← All hardware
256 GB of HBM3e for memory-hungry inference at lower cost than NVIDIA.
Pros
- More memory than H200/B200 for big models
- Competitive total cost of ownership
- Open-source ROCm software
Cons
- Superseded by MI355X (CDNA 4)
- ROCm ecosystem maturity gap vs CUDA
- No native FP4 (unlike MI355X / Blackwell)
✓ Where it shines / best for
- Large-model inference where memory capacity matters
- Cost-conscious NVIDIA alternative deployments
- Enterprises standardizing on AMD platforms
✕ Not the best fit for
- FP4-optimized frontier inference
- CUDA-locked software stacks
- Buyers wanting the newest AMD generation
Features
- ✓ 256 GB HBM3e memory, 6 TB/s bandwidth
- ✓ 19,456 stream processors (304 CUs) up to 2.10 GHz
- ✓ CDNA 3 dual-chiplet GPU architecture
- ✓ FP16/FP8 matrix acceleration
- ✓ Infinity Fabric interconnect
- ✓ ROCm open software stack
- ✓ 8-GPU OAM platform configurations
- ✓ 1,000W peak board power
Pricing
| Plan | Price | Billing | Notes |
|---|---|---|---|
| MI325X GPU (purchase) | Contact AMD/OEM | one-time | No public list price; sold via OEM platforms and cloud (Vultr, etc.) |
| Cloud rental | $1.95-$3.43 | per GPU-hour | From ~$1.95-$2.29/hr (current) across providers like Vultr; on-demand has fallen ~33% from ~$3.43/hr (mid-2025) |
Pricing verified from the official source. Prices change often — confirm on the vendor's site before buying.
Specifications
| clock | up to 2.10 GHz |
| power | ~1,000W peak board power |
| memory | 256 GB HBM3e, 6 TB/s |
| architecture | CDNA 3 (dual-chiplet) |
| stream_processors | 19,456 (304 CUs) |
Sponsored
A full review is being generated for this product and will appear here shortly.