Bimos

I would assume Guilin Noodles are cousin of Vietnam Pho, based on the similarity of their ingredients and their geographic positions.

Maybe add Chimera as well?

https://arxiv.org/pdf/2107.06925

I heard that it is possible to achieve better performance than cuBLAS using CUTLASS? I thought they chose the better one among cuBLAS and CUTLASS as baseline.

I edited it.

I think most AI players rely on high performance GEMM. But most people would be satisfied with cutlass or cublas, and the others implement gemm themselves, but not necessarily use undocumented features?

> FFMA SASS interleaving

> We observe a performance improvement in the CUTLASS FP8 kernel between NVCC 12.2 and 12.3. By comparing the compiled SASS, we discover that one bit in a series of FADD instructions is flipped in an interleaving pattern. After referencing some open-source CUDA assembler implementations, we identified that this bit controls yield, which may enhance warp-level parallelism (just a guess, yielding the current warp and let other warps work).

> To leverage this, we develop a similar script to modify the FFMA instructions in the compiled binary. Besides simply modifying the yield bit, we also flip the reuse bit (registers cannot be reused if the warp is yielded). This adjustment improves performance (10%+ in some cases) for fine-grained scaling FP8 GEMMs by creating more opportunities to overlap MMA instructions with promotion FFMA instructions.

I would say it is really mind-blowing.

The PTX instructions they talked about in the tech report should be pointing to the code here?

How is it different from eliminating y?

As a Chinese I want to mention that foreign people don't usually call it "PPT".

Let's see how US reacts to this incident in Munich meeting

> If GPUs Are So Good

Who ever claimed that?

> In Shanghai, the Huangpu River divides old and new.

Pudong is new, but Puxi is not "the old one". It is mixed with old historical sites and modern buildings. It lives with its history, and it grows actively. It doesn't have to destroy all the old stuff, but it also doesn't have to slow down its pace for them. I believe it is the same for any city (e.g. in Europe) with history.

button^2

I'm a native Chinese speaker. I knew Detexify and used it a lot. This is the first time I see this software. I tried with hand written \succeq, and until its anti-crawler mechanism is triggered, it fails to give the answer. You can argue it is a software with a different purpose (e.g. to convert a piece of content rather than a single symbol), but to me it is not "even better" than detexify.

> just an hour or so up the river

Bullet train takes nearly 2 hours. So I assume it would take about half a day by boat. But still good point.

For a country with nearly one billion poor people, one will never make all of them rich with one step. That is why China "allows some peasants to get rich first", and why they say "Common Prosperity Does Not Mean Equal Wealth" and "Those Who Become Prosperous First Helping Those Who Lag Behind". Although it is truly questionable if those get rich first want to help those lag behind, with a big government, it is more or less realized.

> very old people are still in power

Just out of curiosity, is there any practical reason to send human to moon considering costs and risk?

It is normal (at least I should consider it normal), but not common (or should I consider it common?). We learn "apple, pear, banana" long before we learn "papaya, jujube, kiwi-fruit".

I agree that using `JUN`s makes it more clear as a non-native English speaker. What can we do without refering to some language? Maybe like 2024-[06]-05 or 2024-06-(05) (vs 2024-05-[06] and 2024-05-(06))?

I once saw some sentences like "My brother-in-law loves my brother" in screenshot of duolingo in social media. I think they are uncessarily confusing, especially for beginners.

Is it my problem or the author wrapped all "C"s in code format and they are awkwardly rendered in white blocks within the black background?

Good point. Also "linear convergence" means the residual reduces "linearly", or |r_{k+1}| = \lambda |r_k| with \lambda \in (0, 1). So it is somehow exponentially converging, and an algorithm with linear convergence is neat and fast.

> One problem with using this to draw 字 is the lack of vertical strata.

Totally agree. Adding the number row on the keyboard should be helpful but make QWANJI less like a language.

I tried to draw some hanzi (Kanji).

rsdcygjnbm ebycdomhl

    你        好

It is like written by someone really drunk though...

But the stroke order is correct and I would say native speakers (like me) can recognize them.

On GPU, it can only solve small instances which one single GPU thread can handle, and can only benefit from GPU by solving multiple instances, right?

The feasible region {x | f(x) = 0} is nonconvex no matter whether f is convex.

Thank you!

Thanks for your information. I think it really bridge the gap between the people who are interested in this algorithm and MILP "users". I have two more questions.

1. Usually we deal with models with both integer and continuous variables (MILP). Conceptually B&B tackles ILP and MILP in similar ways. Is there any difficulty for lattice based method to be extended to solve MILP?

2. How likely do you think this lattice type algorithm will overcome the difficulties you mentioned and eventually replace B&B, totally or partly (like barrier vs simplex methods)?

I have a dumb question: how long will it take before this result becoming a pratical MIP solver beating SCIP or gurobi?