AMD's next-gen CDNA 4 competitor to Nvidia Blackwell.
Weekly median price per GPU per hour. AMD MI355X price history (JSON).
Of 4 listings across 4 cloud providers, none are verified in stock. On-demand is listed from $8.60 per GPU per hour and spot from $4.50.
USA
1 config
8x
USA
1 config
8x
USA
1 config
8x
USA
1 config
8x
No providers match .
Heads up: Prices are estimates from published rates for common setups and vary by region and usage; verify with the provider before provisioning.
One MI355X has 288 GB of VRAM. In practice, that's enough memory for roughly 460B parameters at 4-bit or 125B at 16-bit, assuming a 32K context. Below are some open-weight LLMs, with the estimated memory and GPUs each one needs.
| Model | Parameters | 4-bitINT4 / FP4 | 8-bitFP8 / INT8 | 16-bitFP16 / BF16 |
|---|---|---|---|---|
Qwen3.8-27B
|
27B
|
8 GPUs
18 GB VRAM
|
8 GPUs
30 GB VRAM
|
8 GPUs
58 GB VRAM
|
Gemma 4 31B
|
30.7B
|
8 GPUs
22 GB VRAM
|
8 GPUs
35 GB VRAM
|
8 GPUs
69 GB VRAM
|
GPT-OSS-120B
|
117B
5.1B active
|
8 GPUs
67 GB VRAM
|
8 GPUs
120 GB VRAM
|
8 GPUs
237 GB VRAM
|
DeepSeek V4 Flash
|
284B
13B active
|
8 GPUs
160 GB VRAM
|
8 GPUs
287 GB VRAM
|
8 GPUs
573 GB VRAM
|
GLM-5.3-Flash
|
320B
18B active
|
8 GPUs
178 GB VRAM
|
8 GPUs
322 GB VRAM
|
8 GPUs
642 GB VRAM
|
MiniMax-M3
|
428B
23B active
|
8 GPUs
239 GB VRAM
|
8 GPUs
432 GB VRAM
|
8 GPUs
862 GB VRAM
|
Kimi K3
|
2.8T
104B active
|
8 GPUs
1,544 GB VRAM
|
11 GPUs
2,804 GB VRAM
|
22 GPUs
5,605 GB VRAM
|
Estimates based on the median on-demand rate. Memory is weights plus FP8 KV cache at 32K context per request (FP16/BF16 in the 16-bit column). GPU counts assume 90% of advertised VRAM is usable. No guarantee of runtime support, usable performance, or that a matching quantized build exists. How we estimate costs.
AMD MI355X · Per GPU
| Compute · dense | |
|---|---|
| FP4 | 10,066.3 TFLOPS |
| FP8 | 5,033.2 TFLOPS 10,066.4 with sparsity |
| FP16 / BF16 | 2,516.6 TFLOPS 5,033.2 with sparsity |
| INT8 | 5,033.2 TOPS 10,066.4 with sparsity |
| FP32 | 157.3 TFLOPS |
| FP64 | 78.6 TFLOPS |
| Precision support | FP4FP6FP8FP16BF16FP32FP64INT8 |
| Memory | |
|---|---|
| Capacity | 288 GB HBM3e |
| Bandwidth | 8,000 GB/s |
| Bus width | 8,192-bit |
| ECC | Yes |
| Silicon | |
|---|---|
| Architecture | CDNA 4 |
| Process | TSMC 3nm / 6nm FinFET |
| Transistors | 185 billion |
| Shader cores | 16,384 Stream processors |
| Matrix cores | 1,024 |
| Compute units | 256 |
| Fabric and host | |
|---|---|
| GPU interconnect | Infinity Fabric 1,075 GB/s |
| Host interface | PCIe 5.0 x16 |
| Power | |
|---|---|
| Board power | 1,400 W |
| Platform | |
|---|---|
| Partitioning | Up to 8 partitions |
Source: official AMD MI355X datasheet.
288GB HBM3e with FP4 support on CDNA 4 architecture. Next-generation AMD compute for both training and inference. Competitive with Nvidia Blackwell on memory capacity.
Latest generation with minimal cloud availability. ROCm software stack, while rapidly improving, may lag behind CUDA for some workloads. Evaluate software compatibility before committing.
The MI355X is listed by 4 providers.
The MI355X supports 8 precision formats. Training: BF16, FP16, FP32. Inference: FP4, FP6, FP8, INT8. Scientific: FP64.
Yes. 1,075 GB/s bidirectional.
As of August 30, 2026, we track 4 listings from 4 providers. Prices are per GPU per hour.
| Billing type | Listings | Median / GPU / hr | Cheapest in stock |
|---|---|---|---|
| On-demand | 1 | $8.60 | |
| Spot | 2 | $4.50 | |
| Custom contract | 1 | On request |
No median for on-demand and spot: we only show this when at least 3 providers list the GPU on that billing type.
Last updated
