$21 billion AI Chip Newcomer Etched Under Scrutiny: Performance Yet to Receive Third-Party Validation

Insightful Beating AI News Flash: Just valued at $21 billion and freshly raised $700 million, Etched has caught the attention of the chip industry.

Renowned hacker George Hotz's AI computing team, the tiny corp, behind the open-source deep learning framework tinygrad, has publicly questioned Etched's technical hype: while there are plenty of investors, orders, and hardware photos, there is too little data available to truly validate performance.

One of Etched's key selling points is LVI, which enables chips to run AI inference at lower voltages. Etched claims that this can push trillion-parameter sparse MoEs to over 80% of theoretical peak compute power.

Chip design expert Wesley Yue has raised doubts that a high ratio necessarily translates to superior absolute performance. Model FLOPS Utilization (MFU) measures the actual computation as a percentage of theoretical peak, meaning that if a chip's own peak compute power is lower, even at 80% utilization, it may still lag behind competitors.

Etched has yet to publicly disclose complete FLOPs, power consumption, and third-party benchmarks. The official website still only states that "early customer testing has achieved leading levels" and promises to release detailed performance data later.

However, there is currently no evidence to prove that Etched is engaged in deception. Both The Wall Street Journal and Reuters have confirmed that its chips are shipping. Jane Street received its first full rack last month and has commenced deployment.

As of now, the biggest question is not "whether the chip exists," but rather if this chip truly lives up to the extensive publicity.