AI 导读
根据 @ArtificialAnlys 对 Gemma 4 31B 的基准测试，Nvidia 的 Groq 3 LPX 系统在小模型快速解码方面表现出色（>3,400 tok/s）。LPX 通过使用少量（128 GB）超高速 SRAM 替代 HBM 来实现这一点。 Nvidia 提议将 LPU 与 GPU 结合，把这一速度带到前沿规模的模型上。下面来看看这可能如何实现。
正文
Nvidia's Groq 3 LPX system excels at fast decode of small models (>3,400 tok/s), according to a @ArtificialAnlys benchmarks of Gemma 4 31B. LPX does this by using a small amount (128 GB) of ultrafast SRAM in place of HBM.
Nvidia proposes combining LPUs with GPUs to bring this speed to frontier-scale models. A look at how that (might) work below.
来源：@EpochAIResearch · x.com