Question The AVX-512 thread

[Tuna-Fish]

What happened in ONE month???

The x86 ecosystem advisory group convened for the first time. I'd bet some money that all the user side people on it just spent their time dunking on ecosystem fragmentation.

 

[Tuna-Fish]

The graph does make it look like Takum Arithmetic might become the one floating point standard to rule them all.

Huh? It just throws away most of the range of ieee754 double for better precision?

 

[igor_kavinski]

Huh? It just throws away most of the range of ieee754 double for better precision?

Isn't that what we want for floating point type?

 

[Tuna-Fish]

Isn't that what we want for floating point type?

Except when we want range? In FP types, you trade off range and precision. Takum decided that 10^77 was both desperately needed, and enough for everybody. I find both of those arguments specious.

 

[igor_kavinski]

Takum decided that 10^77 was both desperately needed, and enough for everybody. I find both of those arguments specious.

Interesting. So you are disagreeing with the following part of the paper? (just asking)

This study is particularly significant as it presents a
potential paradigm shift in computer arithmetic. Given
further research and evaluation, it suggests a future where
both low- and high-precision computations could be uni-
fied under a single number format, thereby substantially
simplifying hardware design. While such a transition may
seem radical, our findings, like many others, highlight a
key insight: IEEE 754 floating-point numbers are poorly
suited for modern low-precision computations.

Curious to know if this is some breakthrough or if the author is high on something.

 

[Schmide]

It's a troubled article.

You can get better precision so many other ways than the use of logarithms. How do you think those logarithms are calculated. Psst it's cordic or taylor and it's done with floating point. It's also one of the slowest operations on a computer.

I also find this funny

While, at the time of writing, no AMD processor supports AVX10

psst. No intel processor or any other processor supports it either.

The whole premise seems to be, at the same bitness takum numbers have slightly better precision. Ignoring the fact that you need extra bits to store the base of the logarithm used in the system. That base is going to have to float to represent all the numbers needed to do basic calculations. Guess what. We've gone full circle and are now back at floating point, just with a different system.

If there is one thing AI has taught us. If the system is set up right, you can get away with absurdly low precision numbers as long as you have a bunch of them.

If you're looking for precision, just use more bits or operands. There are quad float programs out there that while slow would run circles around any logarithm based system.

The real irony in the paper. There is no hardware logarithm in any simd instruction. They have a sequence. But that's just a software operation that is going to take as long if not longer.

Intel quote: This intrinsic generates a sequence of instructions, which may perform worse than a native instruction. Consider the performance impact of this intrinsic.

 

[soresu]

psst. No intel processor or any other processor supports it either.

Actual footage of Schmide throwing shade at Intel:

 

[soresu]

Latest version of the AVX-10 spec is out, and it drops the 256-bit only implementation.

That is, AVX-10 will in the future just be a successor to AVX-512, with matching vector width and full support for all existing AVX-512 code.

What happened in ONE month???

Likely a combination of negotiations/talks within the new x86 ecosystem group wanting to simplify the AVX10 standard, and the fact of AMD implementing AVX512 fairly efficiently (if not as performantly) with less beefy ALUs in some of its Zen5 CPU cores.

Intel is probably aiming at following AMD's lead here with a future *mont µArch to implement AVX10 across its entire lineup with the same code.

After all the fragmentation of the AVX512 era, this is no doubt an attractive prospect.