CMP 170HX 64GB 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 64GB is a modded 2021 NVIDIA mining card that, patched on Linux, runs 70B-class models at 4-bit. Under Windows it shows only 8GB.
eBay
CMP 170HX 64GB 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 64GB specifications
- Memory64 GB · HBM2e
- Bandwidth1,493 GB/s
- FP16 / BF16202 TFLOPS
- INT8404 TOPS
- Bus width
- 4,096-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
- Jun 1, 2021
Best CMP 170HX 64GB 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.
UsedNVIDIA CMP 170HX 8GB / 64GB Unlockable Mining/Compute Card$2,800Buy on eBay
UsedNVIDIA CMP 170HX 8GB (Unlockable To 64GB, 74+ SMs) HBM2e + Power Adapter - Used$2,950+ shippingBuy on eBay
UsedUS Stock NVIDIA CMP 170HX 8GB HBM2e 64GB Unlocked AI GPU Tested PCIe 2.0 x16 Mod$3,999Buy on eBay
UsedNVIDIA CMP 170HX 8GB HBM2 GPU Gen2 x4 Unlocked 64GB AI Qwen 3.8 Card ComfyUI GPT$5,000Buy on eBay
UsedNVIDIA CMP 170HX 8GB 64GB HBM2 GPU + 3D Shroud Gen2 x4 Unlocked AI Card ComfyUI$5,015Buy on eBay
What LLMs can a single CMP 170HX 64GB run?
A single CMP 170HX 64GB tops out at 70B-class dense models at 4-bit. Up to 35B, use 8-bit with full context; 30B-class MoE tops 120 tokens/s.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 160.1 | 256K |
 Qwen3.6 27B27.8B | 11.8 GB | Runs | 83.7 | 256K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 98.1 | 256K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 168.7 | 198K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 80.9 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 176.4 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 141.1 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 61.3 | 256K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 62.3 | 256K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 154.1 | 198K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 55.4 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 156.7 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Runs | 119 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Runs | 37.6 | 256K |
| Qwen3.8 27B27.8B | 29 GB | Runs | 37.1 | 256K |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Runs | 130.4 | 198K |
| Gemma 4 31B30.7B | 32.6 GB | Runs | 32.8 | 256K |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Runs | 134 | 256K |
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 64GB cards run?
Two CMP 170HX 64GB cards run Qwen3.8 Flash Next at 4-bit, about 111 tokens/s in one user's test; DeepSeek V4 Flash needs four. No NVLink.
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | 2 cards | $6,000 | 128 GB | 135.2 | 500 W |
 DeepSeek V4 Flash284B (13B active) | 96.8 GB | 2 cards | $6,000 | 128 GB | 115.6 | 500 W |
 Hy3295B (21B active) | 107.2 GB | 2 cards | $6,000 | 128 GB | 62.7 | 500 W |
| GLM 5.3 Flash320B (18B active) | 108.7 GB | 2 cards | $6,000 | 128 GB | 98 | 500 W |
| GLM 5.3744B (40B active) | 253.9 GB | 8 cards | $24,000 | 512 GB | 55.9 | 2,000 W |
| Hy4 preview770B (49B active) | 307.2 GBest. | 8 cards | $24,000 | 512 GB | 43.4 | 2,000 W |
 Muse Spark Open816B (42B active) | 311.2 GB | 8 cards | $24,000 | 512 GB | 50.6 | 2,000 W |
 Kimi K2.7 Code1,000B (32B active) | 339.5 GB | 8 cards | $24,000 | 512 GB | 65.9 | 2,000 W |
| Kimi K2.61,000B (32B active) | 340.5 GB | 8 cards | $24,000 | 512 GB | 65.8 | 2,000 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 111.3 GB | 2 cards | $6,000 | 128 GB | 119.6 | 500 W |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | 4 cards | $12,000 | 256 GB | 91.1 | 1,000 W |
| Hy3295B (21B active) | 182.2 GB | 4 cards | $12,000 | 256 GB | 48.5 | 1,000 W |
| GLM 5.3 Flash320B (18B active) | 199.7 GB | 4 cards | $12,000 | 256 GB | 68.1 | 1,000 W |
| GLM 5.3744B (40B active) | 467.3 GB | 8 cards | $24,000 | 512 GB | 35.8 | 2,000 W |
| Hy4 preview770B (49B active) | 467.3 GB | 8 cards | $24,000 | 512 GB | 31.3 | 2,000 W |
| DeepSeek V4.1 Flash763B (8 / 16B active) | 510.3 GB | 8 cardsEngram in RAM | $24,000 | 512 GB | 60 | 2,000 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash284B (13B active) | 161.9 GB | 4 cards | $12,000 | 256 GB | 88.9 | 1,000 W |
| Qwen3.8 Flash Next176B (6B active) | 188.2 GB | 4 cards | $12,000 | 256 GB | 93.9 | 1,000 W |
| Hy3295B (21B active) | 317.7 GB | 8 cards | $24,000 | 512 GB | 34.4 | 2,000 W |
| GLM 5.3 Flash320B (18B active) | 341 GB | 8 cards | $24,000 | 512 GB | 46.3 | 2,000 W |
Rows marked “Engram in RAM” need about 203 GB of extra system RAM. Full calculation rules are on the Methodology page.
Who it suits
Good fit
- Running 70B- to 120B-class models on one card
Llama 3.3 70B at 4-bit fits with room to spare, and gpt-oss 120B at 4-bit just fits
- You want 30B-class models at 8-bit with long context
Gemma 4 31B at 8-bit can use the full 256K, at about 33 tokens/s
- You want to run Qwen3.8 Flash Next on as few cards as possible
Two cards hold it at 4-bit
Not a good fit
- You use Windows
There's no Windows version of the patch, so under Windows the card works only as an 8GB card
- You switch models often
PCIe Gen2 x4 manages only about 2 GB/s, so loading a 64GB model takes more than half a minute
- You have no server airflow and dislike noise
It's passively cooled and throttles at 85°C, and only a high-static-pressure fan keeps it in check
Before you buy
Where the 64GB comes from
“64G” in a listing title usually means the card can be unlocked to 64GB. The card itself is still a mining card that ships with only 8GB enabled, with the rest of the memory locked by firmware inside the chip package. The unlocking tool, cmpunlocker, modifies the open-source NVIDIA driver on your Linux system and doesn't flash the card's firmware. Once installed, it lets you use the full 64GB, while also lifting the compute cap and raising PCIe to Gen2.
The 8GB version unlocks more reliably, but test it yourself
Community experience is that the 8GB version (SK hynix memory) is fairly stable unlocked to 64GB, while the 10GB version (Samsung memory) is unstable at 80GB. Still, every card needs its own testing: after unlocking, the Level1Techs user first filled and tested all 64GB section by section with cmp_memory_tester, and only trusted the card once everything passed. Early versions of the patch threw Xid 31 errors when VRAM was nearly full; that was only fixed by a patch merged in August 2026 (cmpunlocker PR #32), so install the new version.
Without unlocking, prompt processing is very slow
The factory firmware caps compute hard: niconiconi measured only about 42 TFLOPS of FP16, which makes prompt processing (prefill) very slow. The unlock patch lifts this cap as well. A user has measured the result after unlocking: on Level1Techs, someone running the INT8 version of Qwen3.8 27B in vLLM got about 1,500 tokens per second of prompt processing, and 95 to 103 tokens/s of generation with MTP speculative decoding on (we haven't re-tested this). Speeds in the table don't account for MTP.
Cooling needs a high-static-pressure fan
The heatsink is so dense that ordinary fans can't push air through it. The Level1Techs user started with a 3000 RPM 80mm fan, and the card hit 85°C within a minute and a half, with clocks dropping to 800 MHz. Switching to a high-static-pressure 92mm fan with a shroud helped, and the final fix was pulling air out from the tail end of the card: five minutes at full load held steady below 82°C with no more throttling. Power comes from the power supply's CPU 8-pin (EPS) cable or an adapter cable, and the card has no video output.
Compared with similar GPUs
CMP 170HX 64GB vs CMP 170HX 40GB, Instinct MI210, and RTX A6000: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | CMP 170HX 64GB | CMP 170HX 40GB | Instinct MI210 | RTX A6000 |
|---|---|---|---|---|
| Memory | 64 GB | 40 GB38% less | 64 GBSame | 48 GB25% less |
| Bandwidth | 1,493 GB/s | 1,560 GB/s4% higher | 1,638 GB/s10% higher | 768 GB/s49% lower |
| BF16 / FP16 compute | 202 TFLOPS | 202 TFLOPSSame | 181 TFLOPS10% lower | 155 TFLOPS23% lower |
| INT8 compute | 404 TOPS | 404 TOPSSame | 181 TOPS55% lower | 310 TOPS23% lower |
| Qwen3.8 27B speed (2-bit) | 98.1 t/s | 101.9 t/s4% faster | 106.2 t/s8% faster | 54.1 t/s45% slower |
| Largest dense model (4-bit) | 107B | 65B39% less | 107BSame | 79B26% less |
| Power | 250 W | 250 WSame | 300 W20% higher | 300 W20% higher |
| GPU interconnect | Not supported | Not supported | — | NVLink 112.5 GB/s |
| Interface | PCIe 2.0 x4 | PCIe 2.0 x4 | PCIe 4.0 x16 | PCIe 4.0 x16 |
| Released | 2021-06 | 2021-09 | 2022-03 | 2020-10 |
| 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.
Amazon.jp
Xianyu