NVIDIA DGX Spark × Llama 3.3 70B Instruct
Can the DGX Spark run Llama 3.3 70B Instruct?
Barely
Yes. The DGX Spark can run Llama 3.3 70B Instruct at up to Q8_0 (needs ~80.7 GB of 115 GB usable), at an estimated ~5.1 tokens/sec at Q4_K_M.
Speed is estimated from memory bandwidth calibrated with a measured 57.21 tok/s Llama 2 7B Q4_0 result (source). No direct public benchmark of this exact pair is recorded yet.
By quantization (8K context)
| Quant | Memory needed | Fits? | Est. tok/s |
|---|---|---|---|
| Q3_K_M | 38.9 GB | ✅ | ~6.4 |
| Q4_K_M | 47.6 GB | ✅ | ~5.1 |
| Q5_K_M | 55.3 GB | ✅ | ~4.3 |
| Q6_K | 63.5 GB | ✅ | ~3.8 |
| Q8_0 | 80.7 GB | ✅ | ~2.9 |
| F16 | 149 GB | ❌ | — |
Fastest hardware for Llama 3.3 70B Instruct
- Apple M5 Max (40-core GPU, 128GB) — ~10.7 tok/s at Q4_K_M
- Apple M2 Ultra (76-core GPU, 192GB) — ~8.4 tok/s at Q4_K_M
- Apple M3 Ultra (80-core GPU, 512GB) — ~8.2 tok/s at Q4_K_M
- Apple M4 Max (40-core GPU, 128GB) — ~7.4 tok/s at Q4_K_M
- Apple M5 Pro (20-core GPU, 64GB) — ~5.9 tok/s at Q4_K_M
No suitable GPU? Rent one by the hour on Vast.ai or RunPod (affiliate links).
Get new models & benchmarks in your inbox
No spam. Unsubscribe anytime by replying to any email.