Gemma 2 27B at FP8 (E4M3) needs 27.0 GB for its weights and 51.7 GB in total at 8K context with 8 concurrent users. That means it needs 2× L40S 48GB. Every figure is computed by tokcalc's open-source capacity model.
L40S 48GB 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× L40S | 48 GB | 51.7 GB | 0.000 | 0.000 | — | — | ✗ |
| 2× L40S | 96 GB | 51.7 GB | 53.04 | 1273 | 4.88 s | $0.3055 | ✓ |
| 4× L40S | 192 GB | 51.7 GB | 106.1 | 5092 | 2.45 s | $0.1527 | ✓ |
| 8× L40S | 384 GB | 51.7 GB | 212.2 | 20367 | 1.24 s | $0.0764 | ✓ |
| Context | Concurrent users | KV per request | Fits? |
|---|---|---|---|
| 4K | 44 | 1.54 GB | ✓ fits |
| 8K | 22 | 3.09 GB | ✓ fits |
| 33K | 5 | 12.35 GB | ✓ fits |
| 66K | 2 | 24.70 GB | ✓ fits |
| 131K | 1 | 49.39 GB | ✓ fits |
| GPU | GPUs | Aggregate tok/s | Cost / 1M tok |
|---|---|---|---|
| RTX 5090 32GB | 2× | 2640 | $0.1052 |
| AMD MI300X 192GB | 1× | 2297 | $0.3024 |
| AMD MI325X 288GB | 1× | 2600 | $0.3739 |
| A100 80GB SXM4 | 1× | 866.7 | $0.3846 |
| Intel Gaudi 3 128GB | 1× | 1430 | $0.3885 |
| H100 SXM5 80GB | 1× | 1452 | $0.4784 |
Yes, but it needs 2× L40S 48GB. At FP8 (E4M3) the total footprint is 51.7 GB (27.0 GB of weights plus 24.7 GB of KV cache at 8K × 8 concurrent users), against 96 GB of available VRAM across 2 GPUs.
53.04 tokens/sec per stream and 1273 tok/s aggregate with continuous batching at 8 concurrent users. Time to first token is 4.88 s, and inter-token latency is 18.9 ms.
$0.3055 per 1M output tokens at 2× L40S 48GB ($0.7000/hr). That is the marginal hardware cost — add idle capacity, ops, and egress for the real figure.
27.0 GB for the weights at FP8 (E4M3), plus 24.7 GB of KV cache for 8 concurrent requests at 8K context — 51.7 GB total. Every extra concurrent user at long context costs more, so leave headroom.
It works. Compare $0.3055 per 1M tokens against the alternatives below — the cheapest option that fits the same workload is 8× RTX 3080 10GB at $0.0248, and the fastest is 2× Cerebras CS-3 (WSE-3) at 29467 tok/s.
Same formulas, your context length, concurrency, and quantisation.