This week: four new capacity-first comparisons put one RTX 5090 32GB desktop GPU beside NVIDIA A40 48GB, A100 40GB/80GB, H100 80GB/94GB and H200 141GB accelerators. The useful result is not a synthetic winner. It is a set of exact file boundaries, platform warnings and controlled next tests.
Delivery note
This is the fourth web/RSS edition of HighVRAM Weekly Digest. Email signup and delivery are not active. Follow the main RSS feed or browse the Digest archive for future editions.
The 32GB-to-141GB decision map
| Comparison | Capacity boundary | Deployment question | Read |
|---|---|---|---|
| RTX 5090 vs A40 | 32GB vs 48GB | Active desktop GPU or passive Ampere server card? | Open comparison |
| RTX 5090 vs A100 | 32GB vs 40GB/80GB | Desktop card, passive PCIe accelerator or HGX/DGX-only SXM? | Open comparison |
| RTX 5090 vs H100 | 32GB vs 80GB/94GB | PCIe, SXM or H100 NVL—and which capacity is exposed? | Open comparison |
| RTX 5090 vs H200 | 32GB vs 141GB | Qualified H200 SXM or H200 NVL server path? | Open comparison |
Boundary one: 32GB can be useful without being roomy
The cited Qwen3.5 27B Q8-plus-projector files total 29.523GB. That leaves only 2.477GB against a nominal RTX 5090 32GB label before KV cache, runtime workspaces, graph buffers, display allocations and safety headroom. Treat that as an unsafe planning boundary, not proof that a complete workload fits.
Use the Qwen3.5 27B VRAM guide to preserve the exact artifact identity and projector arithmetic.
Boundary two: 40GB is not automatically a large-model tier
The cited Qwen3.5 35B Q8-plus-projector files total 37.802GB. A nominal A100 40GB leaves 2.198GB before runtime state, so the extra eight gigabytes above RTX 5090 do not create a comfortable general-purpose boundary. A40 48GB leaves 10.198GB for the same files; the A40 comparison then adds passive airflow, CPU 8-pin power, vGPU/display and NVLink checks.
Read the Qwen3.5 35B-A3B guide before treating a file-only margin as deployable memory.
Boundary three: 80GB and 94GB still have traps
The listed Qwen3.5 122B official GPTQ artifact is 78.864GB and the community Q4_K_M-plus-projector total is 77.446GB. Against 80GB, those leave only 1.136GB and 2.554GB—unsafe before runtime allocations. H100 NVL’s 94GB improves the nominal margins to 15.136GB and 16.554GB, but that still requires validation with the intended context, concurrency and runtime.
Boundary four: 141GB opens FP8 and Q8 candidates, not every 122B artifact
H200 141GB leaves 13.837GB after the listed 127.163GB official FP8 files and 10.219GB after the listed 130.781GB community Q8-plus-projector total. Those are materially different from “does not fit,” but they remain bounded candidates. The listed 250.173GB BF16 artifact still does not fit on one H200.
Use the Qwen3.5 122B-A10B guide for provenance, multimodal-projector treatment and the full 80GB-through-320GB map.
The platform checks that survive every comparison
- Bandwidth is not tokens per second. Published memory bandwidth does not establish application speed.
- Passive is not cooling-free. A40, A100 PCIe and H100 PCIe require directed chassis airflow.
- SXM is not PCIe. Bare A100, H100 or H200 SXM modules need compatible baseboards, cooling, power delivery and firmware.
- MIG changes visible capacity. A process assigned a MIG slice may see only the slice, not the headline device memory.
- NVLink/NVSwitch is not transparent pooling. The runtime must partition the model, and some allocations may be duplicated.
- Complete-system cost matters. Include chassis, cooling, power, lanes, management, support, software and return risk.
One useful action: contribute the missing controlled pair
The current benchmark database remains a fixed 43-observation snapshot from seven sources; it does not contain a controlled RTX 5090-versus-A40/A100/H100/H200 ranking. If you can run a comparable pair, hold the artifact, runtime, context, concurrency and timing boundary constant. Record exact device/form factor, exposed memory, peak use, cooling/power mode, MIG state and interconnect.
Build a reproducible submission · inspect the current source-reported database
Continue reading and follow the market
- Compare NVIDIA H100, H200, B200 and B300 tiers
- Choose another hardware path
- Browse current screened offers
- Save a private browser watchlist or follow RSS
Primary specification sources
- NVIDIA GeForce RTX 5090 specifications
- NVIDIA A40 datasheet
- NVIDIA A100 datasheet
- NVIDIA H100 product specifications
- NVIDIA H200 product specifications
Published October 25, 2026. Specifications are manufacturer-published. File-size arithmetic is not measured peak memory, a performance claim or a complete-system recommendation. HighVRAM has not independently benchmarked every listed pair.