Table of Contents
RTX 3090 (Ampere, £159/mo) and RTX 4090 (Ada, £289/mo) both have 24 GB VRAM. The 4090 is meaningfully faster, but the cost premium changes the per-token math.
TL;DR
RTX 4090 wins on absolute throughput; RTX 3090 wins on cost-per-token for FP16 7B chatbots. For 13B+ AWQ-INT4 the 4090 is faster enough to win per-pound. Pick based on workload.
Specs
| Spec | 3090 | 4090 |
|---|---|---|
| VRAM | 24 GB | 24 GB |
| Bandwidth | 936 GB/s | 1,008 GB/s |
| FP16 TFLOPS | ~36 | ~165 |
| Monthly | £159 | £289 |
Benchmarks
| Workload | 3090 | 4090 | Cost per 1M (3090) | Cost per 1M (4090) |
|---|---|---|---|---|
| Mistral 7B FP16 | 720 tok/s | 950 tok/s | £0.16 | £0.19 |
| Llama 3 8B FP16 | 680 tok/s | 910 tok/s | £0.17 | £0.20 |
| Qwen 2.5 14B INT4 | 410 tok/s | 590 tok/s | £0.28 | £0.30 |
| SDXL 1024² (s/img) | 14 s | 4 s | higher | much lower |
Verdict
3090 wins per-pound on FP16 7B chatbots. 4090 wins per-pound on image generation and 13B+ INT4. For most production deployments at moderate scale they're close.
Bottom line
The 3090 is the budget LLM card; the 4090 is the budget image-gen card. See 3090 vs 5090.
