Table of Contents
The RTX 5070 is NVIDIA’s mid-range Blackwell card: 12 GB GDDR7 at ~672 GB/s, 6,144 CUDA cores, 5th Gen Tensor Cores with native FP4/FP8 support, full CUDA ecosystem, at £139/mo. The constraint is VRAM — 12 GB means 7B LLMs at FP16 don’t fit, and Flux.1 needs GGUF quantisation. Within those limits it’s the cheapest full-CUDA Blackwell card on GigaGPU.
Full Spec Table
| Spec | RTX 5070 |
|---|---|
| Architecture | NVIDIA Blackwell (GB205) |
| CUDA Cores | 6,144 (48 SMs) |
| Tensor Cores | 5th Gen — INT8, FP8, FP4, BF16 |
| RT Cores | 4th Gen |
| VRAM | 12 GB GDDR7 |
| Memory bus | 192-bit |
| Memory bandwidth | ~672 GB/s |
| FP32 compute | ~36 TFLOPS |
| FP16 Tensor (no sparsity) | ~72 TFLOPS |
| INT8 Tensor (no sparsity) | ~144 TOPS |
| TDP | ~150 W |
| PCIe | 5.0 x16 |
| DLSS | DLSS 4 + Multi Frame Generation |
| NVLink | No |
| Host CPU | AMD Ryzen 7 / 9 |
| Host RAM | Up to 128 GB DDR5 |
| Storage | 1 TB NVMe + 4 TB SATA |
| Network | 1 Gbps unmetered |
| Location | UK data centre |
| GigaGPU price | £139/mo |
What Blackwell Changes for AI
Blackwell (GB series) introduces several changes that directly affect AI inference workloads vs the previous Ada Lovelace generation:
- Native FP4 Tensor compute: Blackwell adds first-class FP4 (4-bit floating point) as a first-party operation on Tensor Cores. For LLM inference with extreme quantisation, this enables higher throughput at lower precision without resorting to INT4 tricks. TensorRT-LLM and llama.cpp both expose FP4 paths on Blackwell.
- FP8 improvements: 5th Gen Tensor Cores handle FP8 with better accuracy than the FP8 experimental path on Ada. The FP8 path in vLLM and TensorRT-LLM is the recommended production path for Blackwell.
- Memory efficiency: GDDR7 delivers significantly higher bandwidth per pin than GDDR6X. At 192-bit, the RTX 5070 achieves ~672 GB/s — higher than the RTX 4080 (16 GB GDDR6X, 736 GB/s on 256-bit).
- Power efficiency: ~150 W TDP vs ~200 W for the RTX 4070 Ti (comparable Ada card). Lower operating cost over 24/7 workloads.
5th Gen Tensor Cores: FP4 and FP8
For AI inference, the Tensor Core generation matters most. The RTX 5070’s 5th Gen Tensor Cores support:
| Precision | Use case | Support on 5070 |
|---|---|---|
| FP32 | Training, full-precision inference | CUDA cores (not Tensor) |
| TF32 | Training on Tensor Cores | Yes |
| FP16 / BF16 | Standard inference, diffusion models | Yes — Tensor Cores |
| INT8 | Quantised LLM inference | Yes — Tensor Cores |
| FP8 | Production LLM inference (vLLM, TRT-LLM) | Yes — native Blackwell |
| FP4 | Ultra-compressed LLM inference | Yes — Blackwell exclusive |
| INT4 | Extreme compression (GGUF Q4 via llama.cpp) | Via CUDA, efficient |
FP8 is the production-recommended precision for LLM inference on Blackwell via vLLM. Models quantised to FP8 retain ~99% of FP16 quality while roughly halving memory footprint — meaning a 7B model at FP8 needs ~7 GB instead of ~14 GB, which fits comfortably in 12 GB.
12 GB GDDR7: The VRAM Trade-off
12 GB is the RTX 5070’s main constraint for AI workloads. What it means in practice:
- 7B models at FP16 don’t fit (need ~14 GB). Use Q4_K_M (~4.4 GB) or FP8 (~7 GB) instead.
- SDXL fits comfortably at ~8 GB FP16, with headroom for a single ControlNet.
- Flux.1 requires GGUF quantisation: Q4 transformer (~7 GB) + T5/CLIP/VAE (~6 GB) = ~13 GB tight — use Q8 transformer or Q4 with tiled approach.
- Max comfortable LLM size at Q4: ~13B (Q4_K_M ~8 GB + KV cache ~3 GB = 11 GB).
The bandwidth story is positive: 672 GB/s GDDR7 is faster than GDDR6X at the same bus width and beats the Arc Pro B60 (456 GB/s) and matches the RTX 3090 (936 GB/s) directionally — you get fast memory for the VRAM you have.
Where the 5070 Sits in the GigaGPU Catalogue
| GPU | VRAM | CUDA? | Bandwidth | Price |
|---|---|---|---|---|
| RTX 5060 Ti | 16 GB GDDR7 | Yes | ~448 GB/s | £119/mo |
| RTX 5070 ← this card | 12 GB GDDR7 | Yes | ~672 GB/s | £139/mo |
| RX 9070 XT | 16 GB GDDR6 | No (ROCm) | ~576 GB/s | £149/mo |
| RTX 3090 | 24 GB GDDR6X | Yes | 936 GB/s | £159/mo |
| RTX 5080 | 16 GB GDDR7 | Yes | ~960 GB/s | £189/mo |
The RTX 5070’s niche: fastest Blackwell CUDA card below £140. The RTX 5060 Ti (£119) has 4 GB more VRAM but much lower bandwidth and compute. The RTX 5080 (£189) has 4 GB more VRAM, much higher bandwidth, and faster compute. The RTX 3090 (£159) has 12 GB more VRAM for £20 more — the key choice is VRAM headroom vs Blackwell speed.
Which AI Workloads Fit
- ✅ SDXL image generation — 8 GB FP16 fits comfortably
- ✅ SD 1.5 — trivial, 4 GB FP16
- ✅ Flux.1 Schnell / Dev via GGUF Q4 — fits with care
- ✅ 7B–13B LLMs via Q4_K_M — comfortable
- ✅ Whisper Large-v3 + 7B LLM simultaneously — ~8 GB total
- ✅ QLoRA fine-tuning of 7B models — comfortable
- ✅ Computer vision / YOLO inference — trivial
- ✅ Embedding generation — trivial
- ⚠️ Flux.1 Dev FP16 — requires quantisation
- ❌ 70B+ LLMs at Q4 — exceeds 12 GB
- ❌ 7B+ FP16 inference — 14+ GB required
Deploy the RTX 5070
12 GB GDDR7 · Blackwell · Full CUDA · £139/mo · Order the RTX 5070 or compare all GPUs.
