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.

Original post →

More from Infra

Infra channel →