Table of Contents
RTX 5070 is faster per FLOP and has native Blackwell FP8/FP4. RTX 3090 has 24 GB vs 12 GB — double the VRAM at £20/mo more. For models that fit in 12 GB, the 5070 is faster. For anything needing 16–24 GB (7B FP16, 13B+ Q4 with large context, Flux.1 FP16, multi-ControlNet stacks), the 3090 wins on capacity. CUDA ecosystem is excellent on both.
Spec Comparison
| Spec | RTX 5070 | RTX 3090 |
|---|---|---|
| Architecture | Blackwell (GB205) | Ampere (GA102) |
| VRAM | 12 GB GDDR7 | 24 GB GDDR6X |
| Memory bandwidth | ~672 GB/s | 936 GB/s |
| CUDA Cores | 6,144 | 10,496 |
| FP32 compute | ~36 TFLOPS | ~35 TFLOPS |
| Tensor Cores | 5th Gen (FP4/FP8) | 3rd Gen (FP16/INT8) |
| FlashAttention 3 | Yes | No (FA2 only) |
| FP8 native | Yes | Limited |
| TDP | ~150 W | ~350 W |
| ECC | No | No |
| Price | £139/mo | £159/mo |
On paper these cards are surprisingly close in FP32 compute. The 3090 has more raw CUDA cores but is an older architecture running at lower efficiency per core. The 5070 has fewer cores but Blackwell’s improvements — better IPC, FlashAttention 3, native FP8 — close the gap on real AI workloads. The dominant difference remains: 12 GB vs 24 GB VRAM, and 672 GB/s vs 936 GB/s bandwidth.
Image Generation
| Workload | RTX 5070 (£139) | RTX 3090 (£159) | Winner |
|---|---|---|---|
| SD 1.5 | ~1 s/img | ~1.1 s/img | Tie |
| SDXL (1024×1024) | ~4–5 s | ~5–6 s | 5070 (bandwidth + Blackwell) |
| Flux.1 Schnell Q4 | ~3 s | ~3 s | Tie |
| Flux.1 Dev Q4 GGUF | Tight (12 GB) | Comfortable (24 GB) | 3090 (capacity) |
| Flux.1 Dev FP16 | No (12 GB) | Yes (24 GB, ~20 s) | 3090 |
| SDXL + 2× ControlNet | No (OOM) | Yes (24 GB) | 3090 |
| Power draw | ~150 W | ~350 W | 5070 |
For SDXL at standard resolutions, the RTX 5070 is actually faster than the RTX 3090 — higher efficiency Blackwell architecture and GDDR7’s per-operation speed compensate for the lower raw bandwidth. For Flux.1 Dev where 24 GB allows the full FP16 pipeline, the 3090 has a meaningful advantage.
LLM Inference
| Model + Quant | RTX 5070 tok/s | RTX 3090 tok/s | Winner |
|---|---|---|---|
| Llama 3.1 8B Q4_K_M | ~75–85 | ~65–70 | 5070 (~15% faster) |
| Mistral 7B Q4_K_M | ~80–90 | ~70–75 | 5070 |
| Qwen 2.5 14B Q4_K_M | ~45–55 | ~45–55 | Tie |
| Llama 3.1 8B FP8 | Yes (8 GB on 12) | Yes (8 GB on 24) | 5070 (faster; FP8 native) |
| Llama 3.1 8B FP16 | No (16 GB) | Yes (16 GB on 24) | 3090 |
| Gemma 2 27B Q4 | No (>12 GB) | Yes (~16 GB on 24) | 3090 |
For models that fit in 12 GB at Q4 or FP8, the RTX 5070 is faster than the RTX 3090. For models between 12 GB and 24 GB (7B FP16, 13B FP8, 27B Q4), the 3090 is the only option at this price tier.
Software Ecosystem
Both cards are NVIDIA CUDA — the full ecosystem applies to both:
- vLLM, TensorRT-LLM, llama.cpp CUDA, Ollama: ✅ both
- bitsandbytes INT4/INT8: ✅ both
- FlashAttention 2: ✅ both
- FlashAttention 3: ✅ 5070 (Blackwell), ❌ 3090 (Ampere)
- FP8 native Tensor: ✅ 5070, ❌ 3090
- xformers: ✅ both
The RTX 5070 has access to new Blackwell-exclusive kernel paths in vLLM and TensorRT-LLM that provide additional throughput improvements not available on Ampere.
When RTX 5070 Wins
- SDXL generation speed is the priority
- 7B–13B Q4 or FP8 LLM inference where 12 GB is sufficient
- Power efficiency matters — RTX 5070 uses ~200 W less per hour
- Budget: £139 vs £159, saving £20/mo (£240/year)
- You want Blackwell features: FlashAttention 3, FP8 native, FP4 Tensor
When RTX 3090 Wins
- 7B models at FP16 (needs 16 GB minimum)
- Flux.1 Dev at FP16 or with Q4 + breathing room
- Multi-ControlNet SDXL stacks (>12 GB)
- 13B+ models at Q4 with large context windows
- Running image generation alongside LLM simultaneously
- 24/7 batch jobs where VRAM ceiling defines throughput
Verdict
For models that fit in 12 GB, the RTX 5070 at £139/mo is the better buy — faster, newer architecture, lower power. For workloads that need 16–24 GB (larger models, Flux.1 FP16, multi-ControlNet), the RTX 3090 at £159/mo is a £20/mo premium that buys twice the VRAM. See the full catalogue including the Arc Pro B60 (24 GB ECC at £129/mo) as a third option if VRAM is the priority and CUDA isn’t required.
Compare and deploy
RTX 5070 — £139/mo, 12 GB · RTX 3090 — £159/mo, 24 GB · All GPUs.
