{"product_id":"carte-graphique-pro-nvidia-a100-pcie-80go-pny-copy","title":"NVIDIA H100 NVL 94GB Pro Graphics Card","description":"\u003cp data-start=\"0\" data-end=\"507\"\u003eThe \u003cstrong data-start=\"3\" data-end=\"28\"\u003eNVIDIA H100 NVL 94 GB\u003c\/strong\u003e is a \u003cstrong data-start=\"36\" data-end=\"67\"\u003edatacenter GPU accelerator\u003c\/strong\u003e (\"Hopper\" class) designed for the most demanding workloads, including \u003cstrong data-start=\"153\" data-end=\"189\"\u003egenerative AI \/ LLM inference\u003c\/strong\u003e, \u003cstrong data-start=\"194\" data-end=\"201\"\u003eHPC\u003c\/strong\u003e, and large-scale analytics. The \u003cstrong data-start=\"247\" data-end=\"254\"\u003eNVL\u003c\/strong\u003e version is designed to operate \u003cstrong data-start=\"283\" data-end=\"297\"\u003ein pairs\u003c\/strong\u003e of GPUs connected by \u003cstrong data-start=\"317\" data-end=\"327\"\u003eNVLink\u003c\/strong\u003e, in order to increase usable memory capacity and inter-GPU throughput, while remaining in a \u003cstrong data-start=\"430\" data-end=\"438\"\u003ePCIe\u003c\/strong\u003e format suitable for many servers. \u003cspan class=\"\" data-state=\"closed\"\u003e\u003c\/span\u003e\u003c\/p\u003e\n\u003ch2 data-start=\"509\" data-end=\"523\"\u003eKey Points\u003c\/h2\u003e\n\u003cul data-start=\"524\" data-end=\"1055\"\u003e\n\n\u003cli data-start=\"524\" data-end=\"704\"\u003e\n\n\u003cp data-start=\"526\" data-end=\"704\"\u003e\u003cstrong data-start=\"526\" data-end=\"557\"\u003eVery large memory per GPU\u003c\/strong\u003e and massive \u003cstrong data-start=\"561\" data-end=\"587\"\u003ememory bandwidth\u003c\/strong\u003e, ideal for large models and low-latency serving. \u003cspan class=\"\" data-state=\"closed\"\u003e\u003c\/span\u003e\u003c\/p\u003e\n\n\n\u003c\/li\u003e\n\n\u003cli data-start=\"705\" data-end=\"897\"\u003e\n\n\u003cp data-start=\"707\" data-end=\"897\"\u003e\u003cstrong data-start=\"707\" data-end=\"741\"\u003eNVLink between two cards (NVL)\u003c\/strong\u003e: allows for creating a pair with a \u003cstrong data-start=\"780\" data-end=\"804\"\u003etotal of 188 GB HBM3\u003c\/strong\u003e (94 GB + 94 GB) and a high-speed link between the GPUs. \u003cspan class=\"\" data-state=\"closed\"\u003e\u003c\/span\u003e\u003c\/p\u003e\n\n\n\u003c\/li\u003e\n\n\u003cli data-start=\"898\" data-end=\"1055\"\u003e\n\n\u003cp data-start=\"900\" data-end=\"1055\"\u003eDesigned for \u003cstrong data-start=\"920\" data-end=\"936\"\u003e24\/7 server\u003c\/strong\u003e operation (passive cooling depending on integration) and the NVIDIA ecosystem for AI\/HPC. \u003cspan class=\"\" data-state=\"closed\"\u003e\u003c\/span\u003e\u003c\/p\u003e\n\n\n\u003c\/li\u003e\n\n\n\u003c\/ul\u003e\n\u003ch2 data-start=\"1057\" data-end=\"1089\"\u003eMain Technical Details\u003c\/h2\u003e\n\u003cul data-start=\"1090\" data-end=\"2041\"\u003e\n\n\u003cli data-start=\"1090\" data-end=\"1175\"\u003e\n\n\u003cp data-start=\"1092\" data-end=\"1175\"\u003e\u003cstrong data-start=\"1092\" data-end=\"1108\"\u003eArchitecture\u003c\/strong\u003e: NVIDIA \u003cstrong data-start=\"1118\" data-end=\"1135\"\u003eHopper (H100)\u003c\/strong\u003e \u003cspan class=\"\" data-state=\"closed\"\u003e\u003c\/span\u003e\u003c\/p\u003e\n\n\n\u003c\/li\u003e\n\n\u003cli data-start=\"1176\" data-end=\"1304\"\u003e\n\n\u003cp data-start=\"1178\" data-end=\"1304\"\u003e\u003cstrong data-start=\"1178\" data-end=\"1189\"\u003eMemory\u003c\/strong\u003e: \u003cstrong data-start=\"1192\" data-end=\"1206\"\u003e94 GB HBM3\u003c\/strong\u003e per GPU (often used in a \u003cstrong data-start=\"1242\" data-end=\"1252\"\u003e188 GB\u003c\/strong\u003e pair via NVLink) \u003cspan class=\"\" data-state=\"closed\"\u003e\u003c\/span\u003e\u003c\/p\u003e\n\n\n\u003c\/li\u003e\n\n\u003cli data-start=\"1305\" data-end=\"1404\"\u003e\n\n\u003cp data-start=\"1307\" data-end=\"1404\"\u003e\u003cstrong data-start=\"1307\" data-end=\"1333\"\u003eMemory Bandwidth\u003c\/strong\u003e: up to \u003cstrong data-start=\"1344\" data-end=\"1356\"\u003e3.9 TB\/s\u003c\/strong\u003e per GPU \u003cspan class=\"\" data-state=\"closed\"\u003e\u003c\/span\u003e\u003c\/p\u003e\n\n\n\u003c\/li\u003e\n\n\u003cli data-start=\"1405\" data-end=\"1480\"\u003e\n\n\u003cp data-start=\"1407\" data-end=\"1480\"\u003e\u003cstrong data-start=\"1407\" data-end=\"1420\"\u003eInterface\u003c\/strong\u003e: \u003cstrong data-start=\"1423\" data-end=\"1440\"\u003ePCIe Gen5 x16\u003c\/strong\u003e \u003cspan class=\"\" data-state=\"closed\"\u003e\u003c\/span\u003e\u003c\/p\u003e\n\n\n\u003c\/li\u003e\n\n\u003cli data-start=\"1481\" data-end=\"1603\"\u003e\n\n\u003cp data-start=\"1483\" data-end=\"1603\"\u003e\u003cstrong data-start=\"1483\" data-end=\"1501\"\u003eInterconnect\u003c\/strong\u003e: \u003cstrong data-start=\"1504\" data-end=\"1514\"\u003eNVLink\u003c\/strong\u003e (NVL pair, high-speed link between two H100 NVL) \u003cspan class=\"\" data-state=\"closed\"\u003e\u003c\/span\u003e\u003c\/p\u003e\n\n\n\u003c\/li\u003e\n\n\u003cli data-start=\"1604\" data-end=\"1724\"\u003e\n\n\u003cp data-start=\"1606\" data-end=\"1724\"\u003e\u003cstrong data-start=\"1606\" data-end=\"1629\"\u003eTDP (max power)\u003c\/strong\u003e: up to \u003cstrong data-start=\"1640\" data-end=\"1649\"\u003e400 W\u003c\/strong\u003e per GPU (according to NVL specifications) \u003cspan class=\"\" data-state=\"closed\"\u003e\u003c\/span\u003e\u003c\/p\u003e\n\n\n\u003c\/li\u003e\n\n\u003cli data-start=\"1725\" data-end=\"1918\"\u003e\n\n\u003cp data-start=\"1727\" data-end=\"1918\"\u003e\u003cstrong data-start=\"1727\" data-end=\"1782\"\u003eAI Acceleration (Tensor Cores + Transformer Engine)\u003c\/strong\u003e: optimized for formats and computations used by LLMs (FP8\/FP16\/BF16\/TF32 depending on modes) \u003cspan class=\"\" data-state=\"closed\"\u003e\u003c\/span\u003e\u003c\/p\u003e\n\n\n\u003c\/li\u003e\n\n\u003cli data-start=\"1919\" data-end=\"2041\"\u003e\n\n\u003cp data-start=\"1921\" data-end=\"2041\"\u003e\u003cstrong data-start=\"1921\" data-end=\"1933\"\u003eDecoding\u003c\/strong\u003e: mention of \u003cstrong data-start=\"1947\" data-end=\"1959\"\u003e7× NVDEC\u003c\/strong\u003e (and \u003cstrong data-start=\"1964\" data-end=\"1975\"\u003e7× JPEG\u003c\/strong\u003e) on the H100 NVL specs \u003cspan class=\"\" data-state=\"closed\"\u003e\u003c\/span\u003e\u003c\/p\u003e\n\n\n\u003c\/li\u003e\n\n\n\u003c\/ul\u003e\n\u003ch2 data-start=\"2043\" data-end=\"2056\"\u003eWho is it for?\u003c\/h2\u003e\n\u003cul data-start=\"2057\" data-end=\"2501\"\u003e\n\n\u003cli data-start=\"2057\" data-end=\"2210\"\u003e\n\n\u003cp data-start=\"2059\" data-end=\"2210\"\u003e\u003cstrong data-start=\"2059\" data-end=\"2083\"\u003eAI \/ LLM (inference)\u003c\/strong\u003e: serving large models, large contexts, significant batches, controlled latency. \u003cspan class=\"\" data-state=\"closed\"\u003e\u003c\/span\u003e\u003c\/p\u003e\n\n\n\u003c\/li\u003e\n\n\u003cli data-start=\"2211\" data-end=\"2348\"\u003e\n\n\u003cp data-start=\"2213\" data-end=\"2348\"\u003e\u003cstrong data-start=\"2213\" data-end=\"2250\"\u003eHPC \/ simulation \/ data analytics\u003c\/strong\u003e: intensive computing and memory-bandwidth bound workloads. \u003cspan class=\"\" data-state=\"closed\"\u003e\u003c\/span\u003e\u003c\/p\u003e\n\n\n\u003c\/li\u003e\n\n\u003cli data-start=\"2349\" data-end=\"2501\"\u003e\n\n\u003cp data-start=\"2351\" data-end=\"2501\"\u003e\u003cstrong data-start=\"2351\" data-end=\"2379\"\u003eServer infrastructures\u003c\/strong\u003e: dense deployments where PCIe format and production efficiency matter.\u003c\/p\u003e\n\n\n\u003c\/li\u003e\n\n\n\u003c\/ul\u003e","brand":"Radiance Systems","offers":[{"title":"Default Title","offer_id":53398363308296,"sku":null,"price":38233.29,"currency_code":"EUR","in_stock":true}],"thumbnail_url":"\/\/cdn.shopify.com\/s\/files\/1\/0702\/4816\/9736\/files\/41FprRdH2RL._AC_SL1260_1024x_1a3d13ab-8f1b-4eee-bd9c-5fec0f8453f0.webp?v=1772210432","url":"https:\/\/www.radiancesystems.eu\/en\/products\/carte-graphique-pro-nvidia-a100-pcie-80go-pny-copy","provider":"Radiance Systems","version":"1.0","type":"link"}