L4 24GB has 24 GB of VRAM and 300 GB/s of memory bandwidth. It runs 9 models from the catalog on a single card — up to GPT-OSS 20B (MoE) — and serves Llama 3.3 70B at 676.1 tok/s aggregate. All figures are computed by tokcalc's open-source capacity model.
| Vendor | NVIDIA |
| Category | workstation |
| VRAM | 24 GB |
| Memory bandwidth | 300 GB/s |
| FP16/BF16 dense compute | 30.3 TFLOPS |
| NVLink bandwidth | none |
| Release year | 2023 |
| Typical cloud price | $0.4/hr |
| Model | Params | Quant | VRAM needed | Aggregate tok/s | Cost / 1M tok |
|---|---|---|---|---|---|
| GTE Large (1024-dim) | 0.44B | fp16 | 7.32 GB | 2659 | $0.0418 |
| Multilingual E5 Large (1024-dim) | 0.56B | fp16 | 7.56 GB | 2089 | $0.0532 |
| BGE-M3 (embed, 1024-dim) | 0.57B | fp16 | 7.58 GB | 2060 | $0.0539 |
| SmolLM2 1.7B Instruct | 1.70B | fp16 | 16.28 GB | 688.2 | $0.1614 |
| Qwen 3 4B | 4B | fp16 | 17.66 GB | 292.5 | $0.3799 |
| Qwen 2.5-VL 7B (VLM) | 7B | fp16 | 17.76 GB | 167.1 | $0.6648 |
| Mistral 7B v0.3 | 7.24B | fp16 | 23.07 GB | 161.6 | $0.6876 |
| Qwen2 7B | 7.62B | fp16 | 19.00 GB | 153.5 | $0.7236 |
| GPT-OSS 20B (MoE) | 20.91B | fp8 | 22.55 GB | 975.0 | $0.1140 |
| 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 |
L4 24GB is a workstation accelerator from NVIDIA with 24 GB of VRAM and 300 GB/s of memory bandwidth, typically around $0.4/hr on cloud rental. It is aimed at on-prem and workstation inference.
9 models in the tokcalc catalog fit on one L4 24GB, 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× L4 24GB serves 676.1 tokens/sec aggregate (14.09 tok/s per stream), with a 19.19 s time to first token. That works out to $0.6573 per 1M output tokens.
Around $0.4/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. L4 24GB has 300 GB/s of bandwidth and 24 GB of VRAM, plus 30.3 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.