Chaining DGX Spark boxes with QSFP cables for dumb data-parallel training actually works
austinvhuang · x · 2026-11-19
While DGX Spark's memory bandwidth is mediocre for n=1 interactivity, Austin Huang found that linearly scaling multiple units with QSFP cables for simple data-parallel training runs works nicely — his new bedtime ritual is kicking off exploratory continued-pretraining runs before sleep.
Related event: Linking Multiple DGX Sparks via QSFP for Data-Parallel Training(2 posts)→
More from Infra
- Oxmiq Offers AI Chip IP and Tools to Indian Teams, Discusses Custom Silicon Economics — RajaXg · 2026-11-19
- SREGym 2.0 recreates the Microsoft Outlook outage: a latent gateway bug plus cascading failures frontier AI can't yet solve — ajratner · 2026-11-19
- Google ships a free fully-local AI notetaker for Mac powered by EmbeddingGemma 2 — dr_cintas · 2026-11-19
- 4x 256GB M5 Ultra Mac cluster with exo serves 32 users at 2000 tok/s on Qwen 27B — StefanoGogioso · 2026-11-19
- Is 262K context worth +0.0011 KLD? A local-deployment KV cache tradeoff question — Distinct-Pie2389 · 2026-11-19
- Anthropic's compute spend grows 4x/year, $10B in Q3 alone, set to overtake OpenAI by May 2027 — scaling01 · 2026-11-19