Unlocking CMP 170HX to 80GB VRAM: most users hit a stable wall at 40GB
Feralzi · reddit · 2026-09-16
The NVIDIA CMP 170HX is essentially an A100 locked to 8/10GB, and tools like cmpunlocker supposedly restore the full HBM2e capacity. The poster wants one for a local LLM build but reports conflict: 10GB cards may theoretically reach 80GB, yet it's unstable and most people run them at 40GB.
They're asking for real experiences: hangs, Xid errors, or kernels dying past 40GB in the 80GB config, and whether lighter workloads can go beyond 40GB at all — to decide if the 40GB ceiling makes the card not worth the hassle.
More from Infra
- Hunting a $50 AMD v340l 16GB from 2018 to run local LLMs amid GPU price surge — According_Study_162 · 2026-09-16
- Skip the $9K RTX 5090: fly to Taipei, buy at ~$4,093, vacation for two weeks — DegenDataGuy · 2026-09-16
- Ben Bajarin doubles down on Credo: photonics-plus-SerDes vertical integration drives growth — BenBajarin · 2026-09-16
- Ocean Protocol launches hourly dedicated-GPU inference, H200 from $2.16/hour — w1kke · 2026-09-16
- NVIDIA Vera CPU completes agentic task lifecycles 1.64x faster than x86, Signal65 finds — ryanshrout · 2026-09-16
- Run your AI agents from anywhere with Tailscale and a simple PWA — johnlindquist · 2026-09-16