EN ▾
SNAPE REP RT Front page →
DevelopingCompute20 Nov, 07:16 UTC

Alibaba to lean more on its own AI chips as US curbs threaten its overseas compute

Photo accompanying Digitimes' report on Alibaba's AI chip and overseas compute strategy.
AFP via Digitimes

Digitimes reports Alibaba plans more reliance on domestic chips as tighter US rules on overseas computing threaten access to Nvidia GPUs; Alibaba Cloud says its Zhenwu V900 goes on sale in volume in the first quarter of next year.

Read at Digitimes Share

Why it matters

The US this month put licence requirements on AI chip and server purchases by datacentre operators in Singapore, Malaysia, Thailand and Indonesia, where Chinese tech companies have driven demand. According to IT Home, the V900 has three times the compute of the Zhenwu M890, 216GB of memory and 1,200GB/s chip-to-chip links, and Alibaba says one cluster can scale to 500,000 chips. How fast the V900 ships, and at what yields, will show whether homegrown chips can fill the gap.

Key facts

  • Zhenwu V900 volume sales from Q1 next year
  • Three times the compute of the Zhenwu M890
  • 216GB memory, 1,200GB/s chip-to-chip interconnect, FP8/FP4
  • One cluster can scale to 500,000 chips
  • M890 superclusters already serve customers in Ulanqab and Zhongwei

Sources

More on Compute
More from Snape ReportFront page →