ryao

23 ms·

331.
by ryao 1y ago
Last year, I took the time to read through public documents and estimated that their annual production was limited to ~300 wafers per year from TSMC. That is not Nvidia level scale, but it is scale. There are many companies that sell tokens
332.
by ryao 1y ago
Thanks for the correction. They are currently using FP16 for inference according to OpenRouter. I had thought that implied that they could not use FP8 given the pressure that they have to use as little memory as possible from being solely r
333.
by ryao 1y ago
I did the math last year to estimate how many wafers per year Nvidia had, and from my recollection it was >50,000. Cerebras with their ~300 per year is not able to handle the inference needs of the market. It does not help that all of th
334.
by ryao 1y ago
Cerebras’ technology works by using an entire wafer as a chip and power draw is 23kW if I recall correctly. Their technology cannot be scaled down and only works when scaling up. They could not be more useless for Apple’s purposes. Acquirin
335.
by ryao 1y ago
The memory bandwidth for that is 150GB/sec. Inference speed is memory bandwidth bound, so that memory is useless for inference. Discrete GPUs will run circles around the CSE-3 at inference if they tried using the external DRAM.
336.
by ryao 1y ago
Their CEO is a felon who plead guilty to accounting fraud: https://milled.com/theinformation/cerebras-ceos-past-felony-... Experienced investors will not touch them: https://www.nbclosangeles.com/news&#
337.
by ryao 1y ago
> At over 2,500 t/s, Cerebras has set a world record for LLM inference speed on the 400B parameter Llama 4 Maverick model, the largest and most powerful in the Llama 4 family. This is incorrect. The unreleased Llama 4 Behemoth is th
338.
by ryao 1y ago
Unfortunately, Microcenter’s selection often leaves me disappointed. In particular, they stock almost no ECC RAM for AM5 machines. They do not stock many graphics cards that have ECC VRAM either. Their best card in stock that has ECC VRAM i
339.
by ryao 1y ago
I thought the purpose of that was to let them avoid price matching on certain items despite having price matching policies. I have never heard of one budging on this. Have any?
340.
by ryao 1y ago
Stores often refuse to stock products unless they are given a comfortable enough margin on them that can be as high as 25%. That is how they can afford to do deals on CPUs to draw people into the store. Amazon also has overhead in shipping
341.
by ryao 1y ago
If it is not broken, do not fix it. Renovations would cost money, which would mean higher prices. It is better to stick with what works than see prices rise to cover pointless renovations that would harm their competitiveness. The money is
342.
by ryao 1y ago
The other guy’s remarks were based on the behavior of ancient compilers. I was describing current ones. In any case, my remarks mostly apply to LLVM too. GCC and LLVM are the only compilers that matter these days. Intel replaced ICC with a
343.
by ryao 1y ago
If every function is marked noexcept, does it make a difference? Either way, the point of C++ exceptions was to make the fast path even faster by moving error handling out of it. Since they had no idea what was wrong, the code evidently was
344.
by ryao 1y ago
Usually, people with higher IQs are the ones who have trouble with such things. There is a logical difference between “before lunch” and “before lunch that day”. This is similar to how higher IQ individuals see the Monty hall problem as hav
345.
by ryao 1y ago
There is nothing wrong with getting a paycheck to be able to eat. I spent a roughly 2 year period after college between 2012 and 2014 where I had no job and just worked on OSS exclusively without any compensation beyond nerd credit (and a s
346.
by ryao 1y ago
Most of my OSS work has been in OpenZFS. I got started on it because of a mix of a hardware failure and Microsoft’s incompetence. Back in college, I had a Windows Media Center machine that had a second drive for recordings. The recording dr
347.
by ryao 1y ago
My OSS contributions are predominantly bug fixes with very little feature work. My primary motivation is either a need I have to fix an issue for myself or the pure intellectual challenge of tackling an issue. I have not had many bugs bothe
348.
by ryao 1y ago
What you wrote about C versus C++ is largely untrue. C++ is not faster than C, even when “using library functions which use void pointers”. There is nothing about a void pointer that prevents inlining in either C or C++. There is also no “o
349.
by ryao 1y ago
Whatever they have their interpreter doing could be done via assembly code without a separate instruction stream that needs to be interpreted. It is like running qemu-user to execute a program by interpreting it versus having the CPU execut
350.
by ryao 1y ago
> If the user has traveled to a different timezone, the time is now being converted to a localtime in the past that is different than it actually was when it happened. I consider this to be a desirable behavior.
351.
by ryao 1y ago
It is unclear to me. It has been ever since I was a child. This is what happens with periodicity. There is no 1 lunch reference point when you just say lunch in this sort of question.
352.
by ryao 1y ago
I always felt the answer to this question is both, since there is always yesterday and tomorrow as far as lunch is concerned, so everything by definition is both before and after lunch. Perhaps the question is ill formed and needs to ask if
353.
by ryao 1y ago
If you record them with some known UTC offset regardless of localtime (e.g. 0), you can convert to local time in the past.
354.
by ryao 1y ago
I find their use of an on-GPU interpreter to be both a bit of an odd choice and interesting at the same time. Usually, you would not want to use an on-GPU interpreter for anything involving high performance. However, it sounds to me like th
355.
by ryao 1y ago
> It can be used to secure your SSH service, or securely share your database server. SSH is one of the most secure network daemons ever devised. This is not to say that there is never any need to harden SSH, but given that people usually
356.
by ryao 1y ago
After presenting their numbers, they mention that CUDA graphs also do much of this, but then say that the launch time is higher for them. It would have been more interesting if they had included comparison numbers. Without numbers, I am lef
357.
by ryao 1y ago
I believe that both Windows Services for UNIX (Interix) and OS/2 application support were NT subsystems too. I am under the impression that Windows Services for UNIX was the foundation for Astoria.
358.
by ryao 1y ago
Here is my take on this from 12 years ago after following the advice of a stack overflow post: https://github.com/gentoo/genkernel/commit/a21728ae287e988a1... With that (minus the gen_die() line unless you co
359.
by ryao 1y ago
Whenever someone misuses the word mad, I think of this: https://www.youtube.com/watch?v=3Ao4WTcMtY8
360.
by ryao 1y ago
https://en.wikipedia.org/wiki/Windows_Subsystem_for_Linux#Hi... It was called Project Astoria previously. Microsoft releasing the Windows Subsystem for Android for Windows 11 is news to me. I thought that they had kill

More ›