This edition: HighVRAM maps the new 288GB accelerator class across NVIDIA B300 and AMD Instinct MI350X/MI355X. The useful comparison is not “which vendor has more AI FLOPS?” It is which qualified system exposes enough usable memory, supports the exact workload and produces the better reproducible result.
Delivery note: this is a web and RSS edition. HighVRAM email signup and delivery are not active, and publication here does not mean an email was sent.
What changed
The new NVIDIA B300 guide separates the 288GB product-class headline from the 270GB individual HGX B300 figure in NVIDIA’s Blackwell Ultra datasheet. Its fit tables use 270GB so a purchasing boundary does not depend on the larger marketing number.
The AMD MI350X/MI355X guide covers their shared 288GB HBM3E and 8TB/s tier while preserving the platform difference: 1000W air-cooled MI350X versus 1400W direct-liquid-cooled MI355X.
The direct B300 versus MI355X comparison uses the same Qwen3.5 files on both sides and routes the final decision through CUDA-versus-ROCm support, complete-system constraints and measured workload behavior.
The exact boundary table
| Qwen3.5 artifact | Files | B300 at 270GB | MI355X at 288GB |
|---|---|---|---|
| 122B BF16 | 250.173GB | 19.827GB nominal margin | 37.827GB nominal margin |
| 397B GPTQ 4-bit | 235.708GB | 34.292GB nominal margin | 52.292GB nominal margin |
| 397B Q4 plus projector | 245.012GB | 24.988GB nominal margin | 42.988GB nominal margin |
| 397B BF16 | 806.796GB | 3 × 270GB leaves 3.204GB: plan 4 | 3 × 288GB leaves 57.204GB: benchmark, with 4 safer |
One useful action
If you can access B300, MI350X or MI355X hardware, run one representative inference workload and submit it through the benchmark builder. Include the exposed memory per device, exact server SKU, driver and CUDA or ROCm version, framework/container, model revision, quantization, context, concurrency, topology, placement, offload, peak device memory, throughput, latency and wall power. A paired result on the same model is especially valuable.
Continue reading
- 288GB guide-cluster announcement
- CUDA versus ROCm for local AI
- Qwen3.5 96GB-to-1TB hardware paths
- Private browser watchlist and RSS feeds
- Weekly Digest archive
Sources and limits
- NVIDIA HGX AI Factory specifications
- NVIDIA Blackwell Ultra datasheet
- AMD ROCm MI350-series reference
- AMD MI355X system-acceptance documentation
Published October 25, 2026. HighVRAM has not independently benchmarked every combination. File fit does not establish compatibility, speed, quality, reliability or value.