News
RaBitQ researchers Cheng Long and Jianyang Gao allege TurboQuant authors published misleading GPU versus CPU performance comparisons
Saturday, March 28, 2026 at 01:13 AM
Authors of the RaBitQ quantization method have issued a formal public complaint against the authors of TurboQuant, alleging incomplete methodology disclosure and misleading empirical comparisons. The dispute centers on the use of Johnson-Lindenstrauss transforms and random rotations in vector quantization, with RaBitQ's team claiming TurboQuant's performance gains over their baseline were reported without disclosing that TurboQuant ran on NVIDIA A100 GPUs while the RaBitQ baseline was restricted to a single CPU thread.
Context
Researchers Cheng Long and Jianyang Gao, authors of the RaBitQ quantization method, have issued a formal public allegation against the authors of Google Research's TurboQuant, claiming the latter published misleading performance comparisons. The dispute, which escalated on March 26, 2026, alleges that TurboQuant achieved its claimed 8x performance boost on NVIDIA A100 and H100 GPUs by comparing its hardware-accelerated results against a RaBitQ baseline intentionally restricted to a single-threaded CPU execution. This controversy follows the high-profile promotion of TurboQuant as a breakthrough capable of reducing KV cache memory by 6x without accuracy loss.
In a formal statement, Cheng Long asserted: "RaBitQ baseline was run on a single CPU with multi-processing disabled... the authors claimed efficiency in their public paper without clearly disclosing this experimental setup." The researchers further allege that TurboQuant lacks sufficient theoretical disclosure regarding its use of random rotations, a core element they claim was derived from RaBitQ without proper attribution. While the TurboQuant team reportedly agreed to address some issues after the ICLR 2026 conference, the RaBitQ authors argue this delay is insufficient to correct the widespread market misunderstanding that has already impacted semiconductor sector sentiment.
Sources (10)
Google's new TurboQuant algorithm speeds up AI memory 8x, cutting costs by 50% or more | VentureBeatTurboQuant: Online Vector Quantization with Near-optimal Distortion RateCan Google's AI Memory Compression Algorithm Help Solve the RAM Crisis? | PCMagASCII News - Swipe Through Tech News - ASCII.co.ukScraper SpiderTURBOQUANT: ONLINE VECTOR QUANTIZATIONCnbc | CNBC: market informationTurboQuant is a new algorithm from Google Research ...
Related Companies
Nvidia
NVDA