AMD MI350P
PCIe CDNA 4 accelerator for AI inference in standard air-cooled servers.
- Arch
- CDNA 4
- Memory
- 144 GB HBM3e
- Bandwidth
- 4.00 TB/s
- Released
- Q2 2026
At a glance
We don't currently track any MI350P offers. See alternative GPUs.
Technical Specifications
AMD MI350P · Per GPU
| Compute · dense | |
|---|---|
| FP4 | 4,600 TFLOPS |
| FP8 | 2,300 TFLOPS 4,600 with sparsity |
| FP16 / BF16 | 1,150 TFLOPS 2,300 with sparsity |
| INT8 | 2,300 TOPS 4,600 with sparsity |
| FP32 | 72 TFLOPS |
| FP64 | 36 TFLOPS |
| Precision support | FP4FP6FP8FP16BF16FP32FP64INT8 |
| Memory | |
|---|---|
| Capacity | 144 GB HBM3e |
| Bandwidth | 4,000 GB/s |
| Bus width | 4,096-bit |
| ECC | Yes |
| Silicon | |
|---|---|
| Architecture | CDNA 4 |
| Process | TSMC 3nm / 6nm FinFET |
| Transistors | 73 billion |
| Shader cores | 8,192 Stream processors |
| Matrix cores | 512 |
| Compute units | 128 |
| Fabric and host | |
|---|---|
| Host interface | PCIe 5.0 x16 |
| Power | |
|---|---|
| Board power | 600 W |
| Cooling | Passive |
| Platform | |
|---|---|
| Partitioning | Up to 4 partitions |
Source: official AMD MI350P datasheet.
Frequently Asked Questions
Why choose the MI350P?
144GB HBM3e with MXFP4 and MXFP6 support on CDNA 4, in a dual-slot, passively cooled PCIe card at 600W that fits standard air-cooled servers rather than an OAM baseboard. One card holds a 70B model at FP8 with room for KV cache.
When is the MI350P not a good fit?
Half the compute units, memory and bandwidth of the OAM MI350X, so a model above roughly 120B parameters at FP8 needs more than one card. For models that fit in 96GB, the RTX PRO 6000 is a 600W PCIe card with CUDA support, and the ROCm software ecosystem, while rapidly improving, is narrower than CUDA.
Alternatives to AMD MI350P
-
-
Nvidia RTX PRO 6000
96 GB GDDR7 · 60% less bandwidth than the MI350P
On-demand from $0.65 /GPU/hr -
-
Last updated