ZGX Fury – How Fast Can it Go?

ZGX Fury – How Fast Can it Go?

More

Summary

Sam Witteveen puts the HP ZGX Fury, a Blackwell-based workstation sponsored by HP and Nvidia, through a series of local inference tests while explaining how running models on your own hardware actually works. The video loads models from DeepSeek, Qwen, GLM, and MiniMax, plus smaller models, to show what the machine can serve for teams and agentic workloads.

He covers the practical math of whether a model will fit: parameters times bits per weight, with roughly 4.5 bits per weight as a realistic figure for 4-bit quantizations. The machine has 252 GB of HBM on the GPU and 496 GB of unified memory attached to the Grace CPU, and the HBM is about 18 times faster, so fitting in HBM is what matters. He also explains why KV cache needs headroom, and why NVFP4 suits Blackwell tensor cores.

On speed, he distinguishes prefill from generation and explains why agentic workloads need around 80 tokens per second per user. He contrasts dense and mixture-of-experts models, showing why a 65B dense model can be slower than a 320B MoE with 18B active parameters, and finds that a one-bit DeepSeek 4.1 Flash quant runs at about 80 tokens per second with surprisingly good output. The video also tests offline coding agents with the network cable unplugged.


📺 Source: Sam Witteveen · Published November 16, 2026
🏷️ Format: Benchmark Test

1 Item

Channels