NVIDIA CMP 170HX 64 GB HBM2e (unlocked)

Modified version with memory unlocked to 64 GB HBM2e. The NVIDIA CMP 170HX is built on the GA100 chip – the same Ampere generation as the data-centre A100. From the factory it has only 8 GB of memory; we sell units with the memory unlocked to the full 64 GB of HBM2e. That makes it one of the most affordable ways today to fit a large language model entirely into fast GPU memory.

At around CZK 1,500 per gigabyte of memory (incl. VAT) it costs roughly a third of what professional large-memory cards cost. Buy it for memory capacity, not for raw compute – that is the fair comparison, and we would rather say it up front.

Who it is for

• Serving one large model that is loaded into memory and runs for hours (inference, chatbot, internal AI assistant).
• Experimenting with large models on a budget – 64 GB of fast memory for a fraction of the price of a professional card.
• Batch inference where memory size, not compute, is the limit.
• Multi-card servers where the model is split by layers (e.g. llama.cpp) – each card has its own 64 GB.

What it is not for

• Training and fine-tuning – compute is restricted compared with a full A100.
• Deployments that need guaranteed, identical performance on every unit – individual cards vary; for that we will gladly offer you the RTX PRO 6000.
• Ordinary desktop PCs – the card has no fan of its own and no video outputs.

Specifications

• GPU: NVIDIA GA100 (Ampere)
• Memory: 64 GB HBM2e (unlocked, 8 GB from the factory)
• Memory bandwidth: 931–970 GB/s (our measurement)
• Power: up to 250 W, 244–245 W sustained in a 60-minute test
• Compute: restricted compared with A100, varies between units
• ECC: disabled
• Cooling: passive – strong airflow required (server chassis, air duct or a fan per card)
• Video outputs: none
• Software: Linux (Ubuntu 22.04 / 24.04), NVIDIA open driver with the unlock service, llama.cpp / ik_llama.cpp
• Condition: used (ex-mining), tested before dispatch

How the unlock works

The 64 GB of memory is made available by a modified driver that a service re-applies after every system start. After a kernel or driver update the unlock has to be restored, otherwise the card shows 8 GB again – so we recommend pinning the driver version. A step-by-step procedure is in the guide Unlocking 64 GB HBM2e on Ubuntu. We are happy to help with the setup.

How many cards for which model

Large mixture-of-experts models fit across several cards: for example GLM-5.3-Flash in Q4 quality fits entirely into four cards. Sizing for 1–6 cards, required RAM and power supplies is in the guide Running GLM-5.3-Flash on CMP 170HX – a sizing guide. For two or more cards we recommend a board with a dedicated slot per card (EPYC / Threadripper) and a 1–2 kW power supply depending on the number of cards.

FAQ

Will my software see all 64 GB?
Yes, after the unlock the full capacity is addressable – that is the main reason to buy the card. Compute, however, is restricted compared with A100 and varies between units.

Can I put several cards in one server?
Yes. Each card has its own 64 GB and the model is split between them by layers. Tell us which model you want to run and we will recommend the number of cards and the build.

Does it need special cooling?
Yes. The card is passive and will overheat in an ordinary case. It needs a server chassis with front-to-back airflow or an air duct with a fan.

Can you build the whole server for me?
Yes – we build and test an AI server with multiple cards, install the system and the unlock, or run it in our data centre.

Supplied to order – we will confirm the delivery date and available quantity after you order or send an enquiry. For multiple units we prepare an individual offer – contact us.