A conservative, transparent sizing model for LLM inference. It counts all weights, KV cache at your real concurrency, 2 GiB CUDA/runtime overhead, and a 15% safety margin — then prices the exact card count on RunPod and Vast.ai. DeepSeek V3/R1 uses its native MLA latent cache, not a naive MHA blow-up.
Estimated VRAM requirement
154.97GiB
(131.50 weights + 1.25 KV + 2.00 CUDA) × 1.15
Topology-safe best fit
4× A100 80GB SXMVast.ai · $3.96/hr
Weights15% safety
Weights131.50 GiB
KV cache1.25 GiB
CUDA / runtime2.00 GiB
15% safety reserve20.22 GiB
KV / token320.00 KiB
KV total1.25 GiB
Sequences1
Subtotal before 15%134.75 GiB
DeepSeek MLA compression is doing the heavy lifting
A naive MHA cache at this shape would need 15.25 GiB; the MLA latent cache needs 0.27 GiB — about 56.9× less. That is why 671B FP8 can fit on 8× H200 instead of a 16-GPU memory wall.
GPU fit & NVLink topology
Raw fit assumes at most 90% of card VRAM is usable by the model. Multi-GPU counts round up to 2/4/8 for tensor-parallel symmetry.
GPU
Raw fit
Recommended
Headroom
Topology
⚡
Topology check
Fixed fact: RTX 4090 and L40S have no NVLink. Multi-GPU tensor parallel on those cards all-reduces over PCIe and can lose most of the scaling. A100/H100/H200 SXM use NVLink; one NVLink domain is typically 8 GPUs, beyond which InfiniBand and pipeline parallelism matter.
RunPod vs Vast.ai ledger
Compute-only estimates for the recommended card count. Sampled 2026-10-18; not a live quote, reservation or guarantee. Storage, network, tax and idle time are extra.
Recommended setup
RunPod
Vast.ai
Lower sampled rate
2-card bundle trapSome Vast.ai marketplace offers are rentable only as a 2-GPU bundle. The ledger charges the whole bundle, so a 1-GPU requirement can cost 2× the per-GPU rate. Check num_gpus before launching.
Next step
Take this card count to a provider.
The buttons carry utm_source=gpu_calculator and your current best-fit card count. Verify the live hourly rate and the bundle size before you confirm a rental.
Affiliate disclosure (FTC 16 CFR Part 255): We may earn a commission when you sign up or rent through the RunPod or Vast.ai links on this page, at no additional cost to you. This does not change the calculated VRAM requirement or the sampled rates shown. Providers set their own prices, availability and terms; verify everything on the provider page before paying. No rental or inference is performed by this calculator.