This GPU can be Hacked to Unlock Secret Memory

Linus Tech Tips

962,778 views • yesterday Save 15 min 4 min read

Video Summary

Nvidia's CMP 170HX mining cards, once dismissed as worthless e-waste due to crippled hardware and locked memory, have been resurrected by modders. By bypassing Nvidia's artificial restrictions—which limited the cards to 8GB of memory and throttled floating-point performance—enthusiasts are unlocking the full 64GB of HBM2e memory and PCI Express capabilities. This transformation turns these discarded mining units into surprisingly potent tools for running large AI models, offering a high-memory alternative to consumer gaming GPUs that often struggle with the memory requirements of modern language models.

However, the path to performance is fraught with risk. Many of these cards were demoted from high-end A100 chips because they contained defective memory or hardware, meaning buyers are essentially gambling on "luck of the draw" when purchasing them on the secondary market. While successful mods—including potential capacitor soldering to restore full PCI Express bandwidth—can yield impressive results, the cards remain prone to instability, heat-related issues, and the inherent volatility of the used hardware market.

Short Highlights

The CMP 170HX can be unlocked to access its full 64GB of HBM2e memory and improved floating-point performance.
Modding requires disabling secure boot, purging existing drivers, and utilizing a custom driver environment.
Performance is limited by a PCI Express Gen 2 x4 interface, which can be further improved via a physical capacitor mod.
Many units suffer from hardware defects, such as bad memory sectors, because they were originally binned as rejects.
The cards are highly effective for agentic AI workloads and large-context tasks due to their massive VRAM capacity.
Secondary market prices have surged, making the financial viability of these cards increasingly questionable.
Stability is not guaranteed, with some users reporting crashes or inconsistencies under heavy thermal loads.

Key Details

The Resurrection of E-Waste [00:00]

Nvidia's CMP 170HX was originally designed strictly for crypto mining with crippled hardware.
The cards were intentionally limited to 8GB of memory and locked floating-point performance.
"Enterprising modders have found ways to unlock the floating point performance the hidden memory and even the pci express interface turning these from e-waste to ee race racing."

Baseline Performance Testing [01:10]

Stock CMP 170HX cards struggle with larger models due to the 8GB memory cap.
A 3090 Ti serves as a comparison for compute speed, though it lacks the massive VRAM of the mining cards.
"It could barely run any decent sized models and to add insult to injury they locked down the floating point performance too."

The Software Unlock Process [02:35]

Unlocking involves cloning a GitHub repository, disabling secure boot, and purging existing Nvidia drivers.
The process uses a modified driver that unlocks any card installed in the system.
"Instead of unlocking the card you're just running like a modified version of the driver software that unlocks any card that you put in the system."

Unlocked Performance Gains [04:25]

Unlocking the cards provides a massive leap in speed compared to their stock state.
The PCI Express bandwidth remains a bottleneck at Gen 2 x4, impacting model load times.
"This is a completely different world of speed compared to what it was doing before."

Managing Large Models [06:10]

The 64GB of HBM2e memory allows these cards to run models that crash consumer-grade GPUs.
The cards are well-suited for agentic workloads where large context windows are prioritized over raw token speed.
"There's a lot of room for context to make it like half a million or a quarter of a million contacts or running additional models."

The Hardware Gamble [07:55]

Many cards sold on the secondary market are defective, having been binned by Nvidia for memory or power issues.
Memtest utilities can reveal bad sectors on the HBM2e memory, which lead to crashes during computation.
"When that happens what nvidia was doing was they were basically fusing off most of it and selling it as a crap mining card."

Advanced Physical Modifications [09:15]

A capacitor mod can restore the PCI Express interface to full x16 functionality.
Cooling remains a significant challenge, often requiring bulky external shrouds or complex water-cooling setups.
"Dan managed to put the 24 little tiny capacitors back on this thing above the pcie slot unlocking full by 16 functionality."