A100 80GB SXM4 has 80 GB of VRAM and 2,000 GB/s of memory bandwidth, plus 600 GB/s of NVLink. It runs 28 models from the catalog on a single card — up to GPT-OSS 120B (MoE) — and serves Llama 3.3 70B at 1127 tok/s aggregate. All figures are computed by tokcalc's open-source capacity model.
| Vendor | NVIDIA |
| Category | datacenter |
| VRAM | 80 GB |
| Memory bandwidth | 2,000 GB/s |
| FP16/BF16 dense compute | 312 TFLOPS |
| NVLink bandwidth | 600 GB/s |
| Release year | 2020 |
| Typical cloud price | $1.2/hr |
| Model | Params | Quant | VRAM needed | Aggregate tok/s | Cost / 1M tok |
|---|---|---|---|---|---|
| GTE Large (1024-dim) | 0.44B | fp16 | 7.32 GB | 17727 | $0.0188 |
| Multilingual E5 Large (1024-dim) | 0.56B | fp16 | 7.56 GB | 13929 | $0.0239 |
| BGE-M3 (embed, 1024-dim) | 0.57B | fp16 | 7.58 GB | 13732 | $0.0243 |
| SmolLM2 1.7B Instruct | 1.70B | fp16 | 16.28 GB | 4588 | $0.0726 |
| Phi-3 Mini 3.8B | 3.80B | fp16 | 33.37 GB | 2053 | $0.1624 |
| Qwen 3 4B | 4B | fp16 | 17.66 GB | 1950 | $0.1709 |
| Llama 2 7B | 6.74B | fp16 | 47.84 GB | 1157 | $0.2880 |
| Qwen 2.5-VL 7B (VLM) | 7B | fp16 | 17.76 GB | 1114 | $0.2991 |
| Mistral 7B v0.3 | 7.24B | fp16 | 23.07 GB | 1077 | $0.3094 |
| Qwen2 7B | 7.62B | fp16 | 19.00 GB | 1024 | $0.3256 |
| Qwen 3 8B | 8B | fp16 | 25.66 GB | 975.0 | $0.3419 |
| Llama 3 8B | 8.03B | fp16 | 24.65 GB | 971.4 | $0.3432 |
| Gemma 2 9B | 9B | fp16 | 40.55 GB | 866.7 | $0.3846 |
| Falcon 3 10B Instruct | 10B | fp16 | 30.74 GB | 780.0 | $0.4274 |
| Pixtral 12B (VLM) | 12B | fp16 | 34.74 GB | 650.0 | $0.5128 |
| OLMo 2 13B | 13B | fp16 | 36.74 GB | 600.0 | $0.5556 |
| Phi-4 14B | 14B | fp16 | 38.74 GB | 557.1 | $0.5983 |
| Qwen 2.5 14B | 14B | fp16 | 40.88 GB | 557.1 | $0.5983 |
| Qwen 3 14B | 14B | fp16 | 38.74 GB | 557.1 | $0.5983 |
| GPT-OSS 20B (MoE) | 20.91B | fp8 | 22.55 GB | 6500 | $0.0513 |
| Codestral 25.08 | 22B | fp8 | 37.03 GB | 1064 | $0.3134 |
| Gemma 2 27B | 27B | fp8 | 51.70 GB | 866.7 | $0.3846 |
| Qwen3.8 27B (VLM) | 27.78B | fp8 | 33.28 GB | 842.3 | $0.3957 |
| Qwen 3 30B-A3B (MoE) | 30B | fp8 | 36.44 GB | 7800 | $0.0427 |
| Qwen 3 32B | 32B | fp8 | 49.18 GB | 731.3 | $0.4558 |
| Mixtral 8x7B (MoE) | 46.70B | fp8 | 55.29 GB | 1814 | $0.1838 |
| Llama 4 Scout 17B-16E (MoE, VLM) | 109B | awq4 | 72.83 GB | 1468 | $0.2270 |
| GPT-OSS 120B (MoE) | 116.83B | awq4 | 66.71 GB | 4894 | $0.0681 |
| 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 |
A100 80GB SXM4 is a datacenter accelerator from NVIDIA with 80 GB of VRAM and 2,000 GB/s of memory bandwidth and 600 GB/s of NVLink, typically around $1.2/hr on cloud rental. It is aimed at production LLM serving.
28 models in the tokcalc catalog fit on one A100 80GB SXM4, up to GPT-OSS 120B (MoE) (117B params) at awq4. Below that, throughput and cost per million tokens improve as the model gets smaller.
At fp8 with 8 concurrent users, 2× A100 80GB SXM4 serves 1127 tokens/sec aggregate (46.95 tok/s per stream), with a 3.76 s time to first token. That works out to $0.5916 per 1M output tokens.
Around $1.2/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. A100 80GB SXM4 has 2,000 GB/s of bandwidth and 80 GB of VRAM, plus 312 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.