OpenAI engineers question V4 hardware recommendations one by one: The chapter where V3 amazed the industry is once again "unexpected"

According to Beating Monitoring, OpenAI engineer Clive Chan stated that the V4 technical report remains top-tier overall, but the hardware recommendation section for chip manufacturers was “surprisingly mediocre or even flawed,” contrasting with V3. The hardware section’s Q&A in V3 was once the hottest discussion topic at the academic conference ISCA, with recommendations specifically addressing industry-standard interconnect protocols; V4, however, is much more vague.

Chan raised questions point by point. Regarding power consumption, the report states that software optimizations cause the chip’s computation, storage, and communication to run at full capacity simultaneously, recommending that chip manufacturers reserve more power headroom. Chan believes this is “just the opposite”: the total power consumption of the chip is limited by physical processes, and leaving more power margin means lowering the operating frequency, which actually reduces computing power. Concerning data transfer methods between GPUs, the report suggests choosing a pull model, where GPUs actively read data, rather than a push model, where data is pushed to them, because push notification overhead is too high. Chan questions this judgment, arguing that pull is actually slower and that network card data processing capabilities should be improved. However, they might be discussing different levels of the problem: the report refers to notification mechanism overhead, while Chan refers to the transmission delay itself. Regarding activation functions, the report recommends replacing SwiGLU with simpler functions to reduce computational load, but Chan believes this is unnecessary, as Sonic MoE has already demonstrated that SwiGLU can still achieve optimal performance. Chan suspects DeepSeek may have “deliberately weakened this section.”