Semiconductor News & Analysis Feed

The company claims its Ascend 950PR delivers approximately 2.87 times the inference performance of Nvidia's H20

As AI systems move from single-turn interactions to coordinated multiagent workflows, low-latency inference becomes increasingly important. Autoregressive LLMs generate tokens sequentially, which can limit GPU utilization and constrain throughput in latency-sensitive serving scenarios. Speculative decoding helps mitigate this bottleneck by using a lightweight model to draft future tokens, which t