Linus Torvalds: "I Hope AVX512 Dies A Painful Death"

[dxin]

Nooooo! Don't take away my Intel GPUs!

iGPU is useful for most of the people but does it have to take up half the area on the chip? Do you guys actually need 3D?

[piotrj3]

How would you compute sqrt(17) with 1000 decimals using your double variables? You are basically making a statement that is as intelligent as saying that it's glue you use as fuel for your car.

Floating point is named as it is, because it has a mantissa and an exponent - and a floating point register is designed to store an *approximation* of a number. In some specific cases, it manages to store the exact value. Most times, not.

Integer arithmetic - as in fixed point - can handle the value 1.33333934565634382349214134124185194999911 exactly. No round off. No unexpected "accident".

In many situation, it's impossible to use floating point just because of the reverse of your claim - the floating point registers can't maintain the precision needed. They can't store the exact values. And if you have a very numerically sensitive algorithm, it may explode from the lack of precision.

Ever wondered why the world has big number libraries, when we already know how large values you can fit in a 64-bit or maybe 80-bit floating point register?

Sorry to tell you that, but there are solutions for double being not precize enough. It is mostly known as triple/quad precision and it is not even that hard to implement.

https://blogs.mathworks.com/cleve/2017/05/22/quadruple-precision-128-bit-floating-point-arithmetic/

Of course big numbers libraries do exist for a reason, but integer like fixed point has so few usages if you know quad precision, that it is essentially pointless.

[zyxxel]

Sorry to tell you that, but there are solutions for double being not precize enough. It is mostly known as triple/quad precision and it is not even that hard to implement.

https://blogs.mathworks.com/cleve/2017/05/22/quadruple-precision-128-bit-floating-point-arithmetic/

Of course big numbers libraries do exist for a reason, but integer like fixed point has so few usages if you know quad precision, that it is essentially pointless.

First off, you shouldn't start sentences with "sorry to tell" - the majority of people who start a response with that phrase are most definitely not sorry. They think they have a smashing answer. Well, your answer isn't a smasher.

Quad precision doesn't get you very far. It has eps 1.9259e-34

That is rather far away from a thousand decimals. 1000 decimal digits is over 3000 bits.

It's quite few situations where double precision is too small while quad precision is enough. Most of the time, you either manage with float/double or you need really much more - and correct - digits.

[kcrudup]

To get AVX-512, you need either a modern Xeon, or their HEDT chips. [...] It is essentially impossible to put together a system for testing AVX-512 utility without putting down quite a chunk of change

Not sure if this is covered in a later comment but I don't consider the i7-1065G7 IceLake CPU in my laptop all that special, but FWIW:

$ tr '\040' '\012' < /proc/cpuinfo | sort -fu | fgrep -i avx
avx
avx2
avx512_bitalg
avx512bw
avx512cd
avx512dq
avx512f
avx512ifma
avx512vbmi
avx512_vbmi2
avx512vl
avx512_vnni
avx512_vpopcntdq

... that being said:

AVX-512 is also a random grab-bag of confusion - and it is "one" set of instruction set extensions!

This. I have no idea what all those AVX512 variants mean or differentiate between another.

[cynyr]

Not sure if this is covered in a later comment but I don't consider the i7-1065G7 IceLake CPU in my laptop all that special, but FWIW:

This. I have no idea what all those AVX512 variants mean or differentiate between another.

Looking at the wikipedia page for AVX512, not all CPUs that support AVX512, support all instruction groups. I'm not even sure that support of all instructions within a group is mandatory.

https://en.wikipedia.org/wiki/AVX-512#CPUs_with_AVX-512

[Paradigm Shifter]

Not sure if this is covered in a later comment but I don't consider the i7-1065G7 IceLake CPU in my laptop all that special, but FWIW:

Yes, I realised after I posted that there are a minor selection (there's a one-shot 10nm dual-core i3 that only came in a few laptops and an Intel NUC which has some, IIRC.

That said, they're barely fit to set up a testing box. I could compile programs on them, but wouldn't be able to really do anything worthwhile due to core or RAM limits. ;(

[szt1980]

Just as if anyone would ever care a slightest bit about some random rant a clown with a finger is producing