CONFIDENTIAL · CONFIDENTIALCONFIDENTIAL · CONFIDENTIAL
Recording timestamp 03:41:02

Model Size Benchmark and the Efficiency Leap from Compiled Languages

Confirms that compiled languages like TileLang bring efficiency gains, not losses, with hardware execution efficiency loss of only 1% to 2%.

Yes, it's a huge efficiency gain. So this is an opportunity.
Investor question

Second question, you mentioned earlier that we're now using some compiled languages other than CUDA for inference. As I understand, we used to rely on the NVIDIA ecosystem, probably using PTX a lot. Now that we're using more of the TileLang you just mentioned, will that significantly reduce our inference efficiency, or at least in the short term? I'm not sure how you view the efficiency loss from this shift to compiled languages, or whether in the long run it's actually a supplement that improves efficiency?

Liang Wenfeng responds

It improves efficiency, significantly improves efficiency.

Investor question

OK, so no negative impact?

Liang Wenfeng responds

Yes, it's a huge efficiency gain. So this is an opportunity. It's like before, you couldn't leave the CUDA ecosystem, but now we can abandon it and use a simpler approach with TileLang. It's a high-level language, writing programs is fast, and the codebase is small; we can rewrite it.

Investor question

I see. So both of these are actually, as you mentioned, a big opportunity brought by AI, not a shortcoming that needs to be addressed.

Liang Wenfeng responds

Yes, it's a big opportunity for technological development. It's not about AI opportunity, because we also have a project where we're using AI to write TileLang.

Investor question

Would that be even faster?

Liang Wenfeng responds

Right now, all TileLang is written by humans, but it's already much faster than writing CUDA was before.

Investor question

I see. So even for low-level hardware execution efficiency, you think there's no impact?

Liang Wenfeng responds

A loss of 1% to 2%, I think that's acceptable.

Investor question

OK, understood. Alright, thank you, that's very clear.

Anyone else have questions? No? Let's wrap up for today. OK, thank you. Thank you all very much for your time. Thank you, Mr. Liang. Thank you. Bye.

Liang Wenfeng responds

Bye.

Open this chapter in the guided reader →
This transcript was produced by automatic speech recognition and edited with AI. Speakers are not separately labelled; chapter titles, summaries and the argument map are editorial aids. Names and figures may contain recognition errors — refer to the original recording. Liang Wenfeng noted during the meeting that some figures are sensitive; please do not redistribute.