DGX B300 vs. DGX Spark: The Case for Decentralized Pretraining

aarthir · x · 2026-08-20

Jon Durbin compared the compute economics of NVIDIA's B300 servers against the DGX Spark. While a B300 server offers 9 PFLOPS dense FP4 compute for $350k, its extreme power and cooling needs make collocation difficult. In contrast, the DGX Spark costs $4,700, delivers 1 PFLOP sparse FP4, and runs on standard household power. With similar cost per PFLOP but vastly lower deployment barriers—and a native fit for sparse FP4 compute with ternary weights—Durbin argues that decentralized pretraining on DGX Spark is a promising direction to test.

Related event: Cost Analysis: DGX Spark Matches B300 Server Per-PFLOP Value(2 posts)→

Original post →

More from Infra

Infra channel →