NVIDIA · Ampere · 40 GB · Qwen3.8 27B · Q4_K_M: ~59 tok/s
NVIDIA A100 40GB SXM4
An A100 out of an HGX server, run on a Chinese SXM4-to-PCIe adapter board: 40 GB of HBM2 at 1.5 TB/s and full Ampere tensor cores for much less than the PCIe card. The board needs a big power supply and real airflow.
Lowest price right now
$2,358.90
eBay · $2,349.95 + $8.95 shipping · $58 per GB · observed 2026-10-18 19:13 UTC
plus an SXM4-to-PCIe adapter, about $570 to $650
Buy the cheapest on eBay, $2,358.90
$2,349.95 + $8.95 shipping · Used · pcrecyclerva 27,240 · 99.7%
Price alert
One email per move, item plus shipping. Clear the price to hear about any drop. Manage alerts.
Price history
Tracking started today. The chart appears once there is a second day of data.
Where to buy
| Vendor | Price | Condition | Observed | |
|---|---|---|---|---|
| eBay cheapest active listing |
$2,358.90 $2,349.95 + $8.95 shipping |
new and used | 2026-10-18 19:13 UTC | Go to eBay |
On eBay now
54 listings passed our filters at the last check. Cheapest delivered first. Seller reputation is feedback count and positive rate.
-
$2,358.90$2,349.95 + $8.95 shipping -
$2,499free shipping -
$2,499free shipping -
$2,859.41$2,849.99 + $9.42 shipping -
$2,894free shipping -
$2,895free shipping
What it runs
For each model, grouped by vendor with the newest first: the largest quant that fits with 8K of context and the estimated speed for one user, at 8K and at 32K of context. On a card, a model too big for its memory can still run with part of the weights in system RAM; we assume 64 GB, and the speed shown accounts for it. The model you picked on the ladder is highlighted. Estimates from the spec sheet, not measurements: see how we estimate.
| Model | Quant | Needs at 8K | Fit | Max context | Decode at 8K | At 32K | Prefill tok/s |
|---|---|---|---|---|---|---|---|
| DeepSeek | |||||||
| DeepSeek V4.1 Flash 552B-A16B Sep 2026 | Q4_K_M | 338 GB | does not fit | ||||
| DeepSeek V4 Flash 284B-A13B Jul 2026 | Q4_K_M | 175 GB | does not fit | ||||
| DeepSeek V3.1 671B-A37B Aug 2025 | Q4_K_M | 411 GB | does not fit | ||||
| Gemma | |||||||
| Gemma 4 12B May 2026 | Q8_0 | 14 GB | fits | 128K+ | ~79 | ~77 | ~4,600 |
| Gemma 4 31B Mar 2026 | Q8_0 | 35 GB | fits | 71K | ~30 | ~29 | ~1,800 |
| Gemma 4 26B-A4B Mar 2026 | Q8_0 | 28 GB | fits | 128K+ | ~130 | ~120 | ~14,000 |
| Gemma 3 27B Mar 2025 | Q8_0 | 31 GB | fits | 112K | ~34 | ~32 | ~2,000 |
| GLM | |||||||
| GLM-5.3 753B-A40B Aug 2026 | Q4_K_M | 462 GB | does not fit | ||||
| GLM-5.3 Flash 320B-A18B Aug 2026 | Q4_K_M | 197 GB | does not fit | ||||
| GLM-4.6 355B-A32B Sep 2025 | Q4_K_M | 221 GB | does not fit | ||||
| GLM-4.5 Air 106B-A12B Jul 2025 | IQ2_M | 39 GB | fits | 12K | ~110 | ~4,500 | |
| gpt-oss | |||||||
| gpt-oss-120b Aug 2025 | Q4_K_M | 73 GB | runs with 34 GB in system RAM | ~24 | ~22 | ~5,400 | |
| gpt-oss-20b Aug 2025 | Q8_0 | 24 GB | fits | 128K+ | ~130 | ~110 | ~15,000 |
| Llama | |||||||
| Llama 4 Scout 109B-A17B Apr 2025 | IQ2_M | 40 GB | fits, short context | 7K | ~94 | ~3,200 | |
| Llama 3.3 70B Dec 2024 | Q3_K_M | 39 GB | fits | 11K | ~27 | ~770 | |
| Llama 3.1 8B Jul 2024 | Q8_0 | 11 GB | fits | 128K+ | ~110 | ~79 | ~6,800 |
| MiMo | |||||||
| MiMo V2.6 Pro 1.02T-A42B Sep 2026 | Q4_K_M | 624 GB | does not fit | ||||
| MiMo V2.6 Flash 309B-A15B Sep 2026 | Q4_K_M | 190 GB | does not fit | ||||
| Mistral | |||||||
| Mistral Medium 3.5 128B Mar 2026 | Q4_K_M | 83 GB | runs with 43 GB in system RAM | ~0.9 | ~0.8 | ~210 | |
| Mistral Small 4 119B-A6.5B Mar 2026 | Q4_K_M | 79 GB | runs with 39 GB in system RAM | ~16 | ~10 | ~4,200 | |
| Mistral Small 3.2 24B Jun 2025 | Q8_0 | 28 GB | fits | 77K | ~38 | ~33 | ~2,300 |
| Nemotron | |||||||
| Nemotron 3.5 Lightning 30B-A3B Aug 2026 | Q8_0 | 35 GB | fits | 128K+ | ~140 | ~140 | ~16,000 |
| Qwen | |||||||
| Qwen3.8 Flash-Next 125B-A6B Aug 2026 | Q4_K_M | 78 GB | runs with 38 GB in system RAM | ~20 | ~19 | ~4,500 | |
| Qwen3.8 27B Aug 2026 (your pick) | Q8_0 | 31 GB | fits | 128K+ | ~35 | ~33 | ~2,000 |
| Qwen3.6 27B Apr 2026 | Q8_0 | 31 GB | fits | 128K+ | ~35 | ~33 | ~2,000 |
| Qwen3.6 35B-A3B Apr 2026 | Q8_0 | 39 GB | fits | 66K | ~150 | ~140 | ~18,000 |
| Qwen3.5 9B Feb 2026 | Q8_0 | 11 GB | fits | 128K+ | ~100 | ~95 | ~6,100 |
| Qwen3 Coder 480B-A35B Jul 2025 | Q4_K_M | 296 GB | does not fit | ||||
| Qwen3 235B-A22B Jul 2025 | IQ2_M | 83 GB | runs with 43 GB in system RAM | ~9.3 | ~8.1 | ~1,200 | |
| Qwen3 32B Apr 2025 | Q8_0 | 38 GB | fits | 13K | ~27 | ~1,700 | |
Specifications
- Memory
- 40 GB HBM2
- Memory bandwidth
- 1,555 GB/s
- Bus width
- 5120-bit
- Shader cores
- 6,912
- Tensor cores
- 432
- FP32
- 19.5 TFLOPS
- FP16 (dense)
- 312 TFLOPS
- Power
- 400 W
- Form factor
- SXM4 module: needs an SXM4-to-PCIe adapter board and serious cooling
- Interface
- PCIe 4.0 x16
- Architecture
- Ampere
- Launched
- 2020
Source: manufacturer specifications.