Ternary weights leverage GPU sparsity for 2x speedup on DGX Spark
bittingthembits · x · 2026-08-23
Ternary weights (-1, 0, 1) naturally fit NVIDIA's 2:4 sparsity pattern, enabling nearly 2x performance gains with minimal quality loss. The DGX Spark, based on the GB10 GPU, offers 1 PFLOP of sparse FP4 performance and 128GB unified RAM for $4,700. Despite slower memory speeds compared to GDDR7, the system minimizes memory manipulation via sparse FP4 compute, FP8 gradients, and limited KV cache. This makes it a cost-effective ($4,700/PFLOP) and decentralized alternative to expensive B300 servers for pretraining.
More from Infra
- Open Source Multi-Host Management for AI Jobs with Agent Orchestration — ii_social · 2026-08-23
- Making Data Centers Likable: Saunas or Lower Electric Bills via Waste Heat? — altryne · 2026-08-23
- Experiment Uses GPU Texture Units to Decode LLM Weights, 1.37× Faster — hp0303 · 2026-08-23
- Fact Check: Claims About Data Center Water Usage Being Overblown Are False — jbwilso1 · 2026-08-23
- Hugging Face launches Buckets for hosting massive 33TB files — NielsRogge · 2026-08-23
- Tesla's Cybercab production line resembles high-speed consumer electronics, outputting a car in under 5 seconds — XFreeze · 2026-08-23