Short answer: choose an RTX 5090 when the complete workload fits comfortably in 32GB and you need a conventional, actively cooled desktop GPU. Choose an H100 only when 80GB or 94GB changes the workload you can run and a qualified data-center installation is justified. H100 PCIe, H100 SXM and H100 NVL are materially different products; a low-priced bare SXM module is not a PCIe card.
The decision in one minute
- RTX 5090: 32GB GDDR7, active cooling, display outputs and a desktop-card role; 575W reference total graphics power.
- H100 PCIe: 80GB HBM3, passive server cooling and 350W maximum power in NVIDIA’s original datasheet.
- H100 SXM: 80GB HBM3, 3.35TB/s published bandwidth and up to 700W in an HGX/DGX-class platform.
- H100 NVL: 94GB HBM3 per GPU, 3.9TB/s and a dual-slot air-cooled PCIe product; confirm whether a listing is one accelerator or a qualified paired configuration.
RTX 5090 vs H100 specifications that matter
| Device | Memory | Published bandwidth | Power | Deployment |
|---|---|---|---|---|
| RTX 5090 Founders Edition | 32GB GDDR7 | 1,792 GB/s | 575W total graphics power | Active-cooled PCIe graphics card |
| H100 PCIe | 80GB HBM3 | 2TB/s in NVIDIA’s original datasheet | 350W maximum; some 300W-input configurations operate at 310W | Passive, dual-slot PCIe accelerator |
| H100 SXM | 80GB HBM3 | 3.35TB/s | Up to 700W, configurable | SXM module for HGX/DGX systems |
| H100 NVL | 94GB HBM3 per GPU | 3.9TB/s | 350–400W, configurable | Dual-slot air-cooled PCIe accelerator |
Published bandwidth is not tokens per second. RTX 5090 is a Blackwell consumer GPU; H100 is a Hopper data-center accelerator with compute capability 9.0. Runtime kernels, quantization, context, batching, CPU, PCIe topology, thermal behavior and software builds all affect application performance.
Exact Qwen3.5 model-file boundaries
These margins subtract cited artifact size from nominal device memory. They are only a file-fit gate: KV cache, runtime workspaces, graph buffers, display use, multimodal state and safety headroom still consume memory.
| Official or identified artifact | Files | RTX 5090 32GB | H100 80GB | H100 NVL 94GB |
|---|---|---|---|---|
| Qwen3.5 27B Q4 + projector | 17.668GB | 14.332GB | 62.332GB | 76.332GB |
| Qwen3.5 27B Q8 + projector | 29.523GB | 2.477GB — unsafe planning margin | 50.477GB | 64.477GB |
| Qwen3.5 27B BF16 | 55.563GB | No file fit | 24.437GB | 38.437GB |
| Qwen3.5 35B Q8 + projector | 37.802GB | No file fit | 42.198GB | 56.198GB |
| Qwen3.5 35B BF16 | 71.904GB | No file fit | 8.096GB — bounded | 22.096GB |
| Qwen3.5 122B official GPTQ-Int4 | 78.864GB | No file fit | 1.136GB — unsafe | 15.136GB |
| Qwen3.5 122B community Q4_K_M + F16 projector | 77.446GB | No file fit | 2.554GB — unsafe | 16.554GB |
| Qwen3.5 122B official FP8 | 127.163GB | No file fit | No file fit | No file fit |
What the larger H100 capacity changes
The H100 80GB tier turns the listed 27B BF16 and 35B BF16 artifacts from impossible on 32GB into file-fit candidates. It does not make the listed 122B 4-bit packages safe: 1.136GB or 2.554GB before runtime state is an unsafe purchase boundary. H100 NVL’s 94GB gives those two 4-bit files materially more room, but still does not hold the listed 122B FP8 artifact on one GPU.
PCIe, SXM and NVL cannot be collapsed into one “H100”
- PCIe: passive cooling requires directed chassis airflow. Plan management/display separately.
- SXM: the module requires a compatible HGX/DGX baseboard, cooling, power delivery and firmware. It cannot be installed in a normal PCIe slot.
- NVL: verify exact memory per GPU, card count, bridge/topology and qualified host. A paired product name does not mean every process sees one automatic 188GB pool.
- MIG: an H100 may be partitioned. Your process can see only its assigned slice rather than the headline device capacity.
Software, multi-GPU and cost safeguards
- Confirm the exact driver, CUDA toolkit, framework, quantization kernels and compiled extensions for GeForce Blackwell or Hopper.
- NVLink/NVSwitch does not automatically create one transparent memory pool. The runtime must shard or parallelize the model, and some state may be duplicated.
- Compare the complete system: chassis, airflow or liquid cooling, power, motherboard lanes, interconnect, management, support and return risk.
- For used hardware, require label photos, exact part number and form factor, current device identity, memory-error evidence, operational proof and clear return terms.
Choose the next evidence
| If this is your constraint | Start here |
|---|---|
| The complete workload fits comfortably below 32GB | RTX 5090 requirements and offers |
| You need an 80GB or 94GB Hopper accelerator | H100 PCIe vs SXM vs NVL guide |
| You need a capacity estimate first | VRAM calculator and recommender |
| You need to define a pass/fail workload | Local-AI workload test |
Continue the decision
- Qwen3.5 27B memory · 35B-A3B memory · 122B-A10B memory
- 24GB, 32GB and 48GB tiers · 48GB through 96GB tiers
- Screened offers · source-reported benchmarks · submit a reproducible benchmark
- Save a private watchlist after confirming the right capacity and form factor.
Official sources
- NVIDIA GeForce RTX 5090 specifications
- NVIDIA H100 product specifications
- NVIDIA H100 Tensor Core GPU datasheet
- NVIDIA H100 PCIe product brief
- NVIDIA MIG supported GPUs
Reviewed October 25, 2026. Specifications are manufacturer-published. HighVRAM has not independently benchmarked this pair. File arithmetic is not a runtime performance claim or fit guarantee.
Does 80GB/94GB still leave a tight boundary? Compare RTX 5090 with H200 141GB to see which listed Qwen3.5 122B files become single-device candidates.