shihab

In one view, the fact that it's a software problem is actually a weakness of (GPU) hardware design.

In the olden, serial computing days, our algorithms were standard, and CPU designers did all sorts of behind-the-scene tricks to improve performance without burdening software developers. It wasn't perfect abstraction, but they tried. Algorithm led the way; hardware had to follow.

CUDA threw that all away, exposed lots of ugly details of GPU hardware design that developers _had to_ take into account. This is why, for a long time, CUDA's primary customers (HPC community & Natl labs) refused to adopt CUDA.

It's interesting that now that CUDA has become a legitimate, widely adopted computing paradigm, how much our view on this has shifted.

> Iran is the principle destabilising element in the middle east

Says Israel, the nation who tore up every single international laws, directly led campaign against UN and ICC, and whose right-wing (ones in power now) have been dreaming about a Greater Israel that threatens territorial integrity of like 10 different ME countries.

I understand Iran is a headache to Israel, but did it have to be an enemy of USA? Isn't Iran's ambition, and its proxies, are all regional in nature? Have they ever attempted to harm an american living in America?

Israel has led an amazingly succesful campaign in presenting their problems (often arising out of their territorial ambitions) as a problem for the entire west.

Her statement directly contradicted her testimony. After recent Trump's open dismissal of her remark, she had to say this to keep her job.

The source is mossad, in case anyone gets fooled by the presence of a citation like me.

This strike didn't happen to protect Americans from nukes, this happened to protect a rogue politician who was about to be impeached by his countrymen, and to make the Greater Israel project come true.

Reminder, a recent survey found 16% American supported an offensive strike against Iran.[1]

[1] https://www.axios.com/2025/06/19/israel-iran-war-americans-p...

Even if Iran were arming regional proxies, that's an Israel problem, not an america problem. Though AIPAC et el makes sure no american is ever aware of that distinction.

opinion like this can only come from someone who have lived a life of privilege, thousands mils away from systematic violence or hunger someone in a war torn country constantly have to face.

For an average individual going about their day, lack of political freedom is a 100x better option than lack of food or security that Iraq (or Libya or Afghanistan) went through.

Will it be widely implemented? This is only for repeat offenders, which the article says would affect less than 2% drivers.

Also, glitch does not look like a big problem, since for now the system will only verbally warn, just once.

What exactly did you mean by "globalists"? who do you refer to by "they"?

Sorry if I got your sentence wrong, but terms like these are routinely used by some truly awful anti-semites, and I sincerely hope Pro-palestinian folks stay clear of those people and their vocabularies.

"Meta has complied with 94% of takedown requests issued by Israel...Meta removed over 90,000 posts to comply with TDRs submitted by the Israeli government in an average of 30 seconds...All of the Israeli government’s TDRs post-October 7th contain the exact same complaint text, according to the leaked information, regardless of the substance of the underlying content being challenged. Sources said that not a single Israeli TDR describes the exact nature of the content being reported"

As someone who came from a pretty authoritarian country- let me assure you that people there do routinely criticize their government, mock them all the time. Governments often do not have the bandwidth to deal with the volume of criticism, and even when they do- they wisely realize that letting people vent a little online is better than complete crackdown. I myself routinely did this in Facebook, where many in my friend list were government employees and (ex-ruling) party members.

I am in fact far more afraid of pro-palestine speech from USA as an immigrant than I was in my home country- and please trust me I am not exaggerating here.

As a recent example, the instagram of guardian journalist Owen jones (well known Israel critic) was suddenly suspended without any explanation today.

It has been since restored, after a predictable twitter storm.

And here is an article on Raffi berg, BBC’s Middle East editor:

https://www.dropsitenews.com/p/bbc-civil-war-gaza-israel-bia...

The NYT's Executive Editor Joe Kahn is the son of a billionaire who was on the board of lobby group CAMERA, a group devoted to pressuring US media to be more pro-Israel.

I went through the article and it seems you're right about the comparison with El Capitan. These performance figures are so bafflingly misleading.

And so unnecessary too- nobody shopping for AI inference server cares at all about its relative performance vs a fp64 machine. This language seems designed solely to wow tech-illiterate C-Suites.

In no modern times did one side complete erase the other side. Hell, even 2000 years ago many regular romans were disgusted by the brutality of roman soldiers against Carthage.

Israeli violence is unprecedented in several objective metrics: in pounds of bumbs dropped, percentage of civilians death, percentage of civilian household destroyed (on par with Dresden now).

Wait, did they get their own library name wrong? CK should be Composable Kernel, I can’t find anything called compute kernel anywhere

Today I came across an interesting case where 3 well-known LLMs (O1, sonnet 3.7 and Deepseek R1) found a "bug" that actually didn't exist.

Very briefly, in a fused cuda kernel, I was using thread i to do some stuff on locations i, i+N, i+2*N of an array. Later in the same kernel, same thread operated on i,i+1,i+2. All LLMs flagged the second part as bug. Not the most optimized code maybe, but definitely not a bug.

It wasn't a complicated kernel (~120 SLOC) either, and the distance between the two code blocks was about only 15 LOC.

Let me say as an international PhD student- I have marched for the freedom of Palestinian people, talk about it, and I won’t allow this Big Brother tactic influence my decision one bit going forward.

The US govt can do whatever it damn pleases.

I’d really like to see vibecoding being used in something other than a crud app.

I do CUDA development. Forget writing whole algorithms, LLMs can’t find common bugs in a 12-lines kernel.

They mean $5.45 earning for every dollar spent. So around 85% profit margin.

Some very illuminating stats from the article:

“ Over the past 24 hours (UTC+8 02/27/2025 12:00 PM to 02/28/2025 12:00 PM), the combined peak node occupancy for V3 and R1 inference services reached 278, with an average occupancy of 226.75 nodes (each node contains 8 H800 GPUs). Assuming the leasing cost of one H800 GPU is $2 per hour, the total daily cost amounts to $87,072…

If all tokens were billed at DeepSeek-R1’s pricing (*), the total daily revenue would be $562,027, with a cost profit margin of 545%. However, our actual revenue is substantially lower for the following reasons…”

I wonder how many Americans know about the night when Kissinger et al activated entire US military, incl nuclear forces, just so some soviet weapons wouldn't reach Egypt during Israel's Yum kippur war- all while Nixon slept. Soviet union was so taken aback, they immediately folded. Of course they would, they treated this thing like a minor regional conflict thousands miles away from home. Their mistake was believing American govt would do the same.

Source- Kissinger: A Biography by Walter Isaacson

61% Americans opposed sending weapons to Israel. They got sent anyway [1]. Similarly at world stage, American govt had to bring out the veto card 49 times in Israel's defense- it was alone against the world, even all of it's strongest European allies [2].

[1] https://www.scribd.com/document/740568401/Cbsnews-20240609-S.... Question No 52.

[2] https://www.middleeasteye.net/news/49-times-us-has-used-veto...

The speedup figures they report are compared to their own cutlass-based baseline. Has anyone done a performance comparison against cuBLAS?

All cutlass results I have seen so far for Gemm are within ~10% of cuBLAS. If the 2x-2.5x speedup they report holds up that would be extremely impressive.

For future readers, note that those 3x and 10x figures are compared to vLLM's own previous release, and NOT compared to Deepseek's implementation.

I am very curious to see how well-optimized Deepseek's code is compared to leading LLM serving softwares like vLLM or SGLang.

The biggest victims of these non-scalable process is people without a good network. As an intl PhD student, I am that person.

So now I have this weird dynamic: I get interview calls only from FAANG companies, the ones with the manpower to do your so called "cursed" scalable interviews. But the smaller companies or startups, ones who are a far better fit for my specialized kills, never call me. You need to either "know someone" or be from a big school, or there is zero chance.

I recently saw Sam Altman bragging about OpenAI's performance on Codeforces (leetcode-like website), which I consider just about the worst benchmark possible.

1. All problems are small- the prompt and solution (<100 LOC, often <60LOC)

2. Solving those problems is more about recollecting patterns and less about good new insights. Now, top level human competitors do need original thinking, but that's only because our memory is too small to store all previously seen patterns.

3. Unusually good dataset- you have tens of thousands of problems, each with thousands of submissions, along with clear signals to train on (right/wrong, time taken etc), a very rich discussion sections etc.

I think becoming 100th best Codeforces programmer is still an incredible achievement for a LLM. But for Sam Altman to specifically note the performance on this- I consider that a sign of weakness, not strength.

From the creator of Triton (OpenAI)-

"PS: The name Triton was coined in mid-2019 when I released my PhD paper on the subject. I chose not to rename the project when the "TensorRT Inference Server" was rebranded as "Triton Inference Server" a year later since it's the only thing that ties my helpful PhD advisors to the project."