GPU Specifications / AMD
AMD Instinct MI250X
CDNA 2 datacenter GPU with 128 GB of HBM2e memory, 3277 GB/s of bandwidth and up to 383 TFLOPS of FP16 tensor compute.
Key Specifications
Memory
128 GB HBM2e
Memory Bandwidth
3,277 GB/s
TDP
560 W
Architecture
CDNA 2
Interconnect
Infinity Fabric 3 · 800 GB/s
Est. On-demand Price
~$2.50/h
Hourly rates are indicative on-demand estimates; actual pricing varies by provider and commitment.
Compute Performance
| Precision | Peak throughput |
|---|---|
| FP64 (double precision) | 47.9 TFLOPS |
| FP32 (single precision) | 47.9 TFLOPS |
| FP24 | 95.8 TFLOPS |
| FP16 (tensor) | 383 TFLOPS |
| INT8 (tensor) | 383 TFLOPS |
| INT4 (tensor) | 383 TFLOPS |
Tensor figures use the vendor's peak numbers (with structured sparsity where supported).
System Requirements
Recommended CPU
AMD EPYC 7763 or AMD EPYC 7A53
Max VRAM per node (8 GPUs)
1,024 GB
System RAM (min / recommended)
512 / 1024 GB
Minimum PSU
1300 W
LLMs on the Instinct MI250X
Number of Instinct MI250X GPUs needed to serve popular models at 8-bit quantization (including 20% overhead for activations and KV cache).
| Model | Params | VRAM (8-bit) | GPUs needed |
|---|---|---|---|
| GPT-5.6 Sol | 2400B | 2682 GB | 21x Instinct MI250X |
| GPT-5 Flagship | 2100B | 2347 GB | 19x Instinct MI250X |
| GPT-5.6 Luna | 1400B | 1565 GB | 13x Instinct MI250X |
| Kimi K3 (1.2T) | 1200B | 1341 GB | 11x Instinct MI250X |
| Kimi K2.6 (1T) | 1000B | 1118 GB | 9x Instinct MI250X |
| GPT-5.6 Terra | 800B | 894 GB | 7x Instinct MI250X |
| DeepSeek V4 Pro (671B) | 671B | 750 GB | 6x Instinct MI250X |
| Llama 4 Behemoth (500B) | 500B | 559 GB | 5x Instinct MI250X |
| Claude 5 Fable (480B) | 480B | 536 GB | 5x Instinct MI250X |
| GLM 5.2 (400B) | 400B | 447 GB | 4x Instinct MI250X |
| Grok 4.5 | 350B | 391 GB | 4x Instinct MI250X |
| Claude 4.8 Opus (300B) | 300B | 335 GB | 3x Instinct MI250X |
| Muse Spark 1.1 | 300B | 335 GB | 3x Instinct MI250X |
| Grok 4 | 270B | 302 GB | 3x Instinct MI250X |
| Gemini 3.1 Pro | 250B | 279 GB | 3x Instinct MI250X |
| Qwen 3.7 Max (235B) | 235B | 263 GB | 3x Instinct MI250X |
| Mistral Large 3 (200B) | 200B | 224 GB | 2x Instinct MI250X |
| Grok 3 Mini | 190B | 212 GB | 2x Instinct MI250X |
| Claude 5 Sonnet (175B) | 175B | 196 GB | 2x Instinct MI250X |
| Gemini 3.5 Flash | 150B | 168 GB | 2x Instinct MI250X |
| Gemini 2.5 Flash | 140B | 156 GB | 2x Instinct MI250X |
| Llama 4 Maverick (128B) | 128B | 143 GB | 2x Instinct MI250X |
| DeepSeek V4 Flash (120B) | 120B | 134 GB | 2x Instinct MI250X |
| Qwen 3.6 Plus (110B) | 110B | 123 GB | 1x Instinct MI250X |
| Nova Premier (80B) | 80B | 89 GB | 1x Instinct MI250X |
| Qwen 3 Coder-Next (80B) | 80B | 89 GB | 1x Instinct MI250X |
| Claude 4.5 Haiku (70B) | 70B | 78 GB | 1x Instinct MI250X |
| Llama 3.3 Instruct (70B) | 70B | 78 GB | 1x Instinct MI250X |
| Mistral Medium 3.5 (70B) | 70B | 78 GB | 1x Instinct MI250X |
| Yi 1.5 (40B) | 40B | 45 GB | 1x Instinct MI250X |
| Nova Core (34B) | 34B | 38 GB | 1x Instinct MI250X |
| DeepSeek V3.1 (32B) | 32B | 36 GB | 1x Instinct MI250X |
| Gemma 3 (27B) | 27B | 30 GB | 1x Instinct MI250X |
| Mistral Small 4 (24B) | 24B | 27 GB | 1x Instinct MI250X |
| Yi 1.5 (15B) | 15B | 17 GB | 1x Instinct MI250X |
| Phi 4 (14B) | 14B | 16 GB | 1x Instinct MI250X |
| Nova Lite (12B) | 12B | 13 GB | 1x Instinct MI250X |
| Llama 3.2 Instruct (11B) | 11B | 12 GB | 1x Instinct MI250X |
| Gemma 3 (9B) | 9B | 10 GB | 1x Instinct MI250X |
| Yi 1.5 Lite (9B) | 9B | 10 GB | 1x Instinct MI250X |
| Phi 4 Mini (7B) | 7B | 8 GB | 1x Instinct MI250X |
| Phi 3.5 (3.8B) | 3.8B | 4 GB | 1x Instinct MI250X |
Frequently Asked Questions
How much VRAM does the AMD Instinct MI250X have?
The AMD Instinct MI250X has 128 GB of HBM2e memory with 3277 GB/s of memory bandwidth.
Which LLMs can run on a single Instinct MI250X?
At 8-bit quantization, a single Instinct MI250X (128 GB) can serve models up to roughly 110B parameters, such as Qwen 3.6 Plus (110B). Larger models require multiple GPUs or more aggressive quantization.
How much does it cost to rent a AMD Instinct MI250X?
On-demand cloud pricing for the Instinct MI250X is around $2.50/hour, i.e. about $1,825/month running 24/7. Actual prices vary by provider, region, and commitment.
What are the power and system requirements of the AMD Instinct MI250X?
The Instinct MI250X has a TDP of 560W. A power supply of at least 1300W per GPU is recommended. Recommended host CPUs: AMD EPYC 7763 or AMD EPYC 7A53.
Deploy on a GPU cloud
Rent the AMD Instinct MI250X by the hour instead of buying hardware.
