Baidu's ERNIE Evolve Hid Its Scorer, Then Posted 6.2x Gain on Kunlun Chips
Reported by Tech Times1 source ↓
Baidu says ERNIE Evolve raised inference speed on Kunlun P800 chips to 623.6% of baseline. Its post reports smaller gains on GPUs, whose starting performance Baidu says had already been tuned by its engineers.
Details
Baidu describes ERNIE Evolve as a two-level research loop with five roles. For Kunlun P800 inference, the system changed the Kunlun adapter plugin, including using fused kernels, enabling speculative decoding, removing a device-to-host synchronization step and correcting dispatch logic. A proposed replacement for collective communication was dropped after it made performance 14.0% slower.
The results use separate baselines for Kunlun and GPUs, so the reported gains are not a like-for-like comparison. Baidu says its GPU baselines had already been tuned by engineers; outside analysis cited in the article also points to limitations in the Kunlun software path as context for its lower starting point. The article reports an independent Kunlun adapter evaluation that reached up to 80% of H20 throughput across most batch sizes after a separate round of fixes.
Baidu's post says the evaluator was hidden from the optimizing agent, but does not give the noise threshold or report how many candidate versions the Reviewer voided for gaming the score. The article says those omissions prevent outside readers from checking how well the defense worked in this case.
Written from coverage as of
Stocks in this story
Change since open: 20 Nov to 21 Nov 2026
Last refresh