Current Kimi workload: the Kimi K2.5 hardware guide compares 4×MI300X, 3×MI325X and 3×MI350X/MI355X paths against NVIDIA alternatives.
Current distributed workload: the Kimi-K2.5 VRAM guide maps its 595.178GB official checkpoint onto 3×256GB and 3×288GB AMD capacity tiers with topology caveats.
Planning MiniMax-M2.5? The official checkpoint guide compares MI325X’s 25.866GB and MI350X/MI355X’s 57.866GB file-only margins.
Planning DeepSeek-V3.2? Use the 689.483GB checkpoint hardware guide to compare MI300X, MI325X, MI350X/MI355X and NVIDIA alternatives with exact file-only margins.
Short answer: AMD’s current high-memory Instinct ladder moves from MI300X at 192GB HBM3 through MI325X at 256GB HBM3E to MI350X and MI355X at 288GB HBM3E. MI300X and MI325X use CDNA 3/gfx942; MI350X and MI355X move to CDNA 4/gfx950. Choose the smallest qualified ROCm platform that fits the complete workload with measured headroom. MI355X is not simply a faster-labeled MI350X: both publish the same 288GB and 8TB/s memory tier, while MI350X is a 1000W air-cooled path and MI355X is a 1400W direct-liquid-cooled path.
These are OAM-class datacenter accelerators. They belong in qualified server platforms with matching power, cooling, firmware, host resources, Infinity Fabric topology and ROCm support. A loose module is not a desktop or ordinary PCIe upgrade.
MI300X vs MI325X vs MI350X vs MI355X specifications
| GPU | Architecture / target | Memory | Published bandwidth | Maximum board power | Cooling distinction |
|---|---|---|---|---|---|
| MI300X | CDNA 3 / gfx942 | 192GB HBM3 | 5.3TB/s maximum peak theoretical | 750W | Qualified server platform |
| MI325X | CDNA 3 / gfx942 | 256GB HBM3E | 6TB/s maximum peak theoretical | 1000W | Passive OAM platform |
| MI350X | CDNA 4 / gfx950 | 288GB HBM3E | 8TB/s peak | 1000W | Air-cooled platform path |
| MI355X | CDNA 4 / gfx950 | 288GB HBM3E | 8TB/s peak | 1400W | Direct-liquid-cooled platform path |
Qwen3.5-122B-A10B: single-device file fit
| Artifact | Files | MI300X 192GB | MI325X 256GB | MI350X/MI355X 288GB |
|---|---|---|---|---|
| GPTQ 4-bit | 78.864GB | 113.136GB margin | 177.136GB margin | 209.136GB margin |
| Q4 plus projector | 77.446GB | 114.554GB margin | 178.554GB margin | 210.554GB margin |
| Official FP8 | 127.163GB | 64.837GB margin | 128.837GB margin | 160.837GB margin |
| Q8 plus projector | 130.781GB | 61.219GB margin | 125.219GB margin | 157.219GB margin |
| Official BF16 | 250.173GB | No | 5.827GB margin; unsafe boundary | 37.827GB margin |
File margin is not runtime headroom. ROCm context, cache, hybrid-attention state, temporary buffers, projector/vision processing, batch and concurrency consume memory. MI325X’s 5.827GB BF16 remainder is a file-size boundary, not a credible deployment guarantee.
Qwen3.5-397B-A17B: arithmetic device counts
| Artifact | Files | MI300X 192GB | MI325X 256GB | MI350X/MI355X 288GB |
|---|---|---|---|---|
| GPTQ 4-bit | 235.708GB | 2 devices | 1; 20.292GB file margin | 1; 52.292GB file margin |
| Q4 plus projector | 245.012GB | 2 devices | 1; 10.988GB file margin | 1; 42.988GB file margin |
| Official FP8 | 406.152GB | 3 devices | 2 devices | 2 devices |
| Q8 plus projector | 422.426GB | 3 devices | 2 devices | 2 devices |
| Official BF16 | 806.796GB | 5 devices | 4 devices | 3 devices total 864GB; 57.204GB aggregate margin |
Those counts are division, not validated deployments. Infinity Fabric does not expose several accelerator memories as one transparent device. The runtime must support tensor, pipeline, expert or other placement, and every device needs its own runtime headroom. MI325X’s 10.988GB single-device Q4/projector margin is especially tight.
Choose the tier by the boundary it removes
- MI300X 192GB: all listed 122B artifacts through Q8 fit by files, while 397B remains multi-device.
- MI325X 256GB: crosses the 397B 4-bit single-device boundary and narrowly crosses 122B BF16 by files; both boundary cases require measurement.
- MI350X 288GB: adds gfx950/CDNA 4 and safer file margins while retaining a 1000W air-cooled platform path.
- MI355X 288GB: targets the same capacity tier with a 1400W direct-liquid-cooled platform path; capacity alone does not establish a workload advantage over MI350X.
ROCm generation and kernel support matter
MI300X and MI325X use gfx942; MI350X and MI355X use gfx950. A framework or container that supports ROCm generally may still lack the exact target, compiled extension, quantization kernel, attention implementation or communication-library version required by the candidate system. Confirm the current compatibility matrix, driver/ROCm release, container, framework, model kernels and multi-device path.
Power, cooling and acquisition are system decisions
Compare complete qualified servers rather than isolated accelerator specifications. Record GPU count, Infinity Fabric topology, memory exposed per accelerator, host CPUs, system RAM, storage, networking, rack power, cooling, firmware, service, delivery and replacement terms. The 1000W-versus-1400W MI350X/MI355X distinction can determine whether a platform is feasible before benchmark speed matters.
Reproducible comparison checklist
- Identify the exact server SKU, accelerator count and memory exposed per device.
- Record gfx target, ROCm/driver, container, framework, kernels and model revision.
- Record Infinity Fabric topology, placement and any CPU-memory or storage offload.
- Hold artifact, precision, context, batch/concurrency and quality target constant.
- Measure peak per-device memory, time to first token, prompt and generation throughput, throttling and wall power.
- Compare complete acquisition, support, cooling and operating cost—not peak-specification ratios.
Detailed guides and next actions
- MI300X 192GB guide
- MI325X 256GB guide
- MI350X vs MI355X 288GB guide
- MI300X vs NVIDIA H200
- NVIDIA B300 vs AMD MI355X
- CUDA vs ROCm compatibility guide
- Build a reproducible benchmark record
- Compare large-model hardware paths
- Browse the hardware guide
- Follow the high-VRAM watchlist
Primary sources
- AMD Instinct MI300X product specifications
- AMD Instinct MI325X product specifications
- AMD ROCm MI350-series architecture reference
- AMD MI350X system-acceptance documentation
- AMD MI355X system-acceptance documentation
Capacity calculations use decimal GB and compare artifact files with nominal device capacity. They are not runtime-memory guarantees. Specifications and software support can change; verify the current qualified system and ROCm documentation before purchase.
Cross-vendor boundary: use the MI325X versus NVIDIA B200 comparison to see how the AMD 256GB tier changes one-device 397B 4-bit arithmetic relative to B200 180GB without implying a performance winner.
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.
Current distributed workload: the DeepSeek-V3.2 VRAM guide applies the exact 689.483GB official checkpoint to 768GB and larger accelerator configurations while preserving topology and offload caveats.