← All hardware
AMD's CDNA 4 flagship with the biggest memory and native FP4/FP6.
Pros
- Largest memory capacity per GPU (288 GB)
- Open ROCm software, less vendor lock-in than CUDA
- Excellent FP64 for HPC plus FP4/FP6 for AI
Cons
- ROCm ecosystem still less mature than CUDA
- 1,400W TDP requires liquid cooling
- Smaller multi-node scale-up fabric than NVLink
✓ Where it shines / best for
- Memory-bound large-model inference
- Mixed AI + HPC workloads needing FP64
- Buyers seeking an NVIDIA alternative
✕ Not the best fit for
- Teams dependent on CUDA-only libraries
- Air-cooled facilities
- Very large NVLink-style single-domain clusters
Features
- ✓ 288 GB HBM3e per GPU with 8 TB/s bandwidth
- ✓ 256 Compute Units on 4th-gen CDNA architecture
- ✓ Native FP4/FP6 support, ~20.1 PFLOPS each
- ✓ 10.1 PFLOPS FP8, 5 PFLOPS FP16, 78.6 TFLOPS FP64
- ✓ 8 CDNA 4 compute dies over Infinity Fabric
- ✓ Up to 1,400W peak board power (liquid-cooled)
- ✓ ROCm open software stack
- ✓ Strong HPC double-precision performance
Pricing
| Plan | Price | Billing | Notes |
|---|---|---|---|
| MI355X GPU (purchase) | Contact AMD/OEM | one-time | No public list price; sold via OEM platforms (Dell, HPE, Supermicro, Lenovo, GIGABYTE) and Oracle Cloud. GA since Oct 2025. |
| Cloud rental | $2.29-$8.60 | per GPU-hour | Across providers (e.g. Vultr): from ~$2.29/hr (36-mo reservation), on-demand ~$2.95-$8.60/hr, average ~$5.45/hr |
Pricing verified from the official source. Prices change often — confirm on the vendor's site before buying.
Specifications
| power | up to 1,400W peak board power |
| memory | 288 GB HBM3e, 8 TB/s |
| architecture | CDNA 4 (4th-gen) |
| compute_units | 256 |
| fp8_performance | 10.1 PFLOPS |
| fp16_performance | 5 PFLOPS |
| fp64_performance | 78.6 TFLOPS |
| fp4_fp6_performance | ~20.1 PFLOPS |
Sponsored
A full review is being generated for this product and will appear here shortly.