Short answer: for Qwen3.5-122B-A10B, a 96GB accelerator is the practical 4-bit starting point, 192GB is the cleaner single-device FP8/Q8 capacity, and roughly 320GB total remains the practical BF16 class. A 256GB MI325X holds the BF16 files with only 5.827GB nominal margin; a B300 planned at its documented 270GB HGX figure leaves 19.827GB; a 288GB MI350X or MI355X leaves 37.827GB. For Qwen3.5-397B-A17B, all three can hold the listed 4-bit files by file size, while practical planning remains roughly 320GB for 4-bit, 512GB for FP8/Q8 and 1TB for BF16. Those totals do not guarantee pooling, expert placement or acceptable speed.
This is a capacity map, not a universal “best GPU” ranking. Confirm exact Qwen3.5 support, quantization kernels, multimodal path, per-device placement, interconnect and a representative benchmark before buying. Unified system memory is not dedicated VRAM.
Official hardware capacity reference
| Hardware path | Published memory | Published bandwidth | Form and power context |
|---|---|---|---|
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 96GB GDDR7 ECC | 1,792GB/s | PCIe 5.0 x16; 600W board power |
| NVIDIA H200 | 141GB HBM3e | 4.8TB/s | SXM or PCIe; up to 700W/600W |
| AMD Instinct MI300X | 192GB HBM3 | 5.3TB/s peak theoretical | OAM; 750W maximum TBP |
| AMD Instinct MI325X | 256GB HBM3E | 6TB/s peak theoretical | OAM; 1000W maximum TBP |
| NVIDIA B300 | 288GB HBM3E tier; 270GB in HGX datasheet | Up to 8TB/s; 7.7TB/s in HGX datasheet | SXM-class HGX/DGX; up to 1100W |
| AMD Instinct MI350X | 288GB HBM3E | 8TB/s | OAM; 1000W air cooled |
| AMD Instinct MI355X | 288GB HBM3E | 8TB/s | OAM; 1400W direct liquid cooled |
| Apple Mac Studio with M5 Ultra | Up to 512GB unified memory | 1.2TB/s | Shared by CPU/GPU/system; not dedicated VRAM |
| NVIDIA HGX H200, 8 GPUs | 1,128GB aggregate HBM3e | 38.4TB/s aggregate local bandwidth | NVLink/NVSwitch platform class |
| AMD MI300X platform, 8 accelerators | 1.5TB aggregate HBM3 | 5.3TB/s per OAM | Fully meshed Infinity Fabric platform |
| AMD MI325X platform, 8 accelerators | 2.048TB aggregate HBM3E | 6TB/s per OAM | UBB 2.0 Infinity Fabric platform |
| NVIDIA HGX/DGX B300, 8 GPUs | 2.3TB marketed aggregate HBM3E | Platform aggregate; verify usable per GPU | Fifth-generation NVLink/NVSwitch |
| AMD MI350-series platform, 8 accelerators | About 2.3TB aggregate HBM3E | 8TB/s per OAM | Infinity Fabric platform |
Qwen3.5-122B-A10B hardware paths
| Artifact path | Exact files | Capacity examples | Interpretation |
|---|---|---|---|
| Official GPTQ-Int4 | 78.864GB | 1×96GB RTX PRO; 1×141GB H200; 1×192GB MI300X; 1×256GB MI325X | 96GB is the practical starting tier; larger devices add workload headroom |
| Official FP8 | 127.163GB | 1×141GB H200; 1×192GB MI300X; 1×256GB MI325X; 2×96GB | 141GB holds files but is tight; 192GB is the cleaner single-device capacity |
| Official BF16 | 250.173GB | 1×256GB MI325X; 1×B300 planned at 270GB; 1×288GB MI350X/MI355X; 3×96GB = 288GB; 2×192GB = 384GB | Nominal file margins are 5.827GB, 19.827GB and 37.827GB respectively; all require peak-runtime measurement, while 384GB clears the 320GB practical class |
Qwen3.5-397B-A17B hardware paths
| Artifact path | Exact files | Capacity examples | Interpretation |
|---|---|---|---|
| Official GPTQ-Int4 | 235.708GB | 1×256GB MI325X; 1×B300 planned at 270GB; 1×288GB MI350X/MI355X; 3×96GB = 288GB; 2×192GB = 384GB | Single-device nominal margins are 20.292GB, 34.292GB and 52.292GB; 320GB remains the practical planning class |
| Official FP8 | 406.152GB | 2×256GB MI325X = 512GB; 2×270GB B300 = 540GB; 2×288GB MI350X/MI355X = 576GB; 5×96GB = 480GB; 512GB unified | 480GB is tight; 512GB, 540GB or 576GB provide more aggregate capacity, subject to placement, memory type and runtime |
| Official BF16 | 806.796GB | 4×256GB MI325X = 1,024GB; 3×270GB B300 = 810GB; 3×288GB MI350X/MI355X = 864GB; 5×192GB = 960GB; 8-device platforms exceed 2TB | Three B300s leave only 3.204GB and are a file-size trap; three 288GB devices leave 57.204GB but still require measurement; roughly 1TB remains the practical tier |
Workstation, unified-memory or datacenter?
- 96GB PCIe workstation: a simple single-device 122B 4-bit capacity path. For larger artifacts, multiple cards require runtime support, adequate slots, power, cooling and interconnect-aware placement.
- 141GB to 288GB datacenter accelerator: higher single-device capacity and HBM bandwidth. H200 can hold 122B FP8 files tightly; MI300X adds headroom; MI325X reaches tight 122B BF16 and 397B 4-bit boundaries; B300 and MI350-series devices add more file margin. Validate exposed memory, CUDA or ROCm support and measured peak allocation.
- 512GB unified-memory Mac Studio: enough published shared capacity for 397B FP8/Q8 files in principle, but the operating system and applications share it. Confirm the runtime, format, maximum usable allocation and measured workload; do not label it 512GB VRAM.
- Eight-accelerator platform: the clearest all-accelerator BF16 capacity class for 397B, but expert placement, parallelism, networking, power and acquisition/hosting cost dominate the decision.
- CPU or storage offload: can make a file load without creating dedicated accelerator memory. Report offload and transfer behavior separately.
Pre-purchase validation checklist
- Choose the exact model repository, revision, artifact and projector.
- Verify the runtime/version supports Qwen3.5 MoE, hybrid attention, MTP, quantization and image/video input on the target hardware.
- Define tensor, pipeline and expert parallelism plus every layer/expert/projector assignment.
- Check local per-device fit; do not rely on summed capacity alone.
- Record interconnect, host RAM, storage, power, cooling, slots and any offload.
- Fix context allocation, visual workload, batch size, output length and concurrency.
- Measure peak per-device memory, system memory, time to first token, throughput, fallback and errors.
Read the 122B-A10B exact guide, 397B-A17B exact guide, complete Qwen3.5 comparison, multi-GPU memory guide, VRAM calculator, workload test or benchmark submission.
Manufacturer sources and limits
- NVIDIA RTX PRO 6000 Workstation Edition product brief
- NVIDIA H200 specifications
- NVIDIA HGX H200 platform specifications
- AMD Instinct MI300X specifications
- AMD MI300X platform documentation
- AMD Instinct MI325X specifications
- AMD MI325X platform specifications
- Apple Mac Studio M5 Ultra technical specifications
- NVIDIA HGX B300 specifications
- NVIDIA Blackwell Ultra datasheet
- AMD MI350-series architecture reference
- HighVRAM methodology
Reviewed October 25, 2026. Hardware specifications are manufacturer-published. HighVRAM has not independently benchmarked every listed topology, runtime, quantization or visual workload. Capacity fit does not establish compatibility, speed, output quality, reliability or value. Prices and availability are intentionally excluded because they change.
Direct 180GB-versus-256GB decision: the MI325X versus B200 guide applies these Qwen3.5 totals to ROCm/gfx942 and CUDA/Blackwell complete-system choices.
Track matching hardware
Save a private browser-only hardware watchlist or choose a focused RSS feed. No account or email address is required; the watchlist stays in this browser and does not send price alerts.
Dedicated 192GB path: the AMD Instinct MI300X local-AI guide covers official specifications, ROCm compatibility checks and exact single-/multi-device Qwen3.5 fit.
Current 512GB unified-memory path: the Mac Studio M5 Ultra local-AI guide maps 96GB, 256GB and 512GB configurations to exact Qwen3.5 artifacts with shared-memory safeguards.
Choosing between the 96GB CUDA workstation and the 512GB shared-memory path? Use the M5 Ultra vs RTX PRO 6000 Blackwell comparison.
\n\nDatacenter accelerator decision: use the MI300X vs H200 comparison to separate 192GB and 141GB per-device capacity from software and topology requirements.
\n\n256GB AMD path: use the MI325X guide to evaluate exact Qwen3.5-122B and 397B artifact boundaries, ROCm support and complete-platform requirements.
For the 288GB CDNA 4 step, use the MI350X vs MI355X model-fit guide.
The CUDA-side 288GB tier is covered in the NVIDIA B300 Blackwell Ultra guide, which uses the lower documented HGX usable capacity for fit decisions.
For a direct CUDA-versus-ROCm decision at the 288GB tier, use the B300 vs MI355X comparison.
Blackwell datacenter choice: compare NVIDIA B200 and B300 using 180GB versus conservative 270GB model-fit math.
180GB Blackwell path: evaluate the NVIDIA B200 for local AI with exact Qwen3.5 fit, multi-device arithmetic and complete-system safeguards.
80GB Hopper versus 180GB Blackwell: use the H100 vs B200 guide for exact Qwen3.5 fit and multi-device boundaries.
NVIDIA datacenter ladder: compare H100, H200, B200 and B300 using exact Qwen3.5 boundaries and complete-system safeguards.
AMD datacenter ladder: compare MI300X, MI325X, MI350X and MI355X using exact Qwen3.5 boundaries and complete-system safeguards.
Compact 128GB path: compare RTX 5090 with DGX Spark before treating unified system memory as dedicated VRAM.
Compact unified-memory decision: compare DGX Spark with Mac Studio across 96GB, 128GB, 256GB and 512GB classes.
Blackwell workstation decision: compare DGX Spark 128GB with RTX PRO 6000 96GB.
DGX Spark alternatives: Compare DGX Spark alternatives across 32GB to 512GB paths in the constraint-first comparison hub.
DGX Spark vs H200: Compare the 128GB and 141GB one-system paths in the exact 128GB-versus-141GB comparison.
For a concrete 122B capacity decision, compare DGX Spark 128GB with A100 40GB/80GB; the guide preserves file-only margins and runtime headroom.