Sakana AI and NVIDIA show a GPU-friendly way to cut LLM data waste

SakanaAILabs · x · 2026-07-22

Sakana AI's joint research with NVIDIA on reducing internal computation in large language models was featured by Nikkei Digital Governance.

• The accompanying graphic explains a new way to reduce data-processing waste while preserving performance.
• Llion Jones said sharing NVIDIA's specialized expertise was a key factor in the project’s success.
• The image contrasts the traditional ELL approach with TwELL, which reorders data within the GPU’s processing window to avoid unnecessary rearrangement.
• The reported idea is that better structural handling of sparse data can make GPU processing faster without sacrificing model quality.

Original post →

More from Infra

• Macaron-V1-Venti ships day one on vLLM with open-source multi-LoRA routing — vllm_project · 2026-07-22
• Developers are starting to rebuild entire software stacks with AI, not just apps — mobileraj · 2026-07-22
• NVIDIA’s Rubin adds tile-level kernel dependency triggers for finer overlap — Moh1tAgarwal · 2026-07-22
• A billboard claims one company can deliver an AI data center in nine months — BenBajarin · 2026-07-22
• India’s chip fabs still depend on imported ultra-pure chemicals — santoshpanda · 2026-07-22
• Cheap just-below-frontier models leave room for distributed serving networks — Ronangmi · 2026-07-22

Infra channel →