NVIDIA RTX 5070 SUPER
GPURTX 5070 SUPER (rumored, CES 2026): 18GB VRAM, GDDR7 28 Gbps, 256-bit bus. Targets 1440p high-refresh + on-device 7-13B Q4 models.
6
Run records
6
Models covered
1
Frameworks covered
L0 Self-reported × 6
Evidence mix
Typical measured performance
Published benchmark averages per model × framework × quantization on this hardware.
- Strata · IQ2_XS Decode 38 · Prefill 285 ·
- Strata · IQ2_XXS Decode 7.8 · Prefill 58.5 ·
- Strata · IQ2_XS Decode 38 · Prefill 285 ·
- Strata · IQ2_XXS Decode 4.4 · Prefill 33 ·
- Strata · IQ2_XS Decode 36.9 · Prefill 276.8 ·
- Strata · IQ4_XS Decode 208.4 · Prefill 1563 ·
Key specs
- Vendor
- NVIDIA
- Architecture
- Blackwell
- VRAM
- 18 GB
- Memory bandwidth
- 896 GB/s
- TDP
- 280 W
- Release date
- 2026-03-01
Run records
6 configurations
| Model | Framework | Quantization | Decode | Gen time | Samples |
|---|---|---|---|---|---|
| Mistral Large 3 70B | Strata | IQ2_XS | 38 tok/s | — | 1 |
| Llama 5 405B | Strata | IQ2_XXS | 7.8 tok/s | — | 1 |
| Llama 5 70B | Strata | IQ2_XS | 38 tok/s | — | 1 |
| Qwen 4 720B | Strata | IQ2_XXS | 4.4 tok/s | — | 1 |
| Qwen 4 72B | Strata | IQ2_XS | 36.9 tok/s | — | 1 |
| Qwen 4 8B | Strata | IQ4_XS | 208.4 tok/s | — | 1 |