The CMP 170HX Gives Back 56GB, and Hydra Lifts RTX 50’s Memory Ceiling

A GPL tool restores the HBM2e NVIDIA disabled on the CMP 170HX mining card, and Hydra's preview build raises RTX 50 memory clocks and power limits. What each one unlocks, and what it costs to use.

Saya4 minute readDrivers & Software

Two pieces of third-party software spent this week undoing NVIDIA's product segmentation from opposite ends of the stack. One hands back 56GB of HBM2e that was sitting on a dead mining card the whole time; the other pushes RTX 50 GDDR7 past the clock ceiling the driver enforces. Both work, and both send a bill.

A mining card hands back the memory it always had

The CMP 170HX was sold with 8GB of usable memory. cmpunlocker, a GPL-2.0 project from the developer publishing as amoghmunikote, restores the rest of the HBM2e that the card's firmware and OTP configuration keep switched off, along with full SM compute throughput and working CUDA, with no hardware modification at all (the tool's own documentation). The install script takes a memory profile: an 8GB card unlocks to 64GB, a 10GB card to 40GB.

The requirements make clear this is a compute project rather than a tweak. It needs Linux on x86-64, root, the nvidia-open driver at 610.43.0x specifically, kernel headers, Secure Boot disabled, and a full power-off reboot. Pinning the driver is the part with long-term cost: the patches are built against those exact kernel modules, so an ordinary driver update breaks the unlock until someone rebuilds them. A community install guide documents the mechanism, opening four privilege-level-mask registers through the signed SEC2 Booter path before the GSP firmware boots, then writing the memory geometry and clearing the compute throttles.

Underneath is a GA100 Ampere part with 4480 CUDA cores and a 250W board power, and NVIDIA pinned its PCIe link to Gen 1 to keep it out of non-mining work (TechPowerUp's test of the 170HX). A seventh community patch reportedly lifts that to Gen 2, roughly 1 GB/s to 2 GB/s over an x4 link, with Gen 3 and Gen 4 fused off in silicon and staying off. That is the honest limit on what the card is for: 64GB of HBM2e behind a 2 GB/s host link suits a model that loads once and stays resident, and punishes anything that streams data. The market priced the headline rather than the caveats. Cards that went for around $250 on eBay are now asking past $1,000 (Tom's Hardware on the unlock).

Hydra raises the RTX 50 memory ceiling, and the power ceiling with it

Hydra 2.3B, from Yuri Bubliy (1usmus), adds a memory offset of up to +3000 MHz and a static power limit up to 125 percent for RTX 50 cards from any vendor (Tom's Hardware on the Hydra update, VideoCardz on the unlocked controls). The memory ceiling lands at 36 Gbps on GDDR7, described as a deliberate safety cap rather than where the chips give up. On the Asus ROG Astral LC RTX 5090 used for the demonstration, going from 28 Gbps to 36 Gbps takes a 512-bit bus from 1,792 GB/s to 2,304 GB/s, about 28 percent more bandwidth (Igor's Lab on the 36 Gbps run).

Two things qualify that. The build is a preview distributed through the developer's supporter channels rather than a general release, so availability may look different by the time most owners go looking. And where Tom's Hardware and VideoCardz put the ceiling flatly at 125 percent on any model, Igor's Lab describes the achievable power limit as a function of each card's BIOS and board design. Treat 125 percent as the tool's cap, not a promise a given card will take it.

The failure mode on the memory side is quiet rather than dramatic. GDDR7 links detect and retry failed transfers rather than passing the corruption through, so a memory clock past its real limit costs throughput instead of producing artifacts, and the tuning has to be checked with a bandwidth test and frame times rather than by watching for corruption. The power side is louder: 125 percent of the RTX 5090's 575W board power is roughly 719W, against a 12V-2×6 connector specified for 600W and a slot good for another 75W. On that card the connector and the cooler, not the slider, decide how much of the range is usable.

Which one is worth the trouble

cmpunlocker earns its effort for someone building a Linux inference box who was already willing to pin a driver version and turn off Secure Boot, and earns nothing in a gaming PC: no useful display path, a Gen 2 x4 link at best, and a driver you cannot update casually. Hydra is the opposite case, a tuning tool for a card already in a gaming machine, where the memory offset holds the real headroom and the power slider mostly finds the limits of the cooler and the connector. What the two have in common is where the limit was written. Firmware is a softer place to put a ceiling than silicon, which is why the 170HX gave up its memory and its fused PCIe generations did not.

Sources