Meituan's cluster setup for large-scale training treats each dual-die 910C package as two separate logical accelerators, which explains why the headline figure lands at 50K dies even though the physical hardware count sits closer to 25K packages spread across multiple CloudMatrix 384 pods.
Underappreciated detail of Meituan claims: as far as I can tell, Yunfan is correct, they run CloudMatrix 384 pods in "910B mode". So it's not 50K GPUs, it's "only" 25K. One 950DT SuperPod with 8192 chips, especially if it's good for fp8, likely outclasses this rickety system.
@teortaxesTex Also note the cross die bandwidth is just 270 GBps so you might actually prefer to let the two dies to function as separate processors.
A single CloudMatrix 384 pod holds 384 physical packages wired in an all-to-all mesh, yet the dual-die design inside each 910C instantly doubles the logical accelerator count to 768 per pod.
Public details stop at the reconciled die-versus-package numbers, leaving the precise total pod count, achieved FLOP, and any sourcing specifics for the chips unaddressed in available reports.
Negative users voiced disillusionment over Meituan's CloudMatrix 384-pod cluster running at reduced scale with only 25K packages instead of 50K dies, as the smaller numbers make the deployment far less impressive.
No Digg Deeper questions have been answered for this story yet.
On the 910C: it's 2x 910B dies. So unclear if 50k 910Cs or 25k 910Cs (i.e. 50k dies).
Either way, these are not domestically manufactured. The dies/chips came from TSMC via the Sophgo incident and the HBM is probably smuggled.
Probably the biggest non-Nvidia pre-training run in China. ≈1e25 FLOP (≈DeepSeek v4 Pro or Qwen3 Max).
50k+ "AI ASICs." Probably Huawei's CloudMatrix-384 superpods with 910Cs (~40 to 80MW).
We're finally seeing data centers with the illicitly procured AI chips from TSMC.
https://www.reuters.com/technology/tsmc-could-face-1-billion-or-more-fine-us-probe-sources-say-2025-04-08/
@teortaxesTex 这数量打折后含金量确实变了
@teortaxesTex 这规模缩水确实有点幻灭了
@ohlennart 25K, certainly. 50K would be an overkill. And 910C has 128GB HBM but they complain about <80 GB per chip and about slow cross-die bandwidth. We've figured this out here.
Underappreciated detail of Meituan claims: as far as I can tell, Yunfan is correct, they run CloudMatrix 384 pods in "910B mode". So it's not 50K GPUs, it's "only" 25K. One 950DT SuperPod with 8192 chips, especially if it's good for fp8, likely outclasses this rickety system.
@ohlennart what makes you so sure that these werent fabbed at SMIC/CXMT?