Codestral 25.08 at FP8 (E4M3) needs 22.0 GB for its weights and 37.0 GB in total at 8K context with 8 concurrent users. That means it needs 2× RTX 5090 32GB. Every figure is computed by tokcalc's open-source capacity model.
RTX 5090 32GB has no NVLink, so 2-way tensor parallelism runs over PCIe. tokcalc applies a flat 0.85× multi-GPU efficiency factor for this, which is realistic for NVLink-equipped datacenter cards but optimistic for PCIe scale-out — treat the throughput and cost figures for this configuration as an upper bound, and measure before committing.
| GPUs | Total VRAM | VRAM needed | Decode tok/s | Aggregate tok/s | TTFT | Cost / 1M tok | Fits? |
|---|---|---|---|---|---|---|---|
| 1× RTX | 32 GB | 37.0 GB | 0.000 | 0.000 | — | — | ✗ |
| 2× RTX | 64 GB | 37.0 GB | 135.0 | 3240 | 3.47 s | $0.0857 | ✓ |
| 4× RTX | 128 GB | 37.0 GB | 270.0 | 12961 | 1.75 s | $0.0429 | ✓ |
| 8× RTX | 256 GB | 37.0 GB | 540.0 | 51844 | 890.1 ms | $0.0214 | ✓ |
| Context | Concurrent users | KV per request | Fits? |
|---|---|---|---|
| 4K | 44 | 962.1 MB | ✓ fits |
| 8K | 22 | 1.88 GB | ✓ fits |
| 33K | 5 | 7.52 GB | ✓ fits |
| 66K | 2 | 15.03 GB | ✓ fits |
| 131K | 1 | 30.06 GB | ✓ fits |
| GPU | GPUs | Aggregate tok/s | Cost / 1M tok |
|---|---|---|---|
| AMD MI300X 192GB | 1× | 2819 | $0.2464 |
| A100 40GB SXM4 | 1× | 824.3 | $0.3033 |
| AMD MI325X 288GB | 1× | 3191 | $0.3047 |
| A100 80GB SXM4 | 1× | 1064 | $0.3134 |
| Intel Gaudi 3 128GB | 1× | 1755 | $0.3166 |
| H100 SXM5 80GB | 1× | 1782 | $0.3898 |
Yes, but it needs 2× RTX 5090 32GB. At FP8 (E4M3) the total footprint is 37.0 GB (22.0 GB of weights plus 15.0 GB of KV cache at 8K × 8 concurrent users), against 64 GB of available VRAM across 2 GPUs.
135.0 tokens/sec per stream and 3240 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 3.47 s, and inter-token latency is 7.4 ms.
$0.0857 per 1M output tokens at 2× RTX 5090 32GB ($0.5000/hr). That is the marginal hardware cost — add idle capacity, ops, and egress for the real figure.
22.0 GB for the weights at FP8 (E4M3), plus 15.0 GB of KV cache for 8 concurrent requests at 8K context — 37.0 GB total. Every extra concurrent user at long context costs more, so leave headroom.
It works. Compare $0.0857 per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 4× RTX 3080 10GB at $0.0404, and the fastest is 1× Cerebras CS-3 (WSE-3) at 10636 tok/s.
Same formulas, your context length, concurrency, and quantisation.