CMP 170HX

Reported 8 GB and 10 GB

The NVIDIA CMP 170HX is a cryptocurrency mining card built on GA100, the same die as the A100 data-center GPU. NVIDIA sold it in 2021 with most of the chip switched off. Since July 2026 the community has been switching it back on, and an unlocked card is now a 40-64 GB HBM card for local AI at a fraction of an A100's price. It exists in two versions, the 8 GB card and the 10 GB card; see Which card do I have.

The card

Property	Value
GPU	GA100 (GA100-105F-A1 on the 8 GB card, GA100-105A-A1 on the 10 GB card), TSMC 7 nm, 54.2 billion transistors, 826 mm²
Compute capability	8.0 (sm_80, same as A100)
SMs as shipped	70 of the die's 128 (4,480 CUDA cores, 280 tensor cores)
Clocks	1,140 MHz base, 1,410 MHz boost
L2 cache	32 MB
Memory	HBM2e: 8 GB on a 4096-bit bus (SK hynix) or 10 GB on a 5120-bit bus (Samsung)
Memory bandwidth (stock, theoretical)	1.49 TB/s (8 GB) / 1.56 TB/s (10 GB)
PCIe as shipped	x16 edge connector, Gen1 x4 electrical
Power input	one EPS12V 8-pin (CPU 8-pin) connector; ships with a 2x PCIe 8-pin to EPS adapter
Power	250 W TDP on the stock VBIOS; idle about 30-40 W
Cooling	passive heatsink, no fan; needs forced airflow
Display outputs	none
Form factor	dual-slot PCIe card
Released	2021-09-01, original MSRP 4,299 USD

Values from the 170th Street specification page by Amogh (GitHub amoghmunikote), the author of cmpunlocker. Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18 The same site's introduction page gives other SM and bandwidth figures for the variants; the specification page values are used here. Memory vendors: SK hynix HBM2e on the 8 GB card, Samsung HBM2e on the 10 GB card Reported. Source: @darkmaniac7, 2026-10-30 (site owner, from his own cards) A read-only two-card fuse study agrees on the vendors. Source: repo github.com/ahnguyen17/ga100-fuse-re by ahnguyen17 on GitHub, posted by @exile_1ck, 2026-10-07 Earlier posts that disagreed on HBM2 versus HBM2e are listed on Which card do I have#Memory vendor.

Board and hardware

A100 PCB with parts left off. The board is the A100 PCIe design with fewer components. Compared with the A100 PCIe board it leaves 5 or 6 Vcore VRM phases unpopulated, Source: @wingrimez_40467, 2026-10-22 along with resistors on the NVLink bus and the parts of the 3.3 V NVLink clock generator. Source: @ghostinthedatacenter_87281, 2026-10-22 The NVLink connector pins themselves are populated and connected to the core. Source: @cute_sc2, 2026-10-05
Out-of-spec edge connector. The PCB shape is outside the PCIe CEM spec and the gold fingers lack the key notch on the left. Source: @dalzyyy, 2026-10-11
Heat spreader. The 170HX package has an integrated heat spreader, like the A100 40 GB; the A100 80 GB has none, so A100 80 GB water blocks use a raised cold plate and do not fit. Source: @_arcanum_, 2026-10-20 SXM4 A100 heatsinks do not fit either. Source: @wakeupmbrc1, 2026-10-16 Under the IHS is a phase-change paste, not solder. Source: @gnasherrr, 2026-10-18 Warning Some cards have been delidded by accident; the stock heatsink cannot be used on a bare die without making up the lid's thickness, or the die cooks. Source: @civiloid, 2026-10-18
Fan header. The unpopulated fan connector is vestigial: no PWM control and no fan curve in the VBIOS; even the A100 PCIe board has no PWM controller. Source: @endsjeremy, 2026-10-22
Power connector. The card has a single EPS12V 8-pin input; the pinout differs from PCIe 8-pin. With the bundled adapter, both PCIe inputs carry current. Source: @civiloid, 2026-10-15 A CPU (EPS12V) cable works directly, but its double-width latch will not clip onto the card and must be trimmed. Source: @civiloid, 2026-10-09 Warning Feeding both adapter inputs from one PCIe cable and its pigtail has burned hardware. Source: @c133, 2026-10-19 See Cooling and power.
Cooling. The cooler is passive and sized for a server's airflow. Without a fan the card overheats even at idle and falls off the bus at 95 C. Source: @analyticeth, 2026-10-07 Throttling starts around 76 C. Source: @zorg33, 2026-10-20
No graphics. No HDMI or DisplayPort, Source: @xeax, 2026-10-27 and GA100 has hardware decode but no NVENC encoder. Source: @djpandaman, 2026-10-27
No FP8. The card is sm_80; it has no FP8 or FP4 tensor cores. Source: @cibernox, 2026-10-15

PCIe as shipped

The card arrives at PCIe Gen1 x4 (about 0.85 GB/s measured) in a x16 slot; lspci shows LnkCap: Speed 2.5GT/s, Width x16 and LnkSta: Speed 2.5GT/s, Width x4 (downgraded). Source: repo github.com/PixelML/club-170hx by PixelML on GitHub, posted by @c133, 2026-10-20 Two separate locks cause this:

Speed. The silicon was never fused to Gen1. The Gen1 limit is a boot-time setting that cmpunlocker bypasses to reach Gen2; Gen3 and Gen4 are limited by real fuses. Source: @djpandaman, 2026-10-08
Width. The AC-coupling capacitors for 12 of the 16 lanes are not fitted, so the link trains at x4. Fitting 24 small 0402 capacitors restores x16. Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18 See PCIe capacitor mod.

What NVIDIA locked

On the stock card:

Compute throttled. FP32 about 0.41 TFLOPS (12.99 unlocked), FP64 0.19 (6.44), INT8 1.64 (48.03); FP16 tensor math (49 TFLOPS) was left alone. The FP32 throttle hits fused multiply-add, the core instruction of most compute work. Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18
SMs cut to 70 of 128 by fuses.
Memory size limited to 8 GB or 10 GB of the HBM on the package.
PCIe at Gen1 x4, as above.
NVLink fused off and its board parts left out. ECC disabled. Resizable BAR limited to 64 MiB. Graphics APIs: the specification page lists CUDA and OpenCL only.

Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18

Contested Graphics in practice. Members disagree, in the order posted: one said there is no DirectX, OpenGL or Vulkan support; Source: @lenxtroz, 2026-10-15 another answered that Vulkan "is what the inference runs on"; Source: @honzapat_cz, 2026-10-15 xeax said somebody tried gaming and "it works, its not great"; Source: @xeax, 2026-10-15 peter643 ran Minecraft with shaders at 1080p at about 160 fps and about 60 W after enabling vertex work on all 74 SMs, marginally faster than his RTX 3060. Source: @peter643, 2026-10-20 Details: Other CMP cards and other NVIDIA GPUs#Graphics on a 170HX.

Why. The CMP line was NVIDIA's answer to the 2021 GPU shortage: dedicated mining cards meant to be useless for gaming and general compute, so they would not compete with GeForce or data-center products. The 170HX was the only CMP model NVIDIA built itself, from GA100 dies that did not make the A100 bin. Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18 The A100 hides bad HBM blocks by itself; on the 170HX owners have to mark them by hand at boot (see 8 GB card#Memory). Source: @cute_sc2, 2026-10-05 The fuses that enforce the cut are electrical fuses inside the GPU die, not on the board. Source: @wakeupmbrc1, 2026-10-14

From mining card to AI card

The card surfaced in China on 2021-09-01, mining Ethereum at about 164 MH/s at 250 W for roughly 4,700 USD. The Ethereum Merge on 2022-09-15 ended GPU mining of Ethereum and prices collapsed to 300-500 USD; by 2025 cards sold on eBay for 200-400 USD. A 2023 workaround (turning off FMA contraction) recovered some FP32 speed, and a 2025 paper measured it for LLM inference. Source: repo github.com/amoghmunikote/170th-Street by amoghmunikote on GitHub, posted by @markshowa, 2026-10-18 In July 2026 the compute unlock (2026-07-12), the public cmpunlocker release (2026-07-14/15), the memory geometry unlock (2026-07-18) and Gen2 (2026-07-23/24) followed in quick succession. Source: repo github.com/Consensus-Protocol/cmp170hx by Consensus-Protocol on GitHub, posted by @zagitta, 2026-10-12 Reported Petri Krohn's account of the releases, posted 2026-10-24 (a summary of one person's view; see Discord): the proof of concept was described in Jon Pry's Canary paper in mid-June 2026 and was never meant to be released. A "Clean Room" team shared a driverless unlocker on GitHub on 2026-07-12; it unlocked compute only, and within days 170HX prices and the global Pearl hash rate roughly doubled. It did not unlock memory because the memory configuration did not survive a function-level reset (FLR); a version 2 that needed no FLR was ready around 2026-07-18. From 2026-07-15 Russian Telegram channels spread a Chinese unlock, which replaced the Clean Room unlocker in Amogh's cmpunlocker repo; per Petri, a diff showed it was a version of the original proof of concept. He prefers the driverless approach, delivered by a Python script before the drivers load or by UEFI, as easier to understand and maintain than patching the open driver. Source: @petrikrohn_26497, 2026-10-24 The unlock reaches the full configuration through a stack overflow in privileged early boot code, sets the card up, then hands over to the normal NVIDIA driver; not everything can be turned on on every card. Source: @endsjeremy, 2026-10-27 The compute unlocks were later ported to GeForce cards, which brought press attention to the project. Source: @djpandaman, 2026-10-26 The cmpunlocker Discord passed 2,000 members on 2026-10-23. Source: @_j3ns, 2026-10-23 Some owners still mine (now Pearl) when the cards are not serving models. Source: @reala7096, 2026-10-23

The two versions

	8 GB card	10 GB card
PCI ID	10de:20c2	10de:2082
Unlocks to	64 GB, 74 SMs, Gen2	40 GB (48 GB on some), up to 128 SMs, Gen3
Why	more memory	RIR fuse records unused, so most fuses can be overridden

Full table: Variant comparison. How to tell them apart: Which card do I have.