Medusa Halo leak points to LPDDR6 and ~691GB/s, builder plans 512GB local LLM rig under $20K
gnukeith · x · 2026-11-08
Leaked specs suggest AMD's Medusa Halo will ship with LPDDR6 (DDR6 was designed with AI in mind, unlike DDR5), with a theoretical peak of 691 GB/s — boosting both inference speed and parallelism.
@gnukeith explains why he'd skip GPUs: they're pricier, and he wants capacity over speed to run bigger models while hyper-optimizing his own inference engine (possibly open-sourced Dec/Jan). He optimistically expects a 512GB RAM option, so four units would cost under $20K.
He notes DeepSeek V4.1 Flash runs 160 tokens/s per session at near-lossless quant, and estimates 30–50 tokens/s across 5 concurrent sessions with heavy parallelism work.
More from Infra
- SpaceX plans 6 GW AI campus near Willcox, Arizona with dry cooling — aronchick · 2026-11-08
- François Fleuret asks: after a loss spike and recovery, is final loss worse, same, or better? — francoisfleuret · 2026-11-08
- ResearchBoost Agent speeds up real research workloads 2.51x on average across 12 tasks — YouJiacheng · 2026-11-08
- Dev builds Cray-2 emulator from scratch, runs a GPT-3-class LLM at 1 token per 4s — inductionheads · 2026-11-08
- Hyperbolic CEO on compute market-making: bridging 3-year contracts and 3-week needs — YiMaTweets · 2026-11-08
- Ben Bajarin mocks 'we don't need data centers' take: CSPs were 'useless' too — BenBajarin · 2026-11-08