Short answer: AMD Instinct MI350X and MI355X each provide 288GB of HBM3E and 8TB/s peak memory bandwidth on one CDNA 4 accelerator. By file size, either can hold every listed Qwen3.5-122B artifact and the 4-bit Qwen3.5-397B artifacts. The decisive difference is the platform: MI350X is a 1000W air-cooled OAM part, while MI355X is a 1400W direct-liquid-cooled OAM part. Neither is a desktop GPU.
Do not buy from capacity alone. Both accelerators require a qualified server, current ROCm support for the gfx950 target, suitable firmware, power and cooling. A model fitting in 288GB of files does not guarantee enough runtime headroom, supported kernels or useful throughput.
MI350X vs MI355X specifications
| Specification | MI350X | MI355X |
|---|---|---|
| Architecture / target | CDNA 4 / gfx950 | CDNA 4 / gfx950 |
| HBM capacity | 288GB HBM3E | 288GB HBM3E |
| Peak memory bandwidth | 8TB/s | 8TB/s |
| Compute units / stream processors | 256 / 16,384 | 256 / 16,384 |
| Matrix cores | 1,024 | 1,024 |
| Form factor / host link | OAM / PCIe 5.0 x16 | OAM / PCIe 5.0 x16 |
| Scale-up links | 7 × 153.6GB/s Infinity Fabric | 7 × 153.6GB/s Infinity Fabric |
| Board power / cooling | 1000W, air cooled | 1400W, direct liquid cooled |
| Eight-accelerator platform capacity | About 2.3TB aggregate | About 2.3TB aggregate |
Selection rule: choose MI350X when the qualified air-cooled platform and its measured performance meet the workload. Choose MI355X only when its additional compute performance justifies a direct-liquid-cooled 1400W design. Memory capacity and bandwidth are identical.
What fits on one 288GB accelerator?
Qwen3.5-122B-A10B
| Artifact | Files | Nominal margin in 288GB | Capacity result |
|---|---|---|---|
| GPTQ 4-bit | 78.864GB | 209.136GB | Fits by files |
| Q4 plus projector | 77.446GB | 210.554GB | Fits by files |
| Official FP8 | 127.163GB | 160.837GB | Fits by files |
| Q8 plus projector | 130.781GB | 157.219GB | Fits by files |
| Official BF16 | 250.173GB | 37.827GB | Fits by files; benchmark peak use |
The BF16 checkpoint has far more nominal headroom here than on a 256GB MI325X, but 37.827GB still has to cover runtime allocations, cache, temporary buffers and implementation overhead. Treat the table as a file-capacity screen, then measure the intended context and concurrency.
Qwen3.5-397B-A17B
| Artifact | Files | One 288GB? | Arithmetic minimum |
|---|---|---|---|
| GPTQ 4-bit | 235.708GB | Yes; 52.292GB nominal margin | 1 |
| Q4 plus projector | 245.012GB | Yes; 42.988GB nominal margin | 1 |
| Official FP8 | 406.152GB | No | 2 × 288GB |
| Q8 plus projector | 422.426GB | No | 2 × 288GB |
| Official BF16 | 806.796GB | No | 3 × 288GB |
Those device counts are division results, not deployment recommendations. Three devices provide 864GB aggregate for the 806.796GB BF16 files—only 57.204GB total nominal margin before runtime overhead—and the software must place the model correctly across separate memory spaces. Four devices may be the more credible planning tier, but only a representative benchmark can settle it.
MI325X vs the MI350 series
| Characteristic | MI325X | MI350X / MI355X |
|---|---|---|
| Architecture / target | CDNA 3 / gfx942 | CDNA 4 / gfx950 |
| Memory | 256GB HBM3E | 288GB HBM3E |
| Peak bandwidth | 6TB/s | 8TB/s |
| Board power | 1000W | 1000W / 1400W |
| Eight-device aggregate | 2.048TB | About 2.3TB |
The extra 32GB changes some boundaries materially, especially BF16 Qwen3.5-122B and 4-bit Qwen3.5-397B. But CDNA 4 is also a software transition: confirm the exact ROCm release, framework, container and kernels for gfx950. Do not assume a binary or optimization built for gfx942 works unchanged.
Precision support is not model support
CDNA 4 adds native low-precision formats including MXFP8, MXFP6 and MXFP4, alongside OCP FP8 support. That does not mean an existing model artifact, inference engine or kernel automatically uses those formats. Verify the complete path from checkpoint format through conversion, kernels and output quality.
A multi-accelerator platform is not one memory pool
An eight-accelerator MI350-series platform provides roughly 2.3TB of aggregate HBM. Each accelerator still owns its memory. The runtime must shard or place weights, experts, cache and activations, while topology and communication affect throughput. Report per-device peak memory, model revision, quantization, context, batch/concurrency, software versions and topology with every result.
Plan the next step
- Use the reproducible benchmark methodology
- Compare the 256GB MI325X boundary
- Review the 192GB MI300X path
- Compare MI300X with NVIDIA H200
- See the Qwen3.5-122B capacity tiers
- See the Qwen3.5-397B capacity tiers
- Compare the full Qwen hardware paths
- Match hardware to your workload
- Check the live hardware market
Primary sources
- AMD ROCm MI350-series architecture reference
- AMD Instinct MI350X system-acceptance documentation
- AMD Instinct MI355X system-acceptance documentation
- AMD ROCm workload optimization documentation
Capacity calculations use decimal GB to match the published file totals and product capacities. Specifications can change; verify the current qualified system and software documentation before purchase.
For the CUDA alternative in the same headline capacity class, compare the NVIDIA B300 288GB guide, including its documented usable-memory caveat.
For the direct CUDA comparison, see NVIDIA B300 vs AMD MI355X with conservative usable-memory arithmetic.
Compare the preceding AMD tiers too: use the MI300X/MI325X/MI350X/MI355X decision hub.
Choosing an AMD path across consumer, workstation and data-center hardware? Use the Best AMD GPU for Local AI guide to compare 24GB through 288GB capacity, physical deployment and ROCm validation gates.