id":66189487407453,"sku":"AMD-R9700","price":981.01,"currency_code":"GBP","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/1080\/9321\/0973\/files\/gpu-r9700.png?v=1791752016"},{"product_id":"amd-radeon-pro-w7900-48gb","title":"AMD Radeon PRO W7900 (48GB)","description":"\u003cp\u003eThe AMD Radeon PRO W7900 is AMD’s flagship workstation graphics card, and its defining feature is simple: 48GB of ECC-protected memory on the full RDNA 3 Navi 31 die, in a dual-slot form factor that fits real workstation chassis. For studios and engineering teams running large AI models, professional rendering pipelines, or scientific visualization workloads that outgrow a 24GB consumer card, the W7900 offers datacenter-class capacity without datacenter pricing or power draw.\u003c\/p\u003e\n\u003ch6\u003e\u003cstrong\u003eKey Features\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cstrong\u003e48GB ECC GDDR6\u003c\/strong\u003e — the full Navi 31 memory configuration on a 384-bit bus, enough to run 30B-parameter class LLMs, hold entire production scenes, or process massive point-cloud and CAD datasets entirely on-card, with error-correcting code protection for professional reliability.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e96 Compute Units, 6,144 Stream Processors\u003c\/strong\u003e — the full Navi 31 die, unlocked for workstation workloads rather than gaming clocks.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e96 second-generation Ray Accelerators, 192 AI Accelerators\u003c\/strong\u003e — dedicated hardware for professional ray-traced rendering and AI-assisted workflows.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e864GB\/s memory bandwidth\u003c\/strong\u003e — keeps the compute units fed when working with the massive datasets this card’s memory capacity enables.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e61 TFLOPS FP32 \/ 122 TFLOPS FP16\u003c\/strong\u003e — strong compute for rendering, simulation, and mixed-precision AI workloads.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eDual-slot workstation design\u003c\/strong\u003e — a considerably more space-efficient footprint than earlier triple-slot professional cards, letting integrators fit up to four W7900s in a single multi-GPU workstation.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eROCm certified\u003c\/strong\u003e — validated for professional AI and ML development, with ISV-certified drivers for major CAD, DCC, and simulation applications.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eQuad DisplayPort 2.1 outputs\u003c\/strong\u003e — three full-size DisplayPort 2.1 plus one mini-DisplayPort 2.1, each capable of 54Gbps for high-resolution multi-monitor CAD and visualization setups.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch6\u003e\u003cstrong\u003eLarge-Memory Compute for AI, Rendering, and Simulation\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cp\u003eThe case for the W7900 is almost entirely about memory headroom. At 48GB, it sits well above every mainstream consumer GPU and matches or beats many competing professional cards at a lower price point, which matters enormously for AI teams working with LLMs and diffusion models that simply won’t fit — or won’t fit comfortably alongside a large batch size — in 24GB. Fine-tuning mid-size open-weight models, running multiple models resident simultaneously for a production inference pipeline, or working with uncompressed 3D scan and point-cloud data all become dramatically easier when the dataset doesn’t have to be chunked or offloaded to system memory.\u003c\/p\u003e\n\u003cp\u003eFor rendering and visualization professionals, the same memory advantage applies directly: complex architectural and product visualization scenes, full-resolution VFX shots, and simulation caches can be held entirely in GPU memory, which keeps interactive viewport performance smooth even on the heaviest projects. AMD’s ISV certification program means the W7900 is validated against the professional applications studios actually rely on, not just consumer benchmarks.\u003c\/p\u003e\n\u003cp\u003eWhere honesty matters most is software maturity: ROCm has made real strides and is officially supported and certified for this card, but the CUDA ecosystem still has a head start in raw framework coverage, especially for bleeding-edge AI research code that assumes NVIDIA hardware by default. For teams already standardized on PyTorch and mainstream inference frameworks, ROCm on the W7900 is a genuinely capable, well-supported path — and one that comes with substantially more memory per pound than the equivalent CUDA-based workstation card. For teams whose workflows are rendering- and CAD-first rather than bleeding-edge AI research, the W7900’s ISV certifications and driver stability make it a low-risk, high-capacity upgrade.\u003c\/p\u003e\n\u003cp\u003eIt’s also a genuinely cost-effective alternative to scaling out with multiple smaller GPUs just to reach a usable memory footprint: one W7900 replaces what would otherwise take two or more 24GB cards, without the added system complexity, driver overhead, or PCIe lane contention of a multi-GPU build. For engineering firms and studios standardizing their workstation fleet, that simplicity — a single ECC-protected, ISV-certified card per seat — is often worth as much as the raw compute numbers on the spec sheet.\u003c\/p\u003e\n\u003ch4\u003e\u003cstrong\u003eSpecifications\u003c\/strong\u003e\u003c\/h4\u003e\n\u003cul\u003e\n\u003cli\u003eArchitecture: AMD RDNA 3 (full Navi 31 die)\u003c\/li\u003e\n\u003cli\u003eCompute Units: 96 (6,144 Stream Processors, 192 AI Accelerators, 96 Ray Accelerators)\u003c\/li\u003e\n\u003cli\u003eMemory: 48GB ECC GDDR6, 384-bit bus\u003c\/li\u003e\n\u003cli\u003eMemory Bandwidth: 864GB\/s\u003c\/li\u003e\n\u003cli\u003eCompute Performance: 61 TFLOPS FP32 \/ 122 TFLOPS FP16\u003c\/li\u003e\n\u003cli\u003eClocks: 1,855MHz core \/ 2,495MHz boost\u003c\/li\u003e\n\u003cli\u003eTotal Board Power: 295W\u003c\/li\u003e\n\u003cli\u003ePower Connectors: 2x 8-pin PCIe\u003c\/li\u003e\n\u003cli\u003eInterface: PCIe 4.0 x16\u003c\/li\u003e\n\u003cli\u003eForm Factor: Dual-slot workstation card, 280mm length\u003c\/li\u003e\n\u003cli\u003eDisplay Outputs: 3x DisplayPort 2.1 + 1x mini-DisplayPort 2.1\u003c\/li\u003e\n\u003cli\u003eSoftware: ROCm certified, ISV-certified professional drivers\u003c\/li\u003e\n\u003c\/ul\u003e","brand":"AMD","offers":[{"title":"Default Title","offer_id":66189487505757,"sku":"AMD-W7900","price":3020.05,"currency_code":"GBP","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/1080\/9321\/0973\/files\/gpu-w7900.png?v=1791752017"},{"product_id":"amd-radeon-rx-7900-xtx-24gb","title":"AMD Radeon RX 7900 XTX (24GB)","description":"\u003cp\u003eThe AMD Radeon RX 7900 XTX is AMD’s flagship RDNA 3 gaming card, and it’s still one of the strongest value propositions in high-end graphics — the first consumer GPU to use an advanced chiplet design, giving it a huge 24GB memory pool and near-flagship performance at a price well below the equivalent NVIDIA tier. For gamers chasing high-refresh 4K, streamers who need serious encode headroom, and Linux users who want a fully open driver stack, the 7900 XTX earns its place at the top of AMD’s lineup.\u003c\/p\u003e\n\u003ch6\u003e\u003cstrong\u003eKey Features\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cstrong\u003eChiplet-based Navi 31 GPU\u003c\/strong\u003e — a 5nm graphics compute die paired with six 6nm memory cache dies (MCDs), the first gaming GPU to split compute and memory logic across separate chiplets for better yields and cost efficiency.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e96 Compute Units, 6,144 Stream Processors\u003c\/strong\u003e — high-throughput parallel compute for rendering, ray tracing, and general GPU workloads.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e24GB GDDR6 on a 384-bit bus\u003c\/strong\u003e — more memory than NVIDIA’s equivalent-tier competition, at up to 20Gbps for 960GB\/s of raw bandwidth.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e96MB AMD Infinity Cache\u003c\/strong\u003e — a large on-die cache that lifts effective memory bandwidth up to 3,500GB\/s, keeping the GPU fed at 4K and beyond without a wider (and costlier) memory bus.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e96 second-generation Ray Accelerators and 192 AI Accelerators\u003c\/strong\u003e — dedicated hardware for ray tracing and AI-assisted upscaling (FSR) workloads.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eUp to 2,500MHz boost clock\u003c\/strong\u003e (2,300MHz game clock) — high sustained clocks for consistent frame rates under load.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e61.4 TFLOPS FP32\u003c\/strong\u003e — strong raw compute for both gaming and GPU-accelerated creative applications.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eROCm support on Linux\u003c\/strong\u003e — an open-source path into PyTorch-based AI\/ML development, something no NVIDIA gaming card offers without proprietary CUDA tooling.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch6\u003e\u003cstrong\u003eFlagship 4K Gaming Without Flagship Pricing\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cp\u003eAt native 4K, the RX 7900 XTX trades blows with NVIDIA’s previous-generation flagship class while typically undercutting it meaningfully on price — for buyers who prioritize raw rasterization performance and memory capacity over the very best ray tracing or AI upscaling quality, that’s a compelling trade. 24GB of VRAM also future-proofs the card against the growing texture and asset sizes of modern AAA titles far better than 16GB competitors, and it means the 7900 XTX can double as a genuinely capable local AI card for image generation and smaller LLM inference — something a 12GB or 16GB card simply can’t do as comfortably.\u003c\/p\u003e\n\u003cp\u003eContent creators benefit from AMD’s dual encode engines and strong OpenCL\/HIP compute performance in supported applications, while streamers get enough encode headroom to broadcast at high quality without tanking in-game frame rates. It’s worth being candid about where AMD’s ecosystem is behind: DLSS and ray tracing quality still generally lead FSR and AMD’s ray accelerators in the most demanding path-traced titles, and ROCm — while genuinely open and improving fast — doesn’t yet have the plug-and-play maturity of CUDA for AI\/ML work, particularly on Windows. Support is strongest on Linux, where ROCm is officially validated for this GPU.\u003c\/p\u003e\n\u003cp\u003eFor buyers who want the most VRAM and raw performance per pound in AMD’s lineup, are comfortable with FSR instead of DLSS, and want the option to dabble in open-source AI development, the RX 7900 XTX remains a genuinely excellent flagship-tier card — chiplet engineering that delivers real value rather than just a manufacturing curiosity.\u003c\/p\u003e\n\u003cp\u003eIt’s also worth weighing this card against upgrading from an older mid-range or previous-generation flagship: the jump in compute units, memory bandwidth, and VRAM capacity is substantial enough that games running at 1440p today have real headroom to move up to native 4K, and titles that felt VRAM-constrained on an 8GB or 12GB card simply won’t be an issue here. Combined with a driver suite (Adrenalin) that’s matured considerably in recent years and now offers strong feature parity for streaming, recording, and performance tuning, the RX 7900 XTX is a well-rounded flagship purchase rather than a card that only wins on a single spec sheet line.\u003c\/p\u003e\n\u003ch4\u003e\u003cstrong\u003eSpecifications\u003c\/strong\u003e\u003c\/h4\u003e\n\u003cul\u003e\n\u003cli\u003eArchitecture: AMD RDNA 3 (Navi 31), chiplet design — 5nm GCD + 6nm MCDs\u003c\/li\u003e\n\u003cli\u003eCompute Units: 96 (6,144 Stream Processors)\u003c\/li\u003e\n\u003cli\u003eRay Accelerators: 96 (2nd-generation) \/ AI Accelerators: 192\u003c\/li\u003e\n\u003cli\u003eMemory: 24GB GDDR6, 384-bit bus, 20Gbps\u003c\/li\u003e\n\u003cli\u003eMemory Bandwidth: 960GB\/s (up to 3,500GB\/s effective with Infinity Cache)\u003c\/li\u003e\n\u003cli\u003eInfinity Cache: 96MB\u003c\/li\u003e\n\u003cli\u003eFP32 Compute: 61.4 TFLOPS\u003c\/li\u003e\n\u003cli\u003eClocks: 2,300MHz game \/ up to 2,500MHz boost\u003c\/li\u003e\n\u003cli\u003eTotal Board Power: 355W\u003c\/li\u003e\n\u003cli\u003ePower Connectors: 2x 8-pin PCIe (reference design)\u003c\/li\u003e\n\u003cli\u003eInterface: PCIe 4.0 x16\u003c\/li\u003e\n\u003cli\u003eDisplay Outputs: DisplayPort 2.1, HDMI 2.1, USB-C (reference design)\u003c\/li\u003e\n\u003cli\u003eLaunch: December 2022, original MSRP $999\u003c\/li\u003e\n\u003c\/ul\u003e","brand":"AMD","offers":[{"title":"Default Title","offer_id":66189487604061,"sku":"AMD-RX7900XTX","price":678.93,"currency_code":"GBP","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/1080\/9321\/0973\/files\/gpu-rx-7900xtx.jpg?v=1791752018"},{"product_id":"nvidia-geforce-rtx-4090-founders-edition-24gb","title":"NVIDIA GeForce RTX 4090 Founders Edition (24GB)","description":"\u003cp\u003eThe NVIDIA GeForce RTX 4090 Founders Edition remains one of the most capable GPUs you can buy — the Ada Lovelace flagship that redefined what a single-card gaming and creative rig could do, and a card that still holds its own against everything short of NVIDIA’s newest Blackwell generation. For gamers who want maxed-out 4K performance, creators pushing 8K video and huge 3D scenes, and enthusiasts running local AI models, the RTX 4090 continues to deliver flagship-class results at a price that has become considerably more accessible.\u003c\/p\u003e\n\u003ch6\u003e\u003cstrong\u003eKey Features\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cstrong\u003e16,384 CUDA Cores\u003c\/strong\u003e — massive parallel compute built on a custom TSMC 4N process, still enough to lead every non-Blackwell GPU on the market in raw shader throughput.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e24GB GDDR6X on a 384-bit bus\u003c\/strong\u003e — running at 21Gbps for over 1TB\/s of memory bandwidth, comfortably handling 4K\/8K textures, large open-world scenes, and local AI models up to the mid-teens in billions of parameters.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e72MB L2 cache\u003c\/strong\u003e — a large on-die cache that reduces memory-bound stalls in both games and compute workloads.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e512 fourth-generation Tensor Cores\u003c\/strong\u003e — up to 1,321 AI TOPS, driving DLSS 3 Frame Generation and strong local inference performance for Stable Diffusion, LLMs, and other generative AI workloads.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e128 third-generation RT Cores\u003c\/strong\u003e — up to 191 TFLOPS of ray tracing performance for fully path-traced titles like Cyberpunk 2077 and Alan Wake 2.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e82.6 TFLOPS FP32 shader compute\u003c\/strong\u003e — a substantial generational gain over the previous RTX 30-series flagship, still highly competitive today.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e2.52GHz boost clock\u003c\/strong\u003e — high out-of-the-box clocks with strong overclocking headroom on partner and Founders Edition boards alike.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eAV1 hardware encode and decode\u003c\/strong\u003e — efficient, high-quality hardware video encoding that cuts export times and improves streaming quality versus older codecs.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch6\u003e\u003cstrong\u003eStill a Flagship for Gaming, Creation, and Local AI\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cp\u003eTwo-plus years after launch, the RTX 4090 remains the card of choice for anyone who wants uncompromising 4K gaming without stepping up to Blackwell pricing. At native 4K with ray tracing enabled, it comfortably delivers high, stable frame rates in the most demanding current titles, and DLSS 3 Frame Generation extends that lead further in supported games. For high-refresh 1440p gaming it has more headroom than most displays can even use, which means this card ages gracefully as games get heavier rather than becoming a bottleneck within a generation.\u003c\/p\u003e\n\u003cp\u003eFor content creators, the 24GB frame buffer and 1TB\/s+ of bandwidth make the RTX 4090 a genuine production tool rather than just a gaming card: 8K video editing, dense Blender and Unreal Engine scenes, and multi-layer compositing all fit and run smoothly, while the NVENC AV1 encoder significantly speeds up exports and live streaming versus CPU-based encoding or older hardware encoders. For local AI experimentation, 24GB is enough to comfortably run Stable Diffusion and similar image models at full speed, and to run inference on quantized LLMs in the 13B–34B parameter range without exotic offloading tricks — a genuinely useful entry point into local generative AI that doesn’t require a dedicated workstation GPU.\u003c\/p\u003e\n\u003cp\u003eThe Founders Edition specifically brings NVIDIA’s own dual-axial-flow cooler and vapor chamber design — a compact, well-engineered thermal solution that keeps the 450W TDP in check without the bulk of some triple-slot partner cards, while still leaving headroom for the transient power spikes (up to 600W) that Ada Lovelace’s flagship is known for. Pair it with an 850W-or-better power supply and a case with decent airflow, and you get a card that, even now, only NVIDIA’s newest Blackwell flagships can outrun.\u003c\/p\u003e\n\u003ch4\u003e\u003cstrong\u003eSpecifications\u003c\/strong\u003e\u003c\/h4\u003e\n\u003cul\u003e\n\u003cli\u003eArchitecture: NVIDIA Ada Lovelace, TSMC 4N\u003c\/li\u003e\n\u003cli\u003eCUDA Cores: 16,384\u003c\/li\u003e\n\u003cli\u003eTensor Cores: 512 (4th-generation)\u003c\/li\u003e\n\u003cli\u003eRT Cores: 128 (3rd-generation)\u003c\/li\u003e\n\u003cli\u003eMemory: 24GB GDDR6X, 384-bit bus, 21Gbps, 72MB L2 cache\u003c\/li\u003e\n\u003cli\u003eMemory Bandwidth: ~1,008GB\/s\u003c\/li\u003e\n\u003cli\u003eFP32 Compute: 82.6 TFLOPS\u003c\/li\u003e\n\u003cli\u003eAI Performance: Up to 1,321 AI TOPS\u003c\/li\u003e\n\u003cli\u003eBoost Clock: 2.52GHz\u003c\/li\u003e\n\u003cli\u003eTotal Graphics Power: 450W (transient spikes up to 600W)\u003c\/li\u003e\n\u003cli\u003eInterface: PCIe 4.0 x16\u003c\/li\u003e\n\u003cli\u003eDisplay Outputs: DisplayPort 1.4a, HDMI 2.1a\u003c\/li\u003e\n\u003cli\u003eRecommended PSU: 850W\u003c\/li\u003e\n\u003cli\u003eLaunch: October 2022, original MSRP $1,599\u003c\/li\u003e\n\u003c\/ul\u003e","brand":"NVIDIA","offers":[{"title":"Default Title","offer_id":66189487669597,"sku":"NVIDIA-RTX4090-FE","price":1207.57,"currency_code":"GBP","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/1080\/9321\/0973\/files\/gpu-rtx-4090.jpg?v=1791752019"},{"product_id":"nvidia-geforce-rtx-5090-founders-edition-32gb","title":"NVIDIA GeForce RTX 5090 Founders Edition (32GB)","description":"\u003cp\u003eThe NVIDIA GeForce RTX 5090 Founders Edition is the card for gamers and creators who refuse to compromise. It is NVIDIA’s flagship Blackwell-generation consumer GPU — the largest GeForce die NVIDIA has ever shipped, wrapped in a Founders Edition cooler engineered to fit that much silicon into two slots. Whether you’re chasing native 4K at maximum settings, editing 8K footage, or running local diffusion models and LLMs on your own hardware, the RTX 5090 is built to be the last GPU upgrade you need for years.\u003c\/p\u003e\n\u003ch6\u003e\u003cstrong\u003eKey Features\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cstrong\u003e21,760 CUDA Cores\u003c\/strong\u003e — the largest GeForce GPU configuration NVIDIA has ever produced, built on a custom TSMC 4NP process for a substantial generational leap over the RTX 4090.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e32GB GDDR7 on a 512-bit bus\u003c\/strong\u003e — massive memory capacity and bandwidth for 4K\/8K textures, huge open-world scenes, and local AI models that would choke a 24GB card.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e98MB L2 cache\u003c\/strong\u003e — a significantly expanded cache hierarchy that keeps the compute cores fed and reduces trips to memory.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e680 fifth-generation Tensor Cores\u003c\/strong\u003e — up to 3,352 AI TOPS, powering DLSS 4.5 with Multi Frame Generation for dramatically higher frame rates at native-quality image fidelity.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e170 fourth-generation RT Cores\u003c\/strong\u003e — full-scene path tracing in supported titles without the frame-rate penalty older RT hardware imposed.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e104.8 TFLOPS FP32 shader compute\u003c\/strong\u003e — a genuine generational jump in raw rasterization and compute throughput over Ada Lovelace.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e2.41GHz boost clock\u003c\/strong\u003e — high clocks despite the enormous core count, thanks to a refined 4NP process node.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eCompact 2-slot Founders Edition design\u003c\/strong\u003e — flagship performance in a chassis-friendly footprint that fits SFF and standard ATX builds alike, unlike the triple-slot coolers common on partner cards this powerful.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch6\u003e\u003cstrong\u003eFlagship Gaming, Creation, and Local AI\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cp\u003eAt native 4K with ray tracing maxed out, the RTX 5090 is currently the fastest consumer GPU on the market, and it’s not close — the jump in CUDA core count over the RTX 4090 alone (21,760 vs 16,384) translates directly into higher sustained frame rates in the most demanding modern titles, before DLSS 4.5 and Multi Frame Generation are even factored in. For competitive players running high-refresh 1440p or 4K monitors, the 5090 has the headroom to keep frame rates pinned near the display’s ceiling even with settings maxed, something narrower-margin cards can’t guarantee as games get heavier year over year.\u003c\/p\u003e\n\u003cp\u003eCreators get an equally strong case: 32GB of GDDR7 means 8K video timelines, multi-layer compositing in DaVinci Resolve or Premiere, and large 3D scenes in Blender or Unreal Engine all fit comfortably in VRAM, and NVIDIA’s NVENC\/NVDEC encoders (backed by AV1 support) cut export and streaming times significantly versus CPU-only workflows. For anyone experimenting with local AI — running Stable Diffusion, fine-tuning smaller LLMs, or doing inference on quantized 30B-class models — 32GB is the difference between a model that fits and one that constantly spills to system RAM. It won’t match a dedicated 96GB workstation card for the largest models, but for local generative AI experimentation alongside gaming and creative work, no other consumer GPU offers this much memory and compute in one card.\u003c\/p\u003e\n\u003cp\u003eThe Founders Edition specifically brings NVIDIA’s own vapor-chamber cooling and flow-through design, tested and validated as a matched pair with the card’s power delivery — a meaningfully different (and typically quieter, more consistent) thermal solution than many third-party coolers on a 575W flagship. Total graphics power lands at 575W, so a genuinely capable 1000W PSU and a case with good airflow are non-negotiable, but in return you get the single most capable GPU NVIDIA has ever put in a GeForce box.\u003c\/p\u003e\n\u003ch4\u003e\u003cstrong\u003eSpecifications\u003c\/strong\u003e\u003c\/h4\u003e\n\u003cul\u003e\n\u003cli\u003eArchitecture: NVIDIA Blackwell, TSMC 4NP\u003c\/li\u003e\n\u003cli\u003eCUDA Cores: 21,760\u003c\/li\u003e\n\u003cli\u003eTensor Cores: 680 (5th-generation)\u003c\/li\u003e\n\u003cli\u003eRT Cores: 170 (4th-generation)\u003c\/li\u003e\n\u003cli\u003eMemory: 32GB GDDR7, 512-bit bus, 98MB L2 cache\u003c\/li\u003e\n\u003cli\u003eFP32 Compute: ~104.8 TFLOPS\u003c\/li\u003e\n\u003cli\u003eAI Performance: Up to 3,352 AI TOPS\u003c\/li\u003e\n\u003cli\u003eBoost Clock: 2.41GHz\u003c\/li\u003e\n\u003cli\u003eTotal Graphics Power: 575W\u003c\/li\u003e\n\u003cli\u003ePower Connector: 1x 16-pin (12V-2×6)\u003c\/li\u003e\n\u003cli\u003eInterface: PCIe 5.0\u003c\/li\u003e\n\u003cli\u003eDisplay Outputs: DisplayPort 2.1b, up to 8K @ 165Hz\u003c\/li\u003e\n\u003cli\u003eRecommended PSU: 1000W\u003c\/li\u003e\n\u003cli\u003eLaunch: January 2025, MSRP $1,999\u003c\/li\u003e\n\u003c\/ul\u003e","brand":"NVIDIA","offers":[{"title":"Default Title","offer_id":66189487767901,"sku":"NVIDIA-RTX5090-FE","price":1509.65,"currency_code":"GBP","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/1080\/9321\/0973\/files\/gpu-rtx-5090.jpg?v=1791752020"},{"product_id":"nvidia-rtx-pro-6000-blackwell-workstation-edition-96gb","title":"NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96GB)","description":"\u003cp\u003eWhen a single professional GPU has to carry an entire AI pipeline, a CAD department, or a VFX render queue, the NVIDIA RTX PRO 6000 Blackwell Workstation Edition is the card built for that job. It is NVIDIA’s flagship professional GPU, pairing the full Blackwell GB202 die with a class-leading 96GB of ECC-protected GDDR7 — enough memory to hold entire large language models, massive point clouds, or feature-film-resolution scenes without offloading to system RAM. For studios, labs, and engineering teams that need workstation-class reliability and datacenter-class capacity in a single desktop tower, this is the card to build around.\u003c\/p\u003e\n\u003ch6\u003e\u003cstrong\u003eKey Features\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cul\u003e\n\u003cli\u003e\n\u003cstrong\u003eFull Blackwell GB202 die\u003c\/strong\u003e — 24,064 CUDA cores across 188 streaming multiprocessors, the largest professional Blackwell configuration NVIDIA ships, built on TSMC’s 4NP process for maximum throughput per watt.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e96GB GDDR7 with ECC\u003c\/strong\u003e — the largest memory pool of any NVIDIA workstation GPU, on a 512-bit bus, with error-correcting code protection for mission-critical simulation and inference workloads where a single flipped bit is unacceptable.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e1,792GB\/s memory bandwidth\u003c\/strong\u003e — feeds the compute cores fast enough for real-time AI inference, complex fluid and structural simulation, and multi-layer 8K compositing without bottlenecking.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e752 fifth-generation Tensor Cores\u003c\/strong\u003e — deliver up to 4,000 AI TOPS (FP4 with structured sparsity) and native FP4 precision support, a Blackwell-exclusive capability that dramatically accelerates modern LLM inference and fine-tuning.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e188 fourth-generation RT Cores\u003c\/strong\u003e — hardware-accelerated ray and path tracing for production rendering, architectural visualization, and real-time look development.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e125 TFLOPS FP32\u003c\/strong\u003e — strong general compute performance for CAD, simulation, and scientific workloads that lean on single-precision math.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003e600W configurable TDP\u003c\/strong\u003e — a single 16-pin PCIe CEM5 power connector, with power tunable by system integrators to balance thermals, acoustics, and density in workstation or server chassis.\u003c\/li\u003e\n\u003cli\u003e\n\u003cstrong\u003eQuad DisplayPort 2.1b outputs\u003c\/strong\u003e — drive up to four high-resolution displays (including 8K panels) from a single card, ideal for multi-monitor CAD and content-creation setups.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch6\u003e\u003cstrong\u003eBuilt for AI, Rendering, and Simulation at Scale\u003c\/strong\u003e\u003c\/h6\u003e\n\u003cp\u003eThe RTX PRO 6000 Blackwell Workstation Edition exists to remove the memory ceiling that forces AI teams into multi-GPU setups or cloud rentals just to fit a model. With 96GB on a single card, you can run 70B-parameter LLMs locally in reduced precision, fine-tune mid-size models without gradient checkpointing gymnastics, or hold multiple diffusion and vision models resident at once for production pipelines. The native FP4 support in its fifth-generation Tensor Cores is a genuine architectural advance over the previous RTX 6000 Ada generation, not just a clock-speed bump — it means inference throughput on quantized modern LLMs scales well beyond what the raw CUDA core count alone would suggest.\u003c\/p\u003e\n\u003cp\u003eFor rendering and visualization teams, the combination of 188 RT cores, 96GB of frame buffer, and CUDA’s mature ecosystem means V-Ray, Arnold, Octane, Redshift, and OptiX-based renderers can hold entire film-resolution scenes — geometry, textures, and simulation caches — in GPU memory simultaneously, cutting out the disk-streaming stalls that plague lower-memory cards on the heaviest shots. Engineering and scientific users get the same benefit for CFD, FEA, and digital-twin simulation, where dataset size, not raw FLOPS, is usually the limiting factor.\u003c\/p\u003e\n\u003cp\u003eThis is a professional card, not a gaming GPU wearing a workstation badge — it ships with NVIDIA’s Studio and RTX Enterprise driver branches, is validated against major CAD, DCC, and AI frameworks, and carries the ECC memory protection that regulated and mission-critical workloads require. Compared against running the equivalent workload across multiple consumer GPUs, the RTX PRO 6000 Blackwell simplifies the software stack considerably: no manual model sharding, no NVLink bridge management, just one card with more addressable memory than most workstations’ entire RAM. At 600W and dual-slot, it also drops into standard workstation chassis without the multi-card cooling and power headaches of a datacenter GPU build. If your bottleneck is VRAM — whether that’s a language model that won’t fit, a scene that keeps swapping to disk, or a simulation mesh you’ve had to decimate just to make it run — this is the card that removes that ceiling in one upgrade.\u003c\/p\u003e\n\u003ch4\u003e\u003cstrong\u003eSpecifications\u003c\/strong\u003e\u003c\/h4\u003e\n\u003cul\u003e\n\u003cli\u003eArchitecture: NVIDIA Blackwell (GB202, full die), TSMC 4NP\u003c\/li\u003e\n\u003cli\u003eCUDA Cores: 24,064\u003c\/li\u003e\n\u003cli\u003eTensor Cores: 752 (5th-generation, native FP4)\u003c\/li\u003e\n\u003cli\u003eRT Cores: 188 (4th-generation)\u003c\/li\u003e\n\u003cli\u003eMemory: 96GB GDDR7 with ECC, 512-bit bus\u003c\/li\u003e\n\u003cli\u003eMemory Bandwidth: 1,792GB\/s\u003c\/li\u003e\n\u003cli\u003eFP32 Compute: 125 TFLOPS\u003c\/li\u003e\n\u003cli\u003eAI Performance: Up to 4,000 AI TOPS (FP4 sparse)\u003c\/li\u003e\n\u003cli\u003ePower: 600W configurable TDP, 1x 16-pin PCIe CEM5 connector\u003c\/li\u003e\n\u003cli\u003eInterface: PCIe 5.0 x16\u003c\/li\u003e\n\u003cli\u003eDisplay Outputs: 4x DisplayPort 2.1b\u003c\/li\u003e\n\u003cli\u003eForm Factor: Dual-slot, workstation blower\u003c\/li\u003e\n\u003cli\u003eLaunch: March 2025, MSRP $8,565\u003c\/li\u003e\n\u003c\/ul\u003e","brand":"NVIDIA","offers":[{"title":"Default Title","offer_id":66189487833437,"sku":"NVIDIA-RTXPRO6000-96GB","price":10006.41,"currency_code":"GBP","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/1080\/9321\/0973\/files\/gpu-rtx-pro-6000.jpg?v=1791752021"}],"url":"https:\/\/saleforgery.com\/collections\/graphics-cards.oembed","provider":"SaleForge","version":"1.0","type":"link"}