Table of Contents
Both are Blackwell with 5th Gen Tensor Cores and GDDR7. The RTX 5080 has 4 GB more VRAM (16 vs 12) and ~43% more bandwidth (~960 vs ~672 GB/s) for £50/mo more. If your workloads fit in 12 GB, the 5070 at £139/mo is excellent value. If you need 16 GB (Flux.1 Dev Q4 comfortably, SDXL + multi-ControlNet, 7B FP8 with context headroom), the 5080 is worth the premium.
Spec Comparison
| Spec | RTX 5070 | RTX 5080 |
|---|---|---|
| Architecture | Blackwell GB205 | Blackwell GB203 |
| CUDA Cores | 6,144 | 10,752 |
| VRAM | 12 GB GDDR7 | 16 GB GDDR7 |
| Memory bandwidth | ~672 GB/s | ~960 GB/s |
| FP32 compute | ~36 TFLOPS | ~60 TFLOPS |
| Tensor Cores | 5th Gen | 5th Gen |
| FP8 / FP4 | Yes | Yes |
| TDP | ~150 W | ~260 W |
| Price | £139/mo | £189/mo |
The RTX 5080 uses the GB203 die (same family as the 5090’s GB202). It’s a substantially faster GPU across the board — ~67% more CUDA cores, ~43% more bandwidth, ~67% more FP32 compute. The 5070 is the entry point to Blackwell AI workloads; the 5080 is the mid-high tier. The premium is £50/mo (£600/year).
Image Generation
| Workload | RTX 5070 (£139) | RTX 5080 (£189) | Winner |
|---|---|---|---|
| SDXL (1024×1024) | ~4–5 s | ~2.5–3 s | 5080 (~60% faster) |
| Flux.1 Schnell Q4 | ~3 s (tight on 12 GB) | ~1.5 s (comfortable on 16 GB) | 5080 |
| Flux.1 Dev Q4 | Very tight (12 GB) | Comfortable (16 GB) | 5080 |
| SDXL + 2× ControlNet | No (OOM) | Yes (16 GB) | 5080 |
| SD 1.5 (512×512) | ~1 s | ~0.6 s | 5080 (but difference negligible) |
LLM Inference
| Model | RTX 5070 tok/s | RTX 5080 tok/s | 5080 premium worth it? |
|---|---|---|---|
| Llama 3.1 8B Q4 | ~75–85 | ~110–120 | Yes, if throughput matters |
| Mistral 7B Q4 | ~80–90 | ~115–125 | Yes |
| Qwen 2.5 14B Q4 | ~45–55 | ~70–80 | Yes |
| Llama 3.1 8B FP8 | ~85 (tight) | ~120 (comfortable) | Yes |
| Llama 3.1 8B FP16 | No (12 GB) | No (16 GB — still short) | Need RTX 3090/5090 |
The RTX 5080’s ~43% bandwidth advantage converts directly to ~40% more tokens per second on the same model. For high-concurrency inference or latency-sensitive applications, the difference is meaningful. For personal/single-user use cases, 75–85 tok/s on the 5070 is already very fast.
Cost Analysis
| Period | RTX 5070 | RTX 5080 | 5080 premium |
|---|---|---|---|
| 1 month | £139 | £189 | +£50 |
| 6 months | £834 | £1,134 | +£300 |
| 12 months | £1,668 | £2,268 | +£600 |
£600/year buys a meaningful compute upgrade. The question is whether that upgrade translates to real value for your workload. For interactive personal use on 7B models, the 5070 is already faster than human reading speed at 75 tok/s — spending £600/year more to go to 110 tok/s is wasteful. For batch inference pipelines or serving multiple concurrent users, the 5080’s throughput advantage pays for itself.
When to Pick the RTX 5070
- Models fit within 12 GB (7B–13B Q4 or FP8, SDXL, Flux.1 Schnell)
- Personal or low-concurrency use — 75 tok/s is more than sufficient
- Budget is the primary consideration — save £50/mo vs 5080
- Single SDXL ControlNet is the extent of image generation complexity
When to Pick the RTX 5080
- Flux.1 Dev Q4 without stress — 16 GB gives comfortable headroom
- SDXL + 2× ControlNet + IP-Adapter stacking — needs 13–16 GB
- High-concurrency LLM serving — 40% more throughput at all model sizes
- 7B at FP8 with long context windows — 16 GB gives more KV cache room
- Future-proofing for larger models as Q4/Q8 model sizes grow
Verdict
For image generation and LLM workloads that stay within 12 GB, the RTX 5070 at £139/mo is excellent value — fast Blackwell CUDA, GDDR7 bandwidth, all CUDA frameworks. The RTX 5080 at £189/mo is the better pick when 16 GB is needed or when you’re running high-concurrency serving and every tok/s counts. See the full catalogue.
Compare Blackwell tiers
RTX 5070 — £139/mo, 12 GB · RTX 5080 — £189/mo, 16 GB · All GPUs.
