Redditor builds ROI calculator matching local LLMs to hardware by memory bandwidth
SnoobieJunes · reddit · 2026-09-19
A Reddit user shared a self-built ROI calculator that estimates which local models run well on which hardware based on memory bandwidth.
- The tool was originally created as the author's own sales tool, quickly generated with Grok, and not yet closely audited
- The author may open-source the repo and plans to update it once benchmarks for next week's new Macs (with faster bandwidth) come online
More from Infra
- Anthropic sets up SF wet lab, plans to scale compute to 5GW by year-end — TinfoilTricorn · 2026-09-19
- How OpenAI Used Its Own LLMs to Design Its Jalapeño Chip — maxall4 · 2026-09-19
- Qwen 27B Hits 1253 tok/s on 2x RTX 5090, 8-10x Cheaper Than API Pricing — me_broke · 2026-09-19
- India hosts ~20% of global chip design engineers, and the real share is likely higher — bookwormengr · 2026-09-19
- Micron and the AI memory cycle: HBM heads to $100B by 2027, 5-year deals reshape the trade — Beth_Kindig · 2026-09-19
- NVIDIA's AIPerf benchmarks LLM inference at scale with multiprocess load testing — NVIDIAAI · 2026-09-19