Shello
20 m
Which shell fits?
A model has to fit in GPU memory. Pick one, and Shello moves into the cheapest live shell that holds it.
Memory first, then price
Llama 3.3 70B Instruct needs about 43 GB of GPU memory. The cheapest GPU free now that holds it: 2 x RTX 3090 at $0.384 an hour.
GPU memory is called VRAMThe GPU's own memory. The whole model has to fit in it, or it will not run. More in the glossary. At int4, the memory splits like this:
weights 35.3 GBKV cache 2.7 GBruntime 4.5 GB
KV cache: 2 x 80 layers x 8 KV heads x 128 head dim x 2 bytes (fp16 cache). Runtime: 10% of the weights plus 1 GB.
Cheapest shell free right now: 2 x RTX 3090 on Nosana (NVIDIA 3090 (premium)) at $0.384/h, 2 cards of 24 GB. .
Show every shell that fits (for developers)
- shell
- the GPU, or a number of them working together
- $ / hour
- for all of those GPUs together
- read
- UTC time of the live read, or the day a cloud list price was read
| shell | where | $ / hour | read |
|---|---|---|---|
| 2 x RTX 3090 | Nosana: NVIDIA 3090 (premium) | $0.384 | 06:01:53 |
| 2 x V100 | Akash: nvidia v100 32Gi PCIe (median of listings, lowest $0.2) | $0.400 | 06:02:01 |
| 3 x T4 | Akash: nvidia t4 16Gi PCIe (median of listings, lowest $0.11) | $0.450 | 06:02:01 |
| RTX 6000 Ada | Nosana: NVIDIA 6000 Ada (premium) | $0.700 | 06:01:53 |
| A100 80GB | Akash: nvidia a100 80Gi SXM4 (median of listings, lowest $1.07) | $1.49 | 06:02:01 |
| RTX PRO 6000 | Akash: nvidia pro6000se 96Gi PCIe (median of listings, lowest $1.52) | $2.07 | 06:02:01 |
| RTX PRO 6000 | Nosana: NVIDIA Pro 6000 Server Edition (CC) (premium) | $2.34 | 06:01:53 |
| H100 | Akash: nvidia h100 80Gi SXM5 (median of listings, lowest $2.55) | $2.56 | 06:02:01 |
| RTX A6000 | cloud list: RunPod (Community Cloud), not stated (pods across 30+ regions) | $0.330 | 2026-11-09 |
| 2 x RTX 3090 | cloud list: RunPod (Community Cloud), not stated (pods across 30+ regions) | $0.440 | 2026-11-09 |
| RTX A6000 | cloud list: RunPod (Secure Cloud), not stated (pods across 30+ regions) | $0.590 | 2026-11-09 |
Model shapes come from each model's config.json on Hugging Face (read 2026-11-09). This estimates memory, not speed: whether a shell serves your traffic fast enough depends on throughput, which Shello does not measure.