RTX 5090 32GB has 32 GB of VRAM and 1,792 GB/s of memory bandwidth. It runs 12 models from the catalog on a single card — up to GPT-OSS 20B (MoE) — and serves Llama 3.3 70B at 4039 tok/s aggregate. All figures are computed by tokcalc's open-source capacity model.
| Vendor | NVIDIA |
| Category | consumer |
| VRAM | 32 GB |
| Memory bandwidth | 1,792 GB/s |
| FP16/BF16 dense compute | 105 TFLOPS |
| NVLink bandwidth | none |
| Release year | 2025 |
| Typical cloud price | $0.5/hr |
| Model | Params | Quant | VRAM needed | Aggregate tok/s | Cost / 1M tok |
|---|---|---|---|---|---|
| GTE Large (1024-dim) | 0.44B | fp16 | 7.32 GB | 15884 | $0.008744 |
| Multilingual E5 Large (1024-dim) | 0.56B | fp16 | 7.56 GB | 12480 | $0.0111 |
| BGE-M3 (embed, 1024-dim) | 0.57B | fp16 | 7.58 GB | 12304 | $0.0113 |
| SmolLM2 1.7B Instruct | 1.70B | fp16 | 16.28 GB | 4111 | $0.0338 |
| Qwen 3 4B | 4B | fp16 | 17.66 GB | 1747 | $0.0795 |
| Qwen 2.5-VL 7B (VLM) | 7B | fp16 | 17.76 GB | 998.4 | $0.1391 |
| Mistral 7B v0.3 | 7.24B | fp16 | 23.07 GB | 965.3 | $0.1439 |
| Qwen2 7B | 7.62B | fp16 | 19.00 GB | 917.2 | $0.1514 |
| Qwen 3 8B | 8B | fp16 | 25.66 GB | 873.6 | $0.1590 |
| Llama 3 8B | 8.03B | fp16 | 24.65 GB | 870.3 | $0.1596 |
| Falcon 3 10B Instruct | 10B | fp16 | 30.74 GB | 698.9 | $0.1987 |
| GPT-OSS 20B (MoE) | 20.91B | fp8 | 22.55 GB | 5824 | $0.0238 |
| GPU | Aggregate tok/s | Cost / 1M tok | $/hr |
|---|---|---|---|
| Cerebras CS-3 (WSE-3) | 45076 | — | — |
| RTX 4080 16GB | 6464 | $0.1031 | $0.3 |
| RTX A4000 16GB | 4039 | $0.1651 | $0.3 |
| RTX 5090 32GB | 4039 | $0.1376 | $0.5 |
| A100 40GB SXM4 | 3493 | $0.2863 | $0.9 |
| RTX 4090 24GB | 2272 | $0.1956 | $0.4 |
| V100 SXM2 32GB | 2028 | $0.2739 | $0.5 |
| RTX 3090 24GB | 2017 | $0.1377 | $0.25 |
| H100 SXM5 80GB | 1888 | $0.7358 | $2.5 |
| TPU v6e Trillium (32GB) | 1846 | $0.4816 | $0.8 |
| T4 16GB | 1836 | $0.2421 | $0.2 |
| A10G 24GB | 1352 | $0.2958 | $0.36 |
RTX 5090 32GB is a consumer accelerator from NVIDIA with 32 GB of VRAM and 1,792 GB/s of memory bandwidth, typically around $0.5/hr on cloud rental. It is aimed at local inference and development.
12 models in the tokcalc catalog fit on one RTX 5090 32GB, up to GPT-OSS 20B (MoE) (21B params) at fp8. Below that, throughput and cost per million tokens improve as the model gets smaller.
At fp8 with 8 concurrent users, 4× RTX 5090 32GB serves 4039 tokens/sec aggregate (84.14 tok/s per stream), with a 5.55 s time to first token. That works out to $0.1376 per 1M output tokens.
Around $0.5/hr for on-demand cloud rental (a typical 2025–2026 estimate across RunPod, Lambda, Modal, AWS, GCP and similar providers). Actual pricing varies by region, commitment, and spot availability — the Build vs Buy tab in the calculator pulls live quotes from several providers.
Memory bandwidth drives decode speed and VRAM drives capacity. RTX 5090 32GB has 1,792 GB/s of bandwidth and 32 GB of VRAM, plus 105 TFLOPS of dense FP16/BF16 compute. The comparison table on this page ranks every accelerator in the catalog on the same reference workload so you can see where it lands.
Pick your model, context length, and concurrency — the calculator computes the rest with the same formulas.