HighVRAM has published a new 288GB accelerator planning cluster covering NVIDIA B300, AMD Instinct MI350X and MI355X, and a direct B300 versus MI355X comparison.
The central finding: “288GB” is not always the same usable-memory claim. NVIDIA markets B300 in the 288GB HBM3E class, while its Blackwell Ultra datasheet lists 270GB for an individual HGX B300 GPU. AMD publishes 288GB for MI350X and MI355X. The new guides preserve that distinction in every model-fit calculation.
What the new guides cover
| Path | Software | Conservative capacity used | Platform distinction |
|---|---|---|---|
| NVIDIA B300 | CUDA / Blackwell Ultra | 270GB from HGX datasheet | SXM-class qualified HGX/DGX system |
| AMD MI350X | ROCm / CDNA 4 | 288GB | 1000W air-cooled OAM platform |
| AMD MI355X | ROCm / CDNA 4 | 288GB | 1400W direct-liquid-cooled OAM platform |
The Qwen3.5 boundary that matters
Qwen3.5-122B-A10B’s 250.173GB official BF16 files fit by arithmetic on both paths, but leave only 19.827GB against 270GB B300 planning capacity and 37.827GB against 288GB MI355X capacity. Neither is a runtime guarantee.
For Qwen3.5-397B-A17B, both listed 4-bit artifacts fit on one accelerator by file size. The 806.796GB BF16 files expose the more important multi-device trap: three 270GB B300 devices leave only 3.204GB aggregate before runtime overhead, while three 288GB MI355X devices leave 57.204GB and still require measured peak-memory evidence.
Use the same workload, not vendor peak claims
The comparison does not rank CUDA and ROCm from theoretical tensor figures. A useful test holds the checkpoint, precision, context, batch/concurrency, runtime configuration and quality target constant, then reports topology, placement, offload, peak device memory, throughput, latency, throttling and wall power.
Operators can capture those fields with the local-AI benchmark builder. Readers tracking hardware availability can use the private browser watchlist and RSS feeds; HighVRAM does not claim that the watchlist sends price alerts.
Continue reading
Sources and limits
Published October 25, 2026. File fit does not establish compatibility, speed, reliability or value. Verify the exact supported system and reproduce the intended workload before purchase.