"Let's start with the matter we discussed before."
Zhao Wenyuan looked at the projected image of the L100 cluster already set up in Zhangjiakou.
"Although inference is basically up and running on this L100 cluster of nearly ten thousand cards, we still hope to unlock as much of its full potential as possible, break through the communication-efficiency bottleneck, and use it to train large models."
Shen Zhonghe shook his head. He felt he had made himself perfectly clear last time. "President Zhao, ambition is a good thing, but we already discussed this. TSMC's packaging parameters are locked down. When it comes to improving the L100, we really can't make something out of nothing."
"The parameters-we may already have them."
Zhao Wenyuan's voice was perfectly level, and even his expression barely changed.
"The parameters? You have them? What do you mean, 'may'?" Shen Zhonghe asked, a little startled.
Everyone else present also found themselves looking at Zhao Wenyuan.
How could he say that so calmly?
Zhao Wenyuan did not answer Shen Zhonghe's question directly, but reaffirmed that he had understood correctly. "Assuming the parameters are genuine and usable, how soon could you reverse-engineer an adaptation and deploy it on this large cluster, Mr. Shen?"
Shen Zhonghe had not quite recovered. His mouth opened, but nothing came out.
He had prepared before the meeting, but all those preparations had been for what came next "if" they could work with Icecore. As for these existing TSMC-packaged chips-
Would Hanlian really get a chance to peek under TSMC's hood?
Yet right out of the gate, Zhao Wenyuan was talking not about next-generation cards, but about the existing cluster.
If Source Intelligence had found a way, did that mean building domestic L100 clusters of a thousand or even ten thousand cards had become a challenge