OpenAI Technicians Critique V4 Hardware Recommendations: A Surprising Shift from V3's Industry Impact

According to monitoring by Beating, OpenAI technician Clive Chan stated that the overall V4 technical report remains top-notch, but the hardware recommendations for chip manufacturers are “surprisingly mediocre and even erroneous,” contrasting sharply with V3. The hardware section of V3’s report featured a Q&A that was the most popular discussion at the academic conference ISCA, with recommendations that were specific to industry standards being developed for interconnects, whereas V4 has become much more vague. Chan raised several points of contention. Regarding power consumption, the report claims that software optimization allows chips to run computation, storage, and communication at full capacity simultaneously, suggesting that chip manufacturers reserve more power headroom. Chan argues that this is “counterproductive”: the total power consumption of a chip is limited by physical processes, and reserving more power headroom would mean lowering the operating frequency, thus reducing computational power. On the topic of data transmission between GPUs, the report suggests that GPUs should actively read data (pull) rather than have it pushed to them, due to the high overhead of push notifications. Chan questions this judgment, believing that pull is actually slower and that the data processing capabilities of network cards should be improved. However, the two may not be discussing the same level of issues: the report addresses the overhead of the notification mechanism, while Chan is concerned with the latency of the transmission itself. Regarding activation functions, the report recommends replacing SwiGLU with simpler functions to reduce computational burden, but Chan believes this is unnecessary, as Sonic MoE has already demonstrated that optimal performance can be achieved using SwiGLU. Chan suspects that DeepSeek may have “deliberately downplayed this section.”