Mixtral 8x22B (MoE) at AWQ 4-bit needs 77.6 GB for its weights and 92.6 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 | 92.6 GB | 0.000 | 0.000 | — | — | ✗ |
| 2× RTX | 64 GB | 92.6 GB | 0.000 | 0.000 | — | — | ✗ |
| 4× RTX | 128 GB | 92.6 GB | 162.5 | 7799 | 3.08 s | $0.0712 | ✓ |
| 8× RTX | 256 GB | 92.6 GB | 324.9 | 31195 | 1.55 s | $0.0356 | ✓ |
| Context | Concurrent users | KV per request | Fits? |
|---|---|---|---|
| 4K | 53 | 962.1 MB | ✓ fits |
| 8K | 26 | 1.88 GB | ✓ fits |
| 33K | 6 | 7.52 GB | ✓ fits |
| 66K | 3 | 15.03 GB | ✓ fits |
| 131K | 1 | 30.06 GB | ✓ fits |
| GPU | GPUs | Aggregate tok/s | Cost / 1M tok |
|---|---|---|---|
| A100 80GB SXM4 | 2× | 2176 | $0.3064 |
| H100 SXM5 80GB | 2× | 3645 | $0.3811 |
| AMD MI300X 192GB | 1× | 1696 | $0.4095 |
| L40S 48GB | 2× | 940.0 | $0.4137 |
| AMD MI325X 288GB | 1× | 1920 | $0.5064 |
| Intel Gaudi 3 128GB | 1× | 1056 | $0.5261 |
Yes, but it needs 4× RTX 5090 32GB. At AWQ 4-bit the total footprint is 92.6 GB (77.6 GB of weights plus 15.0 GB of KV cache at 8K × 8 concurrent users), against 128 GB of available VRAM across 4 GPUs.
162.5 tokens/sec per stream and 7799 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 3.08 s, and inter-token latency is 6.2 ms.
$0.0712 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.
77.6 GB for the weights at AWQ 4-bit, plus 15.0 GB of KV cache for 8 concurrent requests at 8K context — 92.6 GB total. Every extra concurrent user at long context costs more, so leave headroom.
It works. Compare $0.0712 per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 8× RTX 4080 16GB at $0.0534, and the fastest is 4× Cerebras CS-3 (WSE-3) at 87040 tok/s.
Same formulas, your context length, concurrency, and quantisation.