Instinct MI100 Price & Used Prices: Which Local LLMs Can It Run?
Price alert
We'll email you when the eBay (US) price reaches your target.
AMD's Instinct MI100 (2020) is a 32GB data center card whose 1.2 TB/s keeps 30B-class models fast. It's Linux-only, and official vLLM skips it.
eBay
Instinct MI100 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 Instinct MI100 specifications
- Memory32 GB · HBM2
- Bandwidth1,229 GB/s
- FP16 / BF1692.3 TFLOPS
- INT892.3 TOPS
- Bus width
- 4,096-bit
- Architecture
- CDNA (Arcturus)
- Cores
- 7,680
- Tensor cores
- 480
- Ecosystem
- ROCm
- Interface
- PCIe 4.0 x16
- Power
- 300 W
- Power connector
- 2× 8-pin
- Slots
- 2
- Length
- 267 mm
- Type
- Data center
- Launch date
- Nov 16, 2020
Best Instinct MI100 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.
UsedAMD Instinct MI100 32GB HBM2 HPC Accelerator (2x 8-pin, 300W) - 102-D34315Top Rated$1,200Buy on eBay
UsedAMD Radeon Instinct MI100 32GB HBM2 PCIe 4.0 x16 Professional Graphics CardTop Rated$1,299Buy on eBay
UsedAMD Radeon Instinct MI100 32GB HBM2 PCIe 4.0 x16 GPU Card Advanced Processing$1,500Buy on eBay
Used** AMD Radeon Instinct MI100 32GB HBM2 PCIe 4.0 x16 Graphics Card **$1,503Buy on eBay
UsedAMD Radeon Instinct MI100 32GB HBM2 PCIe 4.0 x16 CDNA Graphics Card$1,505Buy on eBay
UsedAMD Radeon Instinct MI100 32GB PCI Express 4.0 x16 HBM2 CDNA Video Graphics Card$1,510Buy on eBay
What LLMs can a single Instinct MI100 run?
A single MI100 fits up to 35B at 4-bit: Qwen3.8 27B at about 52 tokens/s, 30B-class MoE near 150. At 8-bit only 27B-class fits.
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
 Gemma 4 E4B8B | 3.8 GB | Runs | 181.3 | 128K |
 Qwen3.5 9B9.7B | 4.1 GB | Runs | 169.7 | 256K |
| Gemma 4 12B12B | 4.7 GB | Runs | 150.7 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 10.5 GB | Runs | 152 | 256K |
| Qwen3.6 27B27.8B | 11.8 GB | Runs | 70.4 | 256K |
| Qwen3.8 27B27.8B | 9.8 GB | Runs | 82.8 | 256K |
 GLM 4.7 Flash30B (3B active) | 11.9 GB | Runs | 161.5 | 198K |
| Gemma 4 31B30.7B | 11.8 GB | Runs | 68 | 246K |
| Qwen3.6 35B A3B36B (3B active) | 12.3 GB | Runs | 170.2 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 5 GB | Runs | 147.3 | 128K |
| Qwen3.5 9B9.7B | 5.7 GB | Runs | 131.7 | 256K |
| Gemma 4 12B12B | 7.1 GB | Runs | 108.9 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 16.9 GB | Runs | 131.6 | 256K |
| Qwen3.6 27B27.8B | 16.8 GB | Runs | 51.2 | 245K |
| Qwen3.8 27B27.8B | 16.5 GB | Runs | 52.1 | 250K |
| GLM 4.7 Flash30B (3B active) | 18.3 GB | Runs | 145.5 | 198K |
| Gemma 4 31B30.7B | 18.3 GB | Runs | 46.3 | 168K |
| Qwen3.6 35B A3B36B (3B active) | 22.1 GB | Runs | 148.3 | 256K |
| Model | Size | Runs? | Speed t/s | Max context |
|---|---|---|---|---|
| Gemma 4 E4B8B | 8.2 GB | Runs | 98.2 | 128K |
| Qwen3.5 9B9.7B | 9.5 GB | Runs | 86 | 256K |
| Gemma 4 12B12B | 12.7 GB | Runs | 66.1 | 256K |
| Gemma 4 26B A4B25.2B (3.8B active) | 26.9 GB | Runs | 108.7 | 256K |
| Qwen3.6 27B27.8B | 28.6 GB | Runs | 31.2 | 69K |
| Qwen3.8 27B27.8B | 29 GB | Runs | 30.8 | 63K |
| GLM 4.7 Flash30B (3B active) | 31.8 GB | Runs | 120.4 | 26K |
| Gemma 4 31B30.7B | 32.6 GB | Too big | — | — |
| Qwen3.6 35B A3B36B (3B active) | 36.9 GB | Needs RAM offload | 35.6 | 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 Instinct MI100 cards run?
Four MI100s (128GB) run Qwen3.8 Flash Next at 4-bit at about 109 tokens/s, without official vLLM; llama.cpp is far slower. Infinity Fabric links four.
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 78.9 GB | 4 cards | $6,012 | 128 GB | 125.4 | 1,200 W |
 DeepSeek V4 Flash284B (13B active) | 96.8 GB | 4 cards | $6,012 | 128 GB | 105.3 | 1,200 W |
 Hy3295B (21B active) | 107.2 GB | 4 cards | $6,012 | 128 GB | 54.4 | 1,200 W |
| GLM 5.3 Flash320B (18B active) | 108.7 GB | 4 cards | $6,012 | 128 GB | 87.8 | 1,200 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| Qwen3.8 Flash Next176B (6B active) | 111.3 GB | 4 cards | $6,012 | 128 GB | 109.3 | 1,200 W |
| DeepSeek V4 Flash284B (13B active) | 155.1 GB | 8 cards | $12,024 | 256 GB | 81.1 | 2,400 W |
| Hy3295B (21B active) | 182.2 GB | 8 cards | $12,024 | 256 GB | 41.6 | 2,400 W |
| GLM 5.3 Flash320B (18B active) | 199.7 GB | 8 cards | $12,024 | 256 GB | 59.5 | 2,400 W |
| Model | Size | Cards | Total price | Total VRAM | Speed t/s | Total power |
|---|---|---|---|---|---|---|
| DeepSeek V4 Flash284B (13B active) | 161.9 GB | 8 cards | $12,024 | 256 GB | 79 | 2,400 W |
| Qwen3.8 Flash Next176B (6B active) | 188.2 GB | 8 cards | $12,024 | 256 GB | 83.8 | 2,400 W |
Full calculation rules are on the Methodology page.
Who it suits
Good fit
- You want something faster than a 3090 with 32GB
Gemma 4 31B at 4-bit runs at about 46 tokens/s with up to 168K of context
- You often process long prompts
It has matrix cores, with 7 times the MI50's FP16 matrix compute
- You want 32GB with official ROCm support
It's still on the ROCm 10.0 support list, and Ollama's Linux ROCm build includes it
Not a good fit
- You want vLLM
The MI100 isn't among the AMD cards listed in its official documentation
- You want Windows
ROCm supports it only on a few Linux distributions
- You want to run original BF16 weights directly
BF16 compute is only half of FP16, so you need to convert to FP16 first
Before you buy
Linux only, and the distribution matters
The MI100 works only under Linux, and ROCm supports it on just three distributions: Ubuntu, RHEL and SLES. Debian, Rocky Linux and Oracle Linux aren't on the list. Settle on your OS version before installing, then match the kernel and driver to AMD's compatibility table.
FP16 is twice as fast as BF16
The MI100's BF16 compute is only half its FP16 compute. Running original BF16 weights directly in PyTorch makes long-prompt processing noticeably slower; convert to FP16 first to make full use of the matrix compute. Running GGUF quantized versions with llama.cpp isn't affected.
Airflow
The card has no fan, so all 300W of heat has to be carried off by strong front-to-back airflow through the case. It has the same form factor as the MI50 (full height, dual slot, 267 mm long), so tail-mounted shroud setups made for the MI50 work as a direct reference. With several cards packed side by side, the one sandwiched in the middle is the most likely to overheat.
Compared with similar GPUs
Instinct MI100 vs Instinct MI50 32GB, Instinct MI210, and Tesla V100 32GB: price, VRAM, bandwidth, compute and local LLM speed compared.
| Spec | Instinct MI100 | Instinct MI50 32GB | Instinct MI210 |  Tesla V100 32GB |
|---|---|---|---|---|
| Memory | 32 GB | 32 GBSame | 64 GB100% more | 32 GBSame |
| Bandwidth | 1,229 GB/s | 1,024 GB/s17% lower | 1,638 GB/s33% higher | 900 GB/s27% lower |
| BF16 / FP16 compute | 92 TFLOPS | 27 TFLOPS71% lower | 181 TFLOPS96% higher | 112 TFLOPS21% higher |
| INT8 compute | 92 TOPS | 53 TOPS43% lower | 181 TOPS96% higher | 56 TOPS39% lower |
| Qwen3.8 27B speed (2-bit) | 82.8 t/s | 70.3 t/s15% slower | 106.2 t/s28% faster | 62.6 t/s24% slower |
| Largest dense model (4-bit) | 51B | 51BSame | 107B110% more | 51BSame |
| Power | 300 W | 300 WSame | 300 WSame | 250 W17% lower |
| Interface | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 4.0 x16 | PCIe 3.0 x16 |
| Released | 2020-11 | 2018-11 | 2022-03 | 2018-03 |
| 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