Qwen 2.5 72B at FP8 (E4M3) needs 72.7 GB for its weights and 94.2 GB in total at 8K context with 8 concurrent users. That means it needs 4× RTX 5090 32GB. Every figure is computed by tokcalc's open-source capacity model.
RTX 5090 32GB has no NVLink, so 4-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 | 94.2 GB | 0.000 | 0.000 | — | — | ✗ |
| 2× RTX | 64 GB | 94.2 GB | 0.000 | 0.000 | — | — | ✗ |
| 4× RTX | 128 GB | 94.2 GB | 81.71 | 3922 | 5.71 s | $0.1416 | ✓ |
| 8× RTX | 256 GB | 94.2 GB | 163.4 | 15689 | 2.87 s | $0.0708 | ✓ |
| Context | Concurrent users | KV per request | Fits? |
|---|---|---|---|
| 4K | 41 | 1.34 GB | ✓ fits |
| 8K | 20 | 2.68 GB | ✓ fits |
| 33K | 5 | 10.74 GB | ✓ fits |
| 66K | 2 | 21.47 GB | ✓ fits |
| 131K | 1 | 42.95 GB | ✓ fits |
| GPU | GPUs | Aggregate tok/s | Cost / 1M tok |
|---|---|---|---|
| A100 80GB SXM4 | 2× | 1094 | $0.6092 |
| H100 SXM5 80GB | 2× | 1833 | $0.7577 |
| AMD MI300X 192GB | 1× | 853.0 | $0.8142 |
| L40S 48GB | 2× | 472.8 | $0.8226 |
| AMD MI325X 288GB | 1× | 965.6 | $1.01 |
| Intel Gaudi 3 128GB | 1× | 531.1 | $1.05 |
Yes, but it needs 4× RTX 5090 32GB. At FP8 (E4M3) the total footprint is 94.2 GB (72.7 GB of weights plus 21.5 GB of KV cache at 8K × 8 concurrent users), against 128 GB of available VRAM across 4 GPUs.
81.71 tokens/sec per stream and 3922 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 5.71 s, and inter-token latency is 12.2 ms.
$0.1416 per 1M output tokens at 4× RTX 5090 32GB ($0.5000/hr). That is the marginal hardware cost — add idle capacity, ops, and egress for the real figure.
72.7 GB for the weights at FP8 (E4M3), plus 21.5 GB of KV cache for 8 concurrent requests at 8K context — 94.2 GB total. Every extra concurrent user at long context costs more, so leave headroom.
It works. Compare $0.1416 per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 8× RTX 4080 16GB at $0.1062, and the fastest is 4× Cerebras CS-3 (WSE-3) at 43774 tok/s.
Same formulas, your context length, concurrency, and quantisation.