Qwen 3 235B-A22B (MoE) at AWQ 4-bit needs 129.3 GB for its weights and 141.9 GB in total at 8K context with 8 concurrent users. That means it needs 8× RTX 5090 32GB. Every figure is computed by tokcalc's open-source capacity model.
RTX 5090 32GB has no NVLink, so 8-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 | 141.9 GB | 0.000 | 0.000 | — | — | ✗ |
| 2× RTX | 64 GB | 141.9 GB | 0.000 | 0.000 | — | — | ✗ |
| 4× RTX | 128 GB | 141.9 GB | 0.000 | 0.000 | — | — | ✗ |
| 8× RTX | 256 GB | 141.9 GB | 576.0 | 55300 | 889.9 ms | $0.0201 | ✓ |
| Context | Concurrent users | KV per request | Fits? |
|---|---|---|---|
| 4K | 160 | 807.5 MB | ✓ fits |
| 8K | 80 | 1.58 GB | ✓ fits |
| 33K | 20 | 6.31 GB | ✓ fits |
| 66K | 10 | 12.62 GB | ✓ fits |
| 131K | 5 | 25.23 GB | ✓ fits |
| GPU | GPUs | Aggregate tok/s | Cost / 1M tok |
|---|---|---|---|
| L40S 48GB | 4× | 6666 | $0.1167 |
| A100 80GB SXM4 | 2× | 3857 | $0.1728 |
| H100 SXM5 80GB | 2× | 6461 | $0.2150 |
| AMD MI300X 192GB | 1× | 3007 | $0.2310 |
| H200 SXM5 141GB | 2× | 9258 | $0.2400 |
| AMD MI325X 288GB | 1× | 3404 | $0.2856 |
Yes, but it needs 8× RTX 5090 32GB. At AWQ 4-bit the total footprint is 141.9 GB (129.3 GB of weights plus 12.6 GB of KV cache at 8K × 8 concurrent users), against 256 GB of available VRAM across 8 GPUs.
576.0 tokens/sec per stream and 55300 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 889.9 ms, and inter-token latency is 1.7 ms.
$0.0201 per 1M output tokens at 8× RTX 5090 32GB ($0.5000/hr). That is the marginal hardware cost — add idle capacity, ops, and egress for the real figure.
129.3 GB for the weights at AWQ 4-bit, plus 12.6 GB of KV cache for 8 concurrent requests at 8K context — 141.9 GB total. Every extra concurrent user at long context costs more, so leave headroom.
It works. Compare $0.0201 per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 8× RTX 5090 32GB at $0.0201, and the fastest is 4× Cerebras CS-3 (WSE-3) at 154298 tok/s.
Same formulas, your context length, concurrency, and quantisation.