October 31, 2025
Silicon savior or snake oil?
New analog chip that is 1k times faster than high-end Nvidia GPUs
Internet split: miracle math box or too-good-to-be-true hype
TLDR: A Chinese team says an analog chip beats top Nvidia GPUs by up to 1,000x on specific math tasks while using far less power. The community is split: cautious fans see a promising niche, skeptics want proof beyond small demos and non-public code before believing the hype.
China just dropped a plot twist: an analog chip that the team says can outrun Nvidia’s top graphics cards by up to 1,000x and use 100x less power, per Nature Electronics. It does math using real electrical signals inside memory (RRAM) instead of the usual 1s and 0s, which cuts down on energy-hungry data shuffling. Sounds epic — but the comments? Spicy. The first reaction was pure practicality: “What’s this good for?” Another user drilled into the fine print: it’s blazing fast on small matrix problems used in wireless signals, not necessarily training giant AI like ChatGPT. And the eyebrow-raiser that fueled suspicion: code is “available on request,” not public.
Cue the split. Skeptics yelled “too good to be true,” warning this could be another lab-only miracle. The hype-police brought out the meme batons: “Huge if true, room temperature semiconductor if false” — basically, believe it when it breaks the internet. Meanwhile, the optimists are playing the long game, saying this might be one of many specialized chips that eventually make AI and 6G networks faster and cheaper. Translation: promising demo, narrow use case, and the internet wants receipts before handing over the crown to this analog comeback kid.
Key Points
- •Peking University researchers reported an analog RRAM-based chip in Nature Electronics on Oct. 13.
- •The chip matched digital processors’ accuracy on complex communications tasks while using about 100× less energy.
- •With adjustments, benchmarks showed up to 1,000× higher throughput versus top-end GPUs like Nvidia H100 and AMD Vega 20.
- •The design computes using continuous electrical currents and performs in-memory processing within RRAM arrays.
- •A two-circuit approach (fast approximate plus iterative refinement) was used to achieve high precision in analog computing.