xin Liu

[xin Liu @navyasm]

Cooperative jfr sampling JDK-8321098 in JVM is like ABS in automobile. If we walk the stack at arbitary position, it is like we slide while wheels lock up. Resuming execution is just like we release brake shortly to avoid a crash (pun intended)

[xin Liu @navyasm]

Triton is the 1st programming language I know that has 'blocking' as a language construct. Everybody can write cuda, the real change is how to optimize locality or arithmetic intensity. They will end up with blocking anyway. just think of it in the first place!

[xin Liu @navyasm]

@SharkCatSolana @PostsOfCats Yes, I do. I shark my day. Nata!

[xin Liu @navyasm]

@torontobigface 勇者愤怒，抽刃向更强者；怯者愤怒，却抽刃向更弱者。不可救药的民族中，一定有许多英雄，专向孩子们瞪眼。这些孱头们！

[xin Liu @navyasm]

and iterator won't change that. It's better to use ranges and views in C++20. we should complete computation using pipe operators.

[xin Liu @navyasm]

For dynamic containers such as stl vector, &v[i] is always a wrong idea. Unlike array, the address of underlying buffer may change.

[xin Liu @navyasm]

what???

[xin Liu @navyasm]

Java is a product. We only see a variety of vendors are selling it as technology or a solution. C++ is a region. that's why we see so many gurus are selling service package that help you digest C++.

[xin Liu @navyasm]

$699 for ps5 pro is the moment I realize we gamers need xbox series X.

[xin Liu @navyasm]

@PS5Updates No! Absolutely no.

[xin Liu @navyasm]

@wangzhian8848 朱镕基有魄力敢想敢干不假，但是他干的所谓改革都是简单粗暴的解决方案。什么分税制就是中央拿走大头儿，国企改革就是破产不救了，所谓央企改革就是搞垄断经营，救金融就是烂帐划走。有点朱可夫“军神”那意思。

[xin Liu @navyasm]

I think we are using C++ in completely wrong way. C++ is everything about curation. Battery-included is the last thing you want in C++.

[xin Liu @navyasm]

I had a lengthy discussion what's really new in the AI compiler. We concluded that nothing but name is new.

[xin Liu @navyasm]

To get the fixed amount of computation(C) done, $ = C / (perf/TCO) . The cost is directly proportional to TCO(W) and inversely proportional to perf. It's worth noting that we can't take performance for granted if we have custom design. We have to optimize the software stack.

[xin Liu @navyasm]

All internet companies want to design their in-house ML accelerators in these days. One key lesson is 3⃣: "design for performance per TCO vs per CapEx". Your "cost-efficient" design should offset nvidia's margin.

[xin Liu @navyasm]

I learn this from Ten lessons from three generations shaped Google's TPUv4i. TCO is CapEx + 3(years) * OpEX. TCO is pretty much just system TDP.

[xin Liu @navyasm]

if different passes are done by different team or personnel, it's very likely we observe miscommunication or misunderstanding. The number of bug combinatorially relates to the line of code.

[xin Liu @navyasm]

Amazon is really good at building distributed systems. They usually use microservice and two-pizza teams. It's dubious that the muscle memory will work again on an optimizing compiler. A compiler pass has implicit interactions with almost downstream passes. Law of power here!

[xin Liu @navyasm]

I think nvidia will be fine even this AI bubble busts. In the nutshell, NV is a company that sells parallelism. Parallelism is more cost efficient than concurrency. They excel in areas where they discover massive parallelism. Graphics/Crypto/ML/AI just happen to be the case.

[xin Liu @navyasm]

I just realize this inconvenient truth: I am not building a compiler. I am a compiler from the perspective of leadership. We engineers are basically hand-pilers on steroids...