LLM Hacker News Apr 8, 2026 2 min read
MegaTrain proposes training 100B+ parameter LLMs at full precision on a single GPU by keeping parameters and optimizer states in host memory and streaming layers through the device. The recent Hacker News interest is notable because the paper reframes the problem as one of memory-system design rather than simple GPU count.