{"product_id":"nvidia-cmp-170hx-64-gb-hbm2e-modified-ex-mining","title":"NVIDIA CMP 170HX 64 GB HBM2e (Modified, Ex-Mining)","description":"\u003cstyle\u003e\n.rtx{--bl:#334fb4;--dk:#26408f;--ink:#17171a;--soft:#f3f7fd;--muted:#6c6660;--line:#dce6f5;font-family:-apple-system,BlinkMacSystemFont,\"Segoe UI\",Roboto,Helvetica,Arial,sans-serif;color:var(--ink);line-height:1.6;font-size:15px;}\n.rtx *{box-sizing:border-box;}\n.rtx-head{background:var(--ink);color:#fff;border-radius:14px;padding:30px 28px;position:relative;overflow:hidden;}\n.rtx-head:before{content:\"\";position:absolute;inset:0;background:radial-gradient(120% 100% at 100% 0,rgba(51,79,180,.32),transparent 55%);}\n.rtx-head\u003e*{position:relative;}\n.rtx-head h2{font-size:clamp(22px,3.2vw,30px);font-weight:800;margin:0 0 10px;color:#fff !important;}\n.rtx-head p{color:#d7d2cb;max-width:64ch;margin:0;}\n.rtx-badge{display:inline-block;background:var(--bl);color:#fff !important;font-size:12px;font-weight:700;letter-spacing:.08em;text-transform:uppercase;padding:5px 11px;border-radius:999px;margin-bottom:14px;}\n.rtx-hi{display:grid;grid-template-columns:repeat(4,1fr);gap:1px;background:var(--line);border:1px solid var(--line);border-radius:12px;overflow:hidden;margin:18px 0 6px;}\n.rtx-hi div{background:#fff;padding:18px 12px;text-align:center;}\n.rtx-hi .v{font-size:20px;font-weight:800;}\n.rtx-hi .v small{color:var(--bl);font-size:12px;}\n.rtx-hi .k{font-size:11.5px;color:var(--muted);margin-top:4px;}\n.rtx-sec{padding:24px 0 4px;}\n.rtx-ey{font-size:12px;font-weight:700;letter-spacing:.16em;text-transform:uppercase;color:var(--dk);}\n.rtx h2.t{font-size:clamp(19px,2.4vw,24px);font-weight:800;margin:6px 0 12px;}\n.rtx p.lead{font-size:16px;color:#3a3a36;}\n.rtx table{width:100%;border-collapse:collapse;font-size:14px;margin:6px 0;}\n.rtx table.spec td{padding:9px 12px;border-bottom:1px solid var(--line);}\n.rtx table.spec tr:nth-child(odd) td{background:#f6f9fc;}\n.rtx table.spec td:first-child{width:42%;color:var(--muted);font-weight:600;}\n.rtx ul.tick{list-style:none;margin:6px 0;padding:0;}\n.rtx ul.tick li{position:relative;padding:7px 0 7px 26px;border-bottom:1px solid var(--line);font-size:14.5px;}\n.rtx ul.tick li:last-child{border-bottom:none;}\n.rtx ul.tick li:before{content:\"\";position:absolute;left:4px;top:14px;width:9px;height:9px;border-radius:2px;background:var(--bl);}\n.rtx-note{background:var(--soft);border-left:3px solid var(--bl);padding:13px 16px;border-radius:0 8px 8px 0;font-size:13.5px;color:#4a4641;margin-top:14px;}\n.rtx-faq details{border-bottom:1px solid var(--line);padding:14px 0;}\n.rtx-faq summary{font-weight:700;cursor:pointer;list-style:none;display:flex;justify-content:space-between;font-size:15px;}\n.rtx-faq summary:after{content:\"+\";color:var(--bl);font-weight:700;}\n.rtx-faq details[open] summary:after{content:\"–\";}\n.rtx-faq p{margin:9px 0 0;color:var(--muted);font-size:14px;}\n@media(max-width:760px){.rtx-hi{grid-template-columns:repeat(2,1fr);}}\n\u003c\/style\u003e\n\u003cdiv class=\"rtx\"\u003e\n\u003cdiv class=\"rtx-head\"\u003e\n\u003cspan class=\"rtx-badge\"\u003eModified · Ex-Mining · 64 GB HBM2e\u003c\/span\u003e\n\u003ch2\u003eNVIDIA CMP 170HX 64 GB HBM2e — cheap VRAM for resident inference\u003c\/h2\u003e\n\u003cp\u003eA GA100-based CMP 170HX with its memory expanded to 64 GB of HBM2e. At €1,600 that is roughly €25 per gigabyte of high-bandwidth VRAM — around a fifth of what large-VRAM professional cards cost. It is modified hardware with real limits, and we spell them out below before you buy.\u003c\/p\u003e\n\u003c\/div\u003e\n\u003cdiv class=\"rtx-hi\"\u003e\n\u003cdiv\u003e\n\u003cdiv class=\"v\"\u003e64 \u003csmall\u003eGB\u003c\/small\u003e\n\u003c\/div\u003e\n\u003cdiv class=\"k\"\u003eHBM2e VRAM\u003c\/div\u003e\n\u003c\/div\u003e\n\u003cdiv\u003e\n\u003cdiv class=\"v\"\u003eGA100\u003c\/div\u003e\n\u003cdiv class=\"k\"\u003eAmpere GPU\u003c\/div\u003e\n\u003c\/div\u003e\n\u003cdiv\u003e\n\u003cdiv class=\"v\"\u003e~25 \u003csmall\u003e€\/GB\u003c\/small\u003e\n\u003c\/div\u003e\n\u003cdiv class=\"k\"\u003eVRAM cost\u003c\/div\u003e\n\u003c\/div\u003e\n\u003cdiv\u003e\n\u003cdiv class=\"v\"\u003ePCIe 1.0\u003c\/div\u003e\n\u003cdiv class=\"k\"\u003e×4 or ×16 link\u003c\/div\u003e\n\u003c\/div\u003e\n\u003c\/div\u003e\n\u003cdiv class=\"rtx-sec\"\u003e\n\u003cdiv class=\"rtx-ey\"\u003eRead this first\u003c\/div\u003e\n\u003ch2 class=\"t\"\u003eWhat this card is — and is not\u003c\/h2\u003e\n\u003cp class=\"lead\"\u003eThe CMP 170HX was sold by NVIDIA as a dedicated mining card: a cut-down GA100 with its PCIe link deliberately locked to PCIe 1.0 and most display and compute functions restricted. These units have been modified after the fact to carry 64 GB of HBM2e. That makes them an unusually cheap way to hold a large model entirely in high-bandwidth memory — and a poor choice for anything that depends on fast host transfers, multi-GPU scaling or training.\u003c\/p\u003e\n\u003cdiv class=\"rtx-note\"\u003e\n\u003cstrong\u003eThis is modified hardware operating outside its intended configuration.\u003c\/strong\u003e Compute is partially fused off compared with a full A100, so treat VRAM capacity — not throughput — as the reason to buy. Individual cards vary: some are more fully unlocked than others, and we cannot promise which you will receive. If you need guaranteed, predictable compute, buy an RTX PRO 6000 or a proper datacenter card instead and we will happily quote one.\u003c\/div\u003e\n\u003c\/div\u003e\n\u003cdiv class=\"rtx-sec\"\u003e\n\u003cdiv class=\"rtx-ey\"\u003eThe PCIe limit\u003c\/div\u003e\n\u003ch2 class=\"t\"\u003eWhy the ×16 option costs €150 more\u003c\/h2\u003e\n\u003cp class=\"lead\"\u003eBoth versions are locked to PCIe \u003cstrong\u003e1.0\u003c\/strong\u003e signalling — that part cannot be undone. What differs is lane count, and it changes host-to-card bandwidth by 4×:\u003c\/p\u003e\n\u003ctable class=\"spec\"\u003e\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd\u003ePCIe 1.0 ×4 — €1,600\u003c\/td\u003e\n\u003ctd\u003e~1 GB\/s · filling 64 GB takes roughly a minute\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003ePCIe 1.0 ×16 — €1,750\u003c\/td\u003e\n\u003ctd\u003e~4 GB\/s · filling 64 GB takes roughly 15 seconds\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003c\/tbody\u003e\u003c\/table\u003e\n\u003cdiv class=\"rtx-note\"\u003eOnce weights are resident in VRAM, inference runs from HBM2e and the PCIe link barely matters. The link speed dominates \u003cem\u003emodel load time\u003c\/em\u003e, host↔device streaming, and any multi-GPU communication. If you load a model once and serve it for hours, ×4 is fine. If you swap models often, stream data continuously, or plan to split a model across cards, pay the €150.\u003c\/div\u003e\n\u003cdiv class=\"rtx-note\"\u003eBoth options come from the same batch of cards — the ×16 conversion is work we carry out in-house before dispatch, which is what the €150 covers. Allow a little extra time on ×16 orders for the conversion and re-testing.\u003c\/div\u003e\n\u003c\/div\u003e\n\u003cdiv class=\"rtx-sec\"\u003e\n\u003cdiv class=\"rtx-ey\"\u003eSpecifications\u003c\/div\u003e\n\u003ch2 class=\"t\"\u003eTechnical data\u003c\/h2\u003e\n\u003ctable class=\"spec\"\u003e\u003ctbody\u003e\n\u003ctr\u003e\n\u003ctd\u003eGPU\u003c\/td\u003e\n\u003ctd\u003eNVIDIA GA100 — Ampere\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eMemory\u003c\/td\u003e\n\u003ctd\u003e64 GB HBM2e (modified — not a stock configuration)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eCompute\u003c\/td\u003e\n\u003ctd\u003ePartially fused off vs A100 — capacity-oriented, not throughput-oriented\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eInterface\u003c\/td\u003e\n\u003ctd\u003ePCIe 1.0 ×4 or PCIe 1.0 ×16 (select above)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eHost bandwidth\u003c\/td\u003e\n\u003ctd\u003e~1 GB\/s (×4) · ~4 GB\/s (×16)\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eDisplay outputs\u003c\/td\u003e\n\u003ctd\u003eNone — compute only\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eCooling\u003c\/td\u003e\n\u003ctd\u003ePassive — requires chassis airflow\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eCondition\u003c\/td\u003e\n\u003ctd\u003eUsed, ex-mining — tested before dispatch\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eWarranty — card\u003c\/td\u003e\n\u003ctd\u003e6 months\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003ctr\u003e\n\u003ctd\u003eWarranty — modified VRAM\u003c\/td\u003e\n\u003ctd\u003e14 days\u003c\/td\u003e\n\u003c\/tr\u003e\n\u003c\/tbody\u003e\u003c\/table\u003e\n\u003c\/div\u003e\n\u003cdiv class=\"rtx-sec\"\u003e\n\u003cdiv class=\"rtx-ey\"\u003eBest for\u003c\/div\u003e\n\u003ch2 class=\"t\"\u003eWhere it fits\u003c\/h2\u003e\n\u003cul class=\"tick\"\u003e\n\u003cli\u003eServing one large model that stays resident in VRAM — load once, run for hours.\u003c\/li\u003e\n\u003cli\u003eExperimenting with big models on a budget, where 64 GB for €1,600 is the whole point.\u003c\/li\u003e\n\u003cli\u003eBatch inference jobs that are VRAM-bound rather than transfer-bound.\u003c\/li\u003e\n\u003cli\u003eLearning and development on large-model workflows without datacenter-GPU spend.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003ch2 class=\"t\" style=\"margin-top:18px\"\u003eWhere it does not fit\u003c\/h2\u003e\n\u003cul class=\"tick\"\u003e\n\u003cli\u003eTraining or fine-tuning — PCIe 1.0 and reduced compute both work against you.\u003c\/li\u003e\n\u003cli\u003eMulti-GPU tensor-parallel setups — the interconnect is far too slow.\u003c\/li\u003e\n\u003cli\u003eWorkloads that stream data continuously from host memory or disk.\u003c\/li\u003e\n\u003cli\u003eAnything needing display output, or a card you can rely on for years of production duty.\u003c\/li\u003e\n\u003c\/ul\u003e\n\u003c\/div\u003e\n\u003cdiv class=\"rtx-sec\"\u003e\n\u003cdiv class=\"rtx-ey\"\u003eQuestions\u003c\/div\u003e\n\u003ch2 class=\"t\"\u003eFAQ\u003c\/h2\u003e\n\u003cdiv class=\"rtx-faq\"\u003e\n\u003cdetails open\u003e\u003csummary\u003eWhy is the VRAM warranty only 14 days?\u003c\/summary\u003e\u003cp\u003eBecause the 64 GB memory configuration is an aftermarket modification the hardware was never designed for. We test every card before dispatch, but we will not pretend a modified memory subsystem carries the same long-term guarantee as a factory part. The card itself is covered for 6 months; the modified VRAM for 14 days. Test it properly as soon as it arrives.\u003c\/p\u003e\u003c\/details\u003e\n\u003cdetails\u003e\u003csummary\u003eWill my framework see all 64 GB?\u003c\/summary\u003e\u003cp\u003eIn our testing the full capacity is addressable, which is the entire reason to buy this card. Compute capability is a different matter — it is partially restricted versus a real A100, and cards vary between units. Tell us your intended workload before ordering and we will give you a straight answer about whether this is the right purchase.\u003c\/p\u003e\u003c\/details\u003e\n\u003cdetails\u003e\u003csummary\u003eCan I get a fully unlocked card?\u003c\/summary\u003e\u003cp\u003eSome units are less restricted than others, but it is genuinely a lottery and we will not sell you a promise we cannot keep. Order on the basis of 64 GB of VRAM at a low price; treat anything beyond that as a bonus.\u003c\/p\u003e\u003c\/details\u003e\n\u003cdetails\u003e\u003csummary\u003eCan I run several in one machine?\u003c\/summary\u003e\u003cp\u003eYou can physically, and each card keeps its own 64 GB, so independent jobs per card work. What does not work well is splitting a single model across cards — PCIe 1.0 makes tensor-parallel communication the bottleneck by a wide margin.\u003c\/p\u003e\u003c\/details\u003e\n\u003cdetails\u003e\u003csummary\u003eDoes it need special cooling?\u003c\/summary\u003e\u003cp\u003eYes. It is a passive card with no fan of its own and expects a server chassis with a proper front-to-back airflow path. It will overheat in a normal desktop case.\u003c\/p\u003e\u003c\/details\u003e\n\u003cdetails\u003e\u003csummary\u003eHow does this compare with buying a professional card?\u003c\/summary\u003e\u003cp\u003eOn VRAM price nothing comes close — roughly €25\/GB here against about €141\/GB for a 96 GB RTX PRO 6000. On reliability, compute, warranty, PCIe bandwidth and resale, the professional card wins on every count. Pick based on which of those matters to you.\u003c\/p\u003e\u003c\/details\u003e\n\u003c\/div\u003e\n\u003cdiv class=\"rtx-note\"\u003eNot sure this is the right card for your workload? Tell us the model and how you intend to serve it and we will say honestly whether this or a Kentino AI build fits better.\u003c\/div\u003e\n\u003c\/div\u003e\n\u003c\/div\u003e","brand":"NVIDIA","offers":[{"title":"PCIe 1.0 x4","offer_id":53617506648392,"sku":null,"price":1600.0,"currency_code":"EUR","in_stock":true},{"title":"PCIe 1.0 x16","offer_id":53617506681160,"sku":null,"price":1750.0,"currency_code":"EUR","in_stock":true}],"url":"https:\/\/kentino.com\/uk\/products\/nvidia-cmp-170hx-64-gb-hbm2e-modified-ex-mining","provider":"Kentino","version":"1.0","type":"link"}