🤖 AI Summary
Alibaba’s Pingtouge (an Alibaba subsidiary) has unveiled a new PPU for AI workloads that CCTV reports as outperforming NVIDIA’s A800 across major specs and being broadly comparable to NVIDIA’s high-end H20. Key hardware figures: 96 GB HBM2e (vs A800’s 80 GB HBM2e, matching H20 capacity though H20 uses next‑gen HBM3), inter‑chip interconnect up to 700 GB/s (vs A800’s 400 GB/s, slightly under H20), PCIe 5.0×15 interface (ahead of A800’s PCIe 4.0×16 and on par with H20), and a 400 W power envelope (equal to A800 and far lower than H20’s 550 W). The report positions Pingtouge as leading other domestic AI chips on many of these metrics.
The announcement matters because it signals accelerating domestic competition to NVIDIA in high‑performance AI accelerators and concrete deployments: China Unicom’s Sanjiangyuan compute center lists 1,747 signed devices totaling 22,832 cards and 3,479 P of compute, including 16,384 Pingtouge cards (1,945P) for Alibaba Cloud. If validated by independent benchmarks, the Pingtouge PPU’s mix of large HBM2e capacity, high interconnect bandwidth, PCIe 5.0 support and lower power draw could reshape data‑center topology and cost models for large‑scale training/inference in China, though HBM3 adoption and full performance/efficiency comparisons versus H20 remain important caveats.
Loading comments...
login to comment
loading comments...
no comments yet