Table of Contents
The RTX 5060 Ti 16 GB at £119/mo is the cheap entry tier. The RTX 4090 24 GB at £289/mo is a meaningful step up — older architecture but more VRAM and roughly 1.7× faster on FP16. For teams choosing between them, the answer is workload-dependent.
Pick 5060 Ti 16 GB if your workload is 7B/8B chatbots, embeddings, Whisper, or single-pipeline image gen. Pick 4090 24 GB if you need 13B FP16, Mixtral 8x7B INT4, multi-model stacks, or higher concurrency. The 4090 buys you ~50% more capacity for ~65% more cost — usually worth it when you cross VRAM thresholds.
Spec comparison
| Spec | RTX 5060 Ti 16 GB | RTX 4090 24 GB |
|---|---|---|
| Architecture | Blackwell | Ada Lovelace |
| VRAM | 16 GB GDDR7 | 24 GB GDDR6X |
| Memory bandwidth | 448 GB/s | 1,008 GB/s |
| FP16 TFLOPS | ~24 | ~83 |
| FP8 hardware | Yes (~184 TOPS) | No (software only) |
| Monthly | £119 | £289 |
Two opposite trade-offs. The 5060 Ti has FP8 hardware and the newer architecture; the 4090 has more VRAM and meaningfully more raw compute.
By workload
| Workload | 5060 Ti 16 GB | RTX 4090 24 GB | Winner |
|---|---|---|---|
| Mistral 7B FP8 | ~880 tok/s | ~1,200 tok/s (sw FP8) | 4090 (faster) |
| Mistral 7B FP16 | ~580 tok/s | ~950 tok/s | 4090 (faster) |
| Llama 3 8B FP8 | ~820 tok/s | ~1,150 tok/s | 4090 |
| Qwen 2.5 14B FP16 | Does not fit | Fits, tight | 4090 |
| Mixtral 8x7B INT4 | Tight | Fits | 4090 |
| Whisper-only | Trivial | Trivial | Tie (5060 Ti cheaper) |
| SDXL | 7s/image | 4s/image | 4090 faster |
| FLUX.1 dev FP8 | 14s/image | 8s/image (sw FP8) | 4090 faster |
| Voice agent stack | Tight, ~6 concurrent | Comfortable, ~12 | 4090 |
| Embedding-only | Plenty | Overkill | 5060 Ti (cheaper) |
| Cost per 1M tokens (Mistral 7B) | £0.12 (FP8) | £0.19 | 5060 Ti |
Verdict
- 5060 Ti: best when 7B/8B is enough and FP8 unlocks the throughput you need.
- 4090: best when you cross 16 GB or need higher absolute throughput.
- Both wrong: if your workload genuinely needs >24 GB or FP8 is critical, go to 5090 for £399/mo.
Bottom line
The 5060 Ti is the better cost-per-token card; the 4090 is the better capability card. For most teams the 5060 Ti is enough until they hit a VRAM wall — at which point the right upgrade is usually the 5090, not the 4090. The 4090 is best when stock pricing makes it cheaper than the 5090.
