Llama 3.1 8B Instruct on GeForce RTX 4090

Verdict
comfortable
Best quantisation
FP16 / BF16
Max context
58,347 tokens
Decode ceiling
58.8 tok/s

Weight bytes are the MEASURED size of a published GGUF file wherever one exists and params × bits/8 only where none does — the two differ by 33% on sub-1B models and by 89% on natively-quantized releases, so the basis travels with every figure. KV cache is layers × kv_heads × head_dim × 2 × bytes-per-element per token, honouring sliding-window layers. Usable VRAM is a fixed fraction of nameplate memory (different for dedicated and unified). The decode ceiling is bandwidth ÷ bytes-read-per-token, which no runtime beats.

Open this on MakerPortal · Free to reuse with attribution