ChinaJD.com says it cut ad inference latency 15% with Huawei.
JD says generative ad models need about 1,000 times more compute, and the bottleneck is general-purpose processors and memory, not graphics chips.
A Huawei Kunpeng supernode links machines into a 300TB shared memory pool, replacing four cache tiers with two.
The 15% figure is JD’s and Huawei’s own account, given at Huawei’s Kunpeng industry summit.
How each outlet framed it
- TMTPost
- details JD's 300ms ad latency budget and the 1000x compute surge from generative recommendation models
Sources: TMTPost