Author: Evolving AI - Bewertung: 4x - Views:22
Huawei just claimed it can get up to 1000× more effective AI performance than Nvidia’s chips—without building a single new GPU. In this video, we break down Flex: ai, Huawei’s new software platform that turns scattered GPUs and NPUs into one giant “analog AI chip” by slicing accelerators into tiny virtual pieces, packing multiple jobs onto the same card, and pushing real-world utilization from 30–40% closer to 70%. Built on Kubernetes and designed to be open source, Flex: ai lets you carve GPUs and Ascend NPUs into 10% slices, mix training and inference on the same hardware, and treat a messy cluster of different accelerators like one unified pool instead of a pile of half-idle cards. We’ll walk through how Hi Scheduler constantly watches the whole cluster, predicts workloads, and decides which jobs can safely share a device, how Flex: ai handles isolation so one noisy model doesn’t destroy latency for everything else, and why Huawei argues that long-term throughput is what really matters—not peak FLOPs on a spec sheet. We also examine how this aligns with China’s broader AI strategy, which involves utilizing older GPUs, restricted Nvidia parts, and domestic Ascend NPUs to work more efficiently together under export controls. If Huawei can squeeze this much extra performance out of existing silicon, does software like Flex: ai matter more than the next GPU generation? And if you want the real story behind the world’s fastest-moving tech and AI breakthroughs, make sure to like and subscribe to Evolving AI for daily coverage.
SOCIAL SHARE CARD GENERATOR