NVIDIA CMP 170HX 64 GB HBM2e (Modified, Ex-Mining)
Enterprise-grade platform, assembled and tested in the EU.
NVIDIA CMP 170HX 64 GB HBM2e — cheap VRAM for resident inference
A 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.
What this card is — and is not
The 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.
Why the ×16 option costs €150 more
Both versions are locked to PCIe 1.0 signalling — that part cannot be undone. What differs is lane count, and it changes host-to-card bandwidth by 4×:
| PCIe 1.0 ×4 — €1,600 | ~1 GB/s · filling 64 GB takes roughly a minute |
| PCIe 1.0 ×16 — €1,750 | ~4 GB/s · filling 64 GB takes roughly 15 seconds |
Technical data
| GPU | NVIDIA GA100 — Ampere |
| Memory | 64 GB HBM2e (modified — not a stock configuration) |
| Compute | Partially fused off vs A100 — capacity-oriented, not throughput-oriented |
| Interface | PCIe 1.0 ×4 or PCIe 1.0 ×16 (select above) |
| Host bandwidth | ~1 GB/s (×4) · ~4 GB/s (×16) |
| Display outputs | None — compute only |
| Cooling | Passive — requires chassis airflow |
| Condition | Used, ex-mining — tested before dispatch |
| Warranty — card | 6 months |
| Warranty — modified VRAM | 14 days |
Where it fits
- Serving one large model that stays resident in VRAM — load once, run for hours.
- Experimenting with big models on a budget, where 64 GB for €1,600 is the whole point.
- Batch inference jobs that are VRAM-bound rather than transfer-bound.
- Learning and development on large-model workflows without datacenter-GPU spend.
Where it does not fit
- Training or fine-tuning — PCIe 1.0 and reduced compute both work against you.
- Multi-GPU tensor-parallel setups — the interconnect is far too slow.
- Workloads that stream data continuously from host memory or disk.
- Anything needing display output, or a card you can rely on for years of production duty.
FAQ
Why is the VRAM warranty only 14 days?
Because 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.
Will my framework see all 64 GB?
In 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.
Can I get a fully unlocked card?
Some 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.
Can I run several in one machine?
You 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.
Does it need special cooling?
Yes. 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.
How does this compare with buying a professional card?
On 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.
The questions buyers ask us most often before ordering a server.
How long does it take?
Machines built from components we hold ship quickly; anything requiring a specific GPU generation depends on supply. We give you a date before you pay, and if it moves we tell you rather than letting you find out.
Can the configuration be changed before you build it?
Almost always. GPUs, memory, storage and cooling are chosen per order, and the listed configuration is a starting point rather than a fixed package. If you need more VRAM, faster storage or a different cooling approach, say so before you order and we will quote the change.
Can I collect the server in person?
You can. Our warehouse is in Prague, and collection in person is welcome — most people who come use the visit to go through the machine with our engineer and ask the questions that are awkward over email. For orders within the Czech Republic we also try to deliver personally and walk you through the setup on site.
Can I talk to someone who actually understands the workload?
Yes. We have an engineer who works on AI systems specifically, not a general sales desk. If your question is about batch sizes, quantisation, interconnect or where your bottleneck will be, ask it — that conversation usually changes the configuration for the better.
Which model can I run on this configuration?
Yes. Every machine is assembled, burn-in tested and benchmarked on real AI workloads before it ships, and it leaves us with an LLM already installed and running. You plug it in, connect it to your network and start work — the only decision left is which project it runs first.
How do you test a server before shipping?
We run it against actual AI workloads rather than synthetic scores: inference throughput, sustained load behaviour and thermals under continuous operation. You get the benchmark results with the machine, so the performance you were promised is the performance you can verify on day one.
Is the server ready to run when it arrives?
Yes. Every machine is assembled, burn-in tested and benchmarked on real AI workloads before it ships, and it leaves us with an LLM already installed and running. You plug it in, connect it to your network and start work — the only decision left is which project it runs first.
Ships from our EU warehouse. Heavy items may require freight arrangement — contact us for a shipping quote and lead time. 2-year limited warranty with advanced RMA support; extended warranty available.
Not exactly what you need?
Tell us your workload and we'll spec this platform around it — GPUs, memory, storage and cooling matched to what you actually run.