by NVIDIA · Hardware · Text & code
A 256K-context window with speculative decoding fully consumed its VRAM when running Qwen3.8 27B in SIE