CMP 170HX 40GB Price & Used Prices: Which Local LLMs Can It Run?
Price alert
We'll email you when the eBay (US) price reaches your target.
The CMP 170HX 40GB is a modded 2021 NVIDIA mining card; a driver patch unlocks 40GB for models up to 35B at 8-bit. The patch is Linux-only.
eBay
CMP 170HX 40GB price history
eBay
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
Full CMP 170HX 40GB specifications
- Memory40 GB · HBM2e
- Bandwidth1,560 GB/s
- FP16 / BF16202 TFLOPS
- INT8404 TOPS
- Bus width
- 5,120-bit
- Architecture
- Ampere (GA100)
- Cores
- 4,480
- Tensor cores
- 280
- Ecosystem
- CUDA
- Interface
- PCIe 2.0 x4
- NVLink
- Not supported
- Power
- 250 W
- Power connector
- 8-pin CPU (EPS-12V)
- Slots
- 2
- Length
- 267 mm
- Type
- Data center
- Launch date
- Sep 1, 2021
Best CMP 170HX 40GB deals on eBay
Hand-picked deals from trusted eBay sellers. Shop with confidence.
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
What LLMs can a single CMP 170HX 40GB run?
Unlocked, one CMP 170HX 40GB fits every model up to 35B at 8-bit, with 30B-class MoE near 130 tokens/s. 70B-class dense needs two cards.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| 4.1 GB | Runs | 202.5 | 256K | |
| 4.7 GB | Runs | 181.1 | 256K | |
| 10.5 GB | Runs | 161.9 | 256K | |
| 11.8 GB | Runs | 87 | 256K | |
| 9.8 GB | Runs | 101.9 | 256K | |
| 11.9 GB | Runs | 170.2 | 198K | |
| 11.8 GB | Runs | 84.1 | 256K | |
| 12.3 GB | Runs | 177.7 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| 5.7 GB | Runs | 159.3 | 256K | |
| 7.1 GB | Runs | 132.8 | 256K | |
| 16.9 GB | Runs | 143.2 | 256K | |
| 16.8 GB | Runs | 63.8 | 256K | |
| 16.5 GB | Runs | 64.9 | 256K | |
| 18.3 GB | Runs | 156 | 198K | |
| 18.3 GB | Runs | 57.7 | 256K | |
| 22.1 GB | Runs | 158.5 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| 9.5 GB | Runs | 105.8 | 256K | |
| 12.7 GB | Runs | 81.9 | 256K | |
| 26.9 GB | Runs | 121.3 | 256K | |
| 28.6 GB | Runs | 39.2 | 197K | |
| 29 GB | Runs | 38.7 | 191K | |
| 31.8 GB | Runs | 132.6 | 181K | |
| 32.6 GB | Runs | 34.2 | 100K | |
| 36.9 GB | Runs | 136.2 | 237K |
RunsFits fully in VRAM; speed is estimated from VRAM bandwidth.
Needs RAM offloadMoE only: expert weights sit in system RAM, so speed is estimated from 70 GB/s RAM bandwidth.
Too bigWon’t fit on one card at this quantization, even with RAM offload.
Full calculation rules are on the Methodology page.
What LLMs can multiple CMP 170HX 40GB cards run?
Two CMP 170HX 40GB cards run gpt-oss 120B at 8-bit at about 133 tokens/s; large MoE needs four. Four PCIe 2.0 lanes per card make syncing slow.
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| 78.9 GB | 4 cards | $11,736 | 160 GB | 137.4 | 1,000 W | |
| 96.8 GB | 4 cards | $11,736 | 160 GB | 117.9 | 1,000 W | |
| 107.2 GB | 4 cards | $11,736 | 160 GB | 64.6 | 1,000 W | |
| 108.7 GB | 4 cards | $11,736 | 160 GB | 100.4 | 1,000 W | |
| 253.9 GB | 8 cards | $23,472 | 320 GB | 57.7 | 2,000 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| 111.3 GB | 4 cards | $11,736 | 160 GB | 121.9 | 1,000 W | |
| 155.1 GB | 8 cards | $23,472 | 320 GB | 93.4 | 2,000 W | |
| 182.2 GB | 8 cards | $23,472 | 320 GB | 50.2 | 2,000 W | |
| 199.7 GB | 8 cards | $23,472 | 320 GB | 70.2 | 2,000 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| 161.9 GB | 8 cards | $23,472 | 320 GB | 91.1 | 2,000 W | |
| 188.2 GB | 8 cards | $23,472 | 320 GB | 96.2 | 2,000 W |
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- You use only Linux, don't mind tinkering and want A100 generation speed
It's as fast as the A100 40GB on the same model, about 65 tokens/s on Qwen3.8 27B at 4-bit
- You want 30B-class models at 8-bit
Qwen3.8 27B at 8-bit (29 GB) fits and runs at about 39 tokens/s, which 24GB cards can't do
- You mainly run MoE models
Qwen3.6 35B A3B and GLM 4.7 Flash run at about 156 to 159 tokens/s at 4-bit
Not a good fit
- You use Windows or don't want to touch the kernel
The unlock patch supports only Linux and specific versions of the open-source driver, and you have to turn off Secure Boot
- You switch models often or need to run weights from system RAM
PCIe is only Gen2 x4, so moving data between VRAM and RAM is slow
- You want a warranty and worry-free long-term use
40GB is not an official spec, so there's no after-sales support once it's sold
Before you buy
The 40GB is unlocked, not soldered on
The “40GB” sellers advertise isn't the card's factory spec. The CMP 170HX ships with only 10GB enabled (another version has 8GB); the rest of the memory is actually inside the chip package, locked by firmware, and flashing the BIOS can't change that. Unlocking relies on the open-source tool cmpunlocker, which patches the driver to release the locked memory at boot, while also lifting the compute cap and raising PCIe to Gen2. So the patch lives in your computer's operating system and the card itself is unchanged: moving to another machine or reinstalling the OS means installing it again. In theory the 10GB version can be unlocked to 80GB, but the community reports that's unstable, so the tool uses the 40GB setting for 10GB cards.
Linux only, with limited driver versions
The patch has very specific system requirements: Linux x86-64, root access, and Secure Boot turned off (the modified driver module isn't signed). It accepts only a few nvidia-open driver versions on its list (currently the 610 and 615 series); if your installed version isn't on the list, the install script simply exits. You need to reinstall it after upgrading the kernel or driver. There's no such patch for Windows, where the card can only be used at its factory 10GB.
Test the memory while you can still return it
Unlocked memory isn't necessarily all good: Tom's Hardware warns that the locked portion may have been defective to begin with, and wccftech notes that some cards are only stable at a smaller capacity. When the card arrives, use a tool like cmp_memory_tester to fill and test all 40GB from start to finish while the return window is still open.
The board is nearly identical to the A100's
Taken apart, its circuit board is almost the same as the A100 40GB PCIe's, just with some components left unpopulated, so A100 water blocks fit directly (one user went with the Bykski N-TESLA-A100-X-V2). Like the A100, it takes power through one 8-pin CPU (EPS) connector, and the card has no fan and no video output.
Compared with similar GPUs
CMP 170HX 40GB vs A100 40GB PCIe, CMP 170HX 64GB, and GeForce RTX 4090: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | ||||
|---|---|---|---|---|
| Memory | 40 GB | 40 GBSame | 64 GB60% more | 24 GB40% less |
| Bandwidth | 1,560 GB/s | 1,555 GB/sSame | 1,493 GB/s4% lower | 1,008 GB/s35% lower |
| BF16 / FP16 compute | 202 TFLOPS | 312 TFLOPS54% higher | 202 TFLOPSSame | 330 TFLOPS64% higher |
| INT8 compute | 404 TOPS | 624 TOPS54% higher | 404 TOPSSame | 661 TOPS64% higher |
| Qwen3.8 27B speed (2-bit) | 101.9 t/s | 101.6 t/sSame | 98.1 t/s4% slower | 69.3 t/s32% slower |
| Largest dense model (4-bit) | 65B | 65BSame | 107B65% more | 37B43% less |
| Power | 250 W | 250 WSame | 250 WSame | 450 W80% higher |
| GPU interconnect | Not supported | NVLink 600 GB/s | Not supported | Not supported |
| Interface | PCIe 2.0 x4 | PCIe 4.0 x16 | PCIe 2.0 x4 | PCIe 4.0 x16 |
| Released | 2021-09 | 2020-06 | 2021-06 | 2022-10 |
| eBay median |
As an eBay Partner, VRAMGlass may be compensated if you buy through the eBay links on this page. The price you pay stays the same.
