Linux founder tells Intel to stop inventing 'magic instructions' and'start fixing real problems'

Linus Torvalds' comments came from this article: https://is.gd/6zpZRL

His comments came in a mailing list (via Phoronix) discussing an article suggesting AVX-512 might not be part of Intel's upcoming Alder Lake architecture. If that comes to pass, it will be just fine by Torvalds.

"I hope AVX512 dies a painful death, and that Intel starts fixing real problems instead of trying to create magic instructions to then create benchmarks that they can look good on. I hope Intel gets back to basics: gets their process working again, and concentrate more on regular code that isn't HPC or some other pointless special case," Torvalds said.

Intel introduced AVX-512 in 2013, initially as part of its Xeon Phi x200 and Skylake-X processor lines. It has also found its way into more current CPU architectures, including Ice Lake.

The instruction set is designed to bolster performance in various types of workloads, such as scientific simulations, financial analytics, artificial intelligence, data compression, and other tasks that can benefit from more robust floating point operations.

Nevertheless, Torvalds views AVX-512 as an example of "special-case garbage," noting that in regards to floating point performance, "absolutely nobody cares outside of benchmarks."

"I absolutely detest FP benchmarks, and I realize other people care deeply. I just think AVX-512 is exactly the wrong thing to do. It's a pet peeve of mine. It's a prime example of something Intel has done wrong, partly by just increasing the fragmentation of the market," Torvalds said.

I think he's absolutely right, and previously we didn't see how much
Intel was wasting its time making these AVX instructions because it's
gaping security flaws were not yet known. We just assumed that the more
sophisticated these floating-point instructions got, the more power they
must draw naturally. But previous generations of FP instructions stayed
well within the power envelope of the processor, whereas these AVX
instructions have been known to go well outside the standard power envelope.

Yousuf Khan

Yousuf Khan wrote:

Linus Torvalds' comments came from this article: https://is.gd/6zpZRL

Full URL:
https://www.pcgamer.com/linux-founde...al-problems%2F

Linus is known for publishing his tirades on Windows, and even on Linux
variants. He lambasts everyone.

Tweaking hardware to look good in benchmarks is news to you? Video chip
makers have been doing this forever, making their hardware or firmware
look better in particular benchmarks (sometimes their own benchmarks
tweaked for their hardware) but for which the benchmarks have no
practical implementation illustrating actual performance in real use.

AVX wasn't just about improving FP instructions. The number of cores
available back then was maybe up to 4 allowing concurrent thread
processing. With more cores to parallelize the computing, AVX becomes
less necessary. The latest CPUs (although far outside the consumer
price range) have 64 cores, maybe more. Sorry, but bitching in
hindsight is the easy way to look superior. I don't see Linux bitching
back *then* when AVX showed up. His forward-looking crystal ball was
just as cloudy as everyone else's. So, how many cores were in your home
computer back in 2013 when AVX came out? AVX isn't just about upping
the bit-width of FP calculations, but also about parallelization. How
many desktops nowadays have any apps on them that can use all 4 cores?

Not all CPUs are waiting to do something for end users. Some are
involved in highly complex computing, like animated computer graphics.
You think Zootopia was composed on a home computer? So, you think Intel
(or AMD) are going to tool up for a completely separate production line
for consumer vs high-graphics design platforms? There is an economy in
production by reusing existing manufacturing processes. Do consumer
platforms utilize AVX? Rarely. Why didn't Linus bitch when Intel added
Streaming SIMD Extensions (SSE)? How about all those non-gaming users
that don't care even about the old SSE extensions? Oh my God, the CPU
has something they don't need.

I suppose next Linus will bitch about increased parallelization in
Mozilla's Firefox. The next engine, Servo, takes advantage of the
memory safety and concurrency features of the Rust programming language.
Servo will use parallelism by using more cores for the rendering engine,
layout, HTML parsing, image processing, decoding, and other tasks that
can be isolated (into separate processes or threads to run on more
cores). Servo also makes further use of GPU-assisted acceleration, so
code running on a different processor. Would the GPU be needed if there
more core CPUs (real or multi-core) to parallelize the FP instructions?

I think GPU-assisted acceleration in web browsers started back in 2010,
but was just for web browsers. I remember some other apps used the GPU
for faster FP processing, but they seemed few and far between. More
video games are using AVX (AVX 2 more than AVX 512) since it is part of
the DirectX12 API. LOTS of users play video games on their home
computers, so AVX is really not that rare for use on low-end computing
platforms. AVX used to be shunned by game devs due to complexity in
coding.

Scalar, non-AVX :

void interpolate(vectorvectorint& mat)
{
for(int i=2; imat.size()-1; i=i+2)
for(int j=0; jmat[0].size(); j++)
{
mat[i][j] = mat[i-1][j] + 0.5f * (mat[i+1][j] - mat[i-1][j]);
}
}


Using AVX:

void interpolate_avx(vectorvectorint& mat)
{
for(int i=2; imat.size()-1; i=i+2)
for(int j=0; jmat[0].size(); j=j+8)
{
_mm256_storeu_si256((__m256i *)&mat[i][j],
_mm256_cvtps_epi32(_mm256_add_ps(_mm256_mul_ps(_mm 256_sub_ps(_mm256_cvtepi32_ps(_mm256_loadu_si256(( __m256i
*)&mat[i+1][j])), _mm256_cvtepi32_ps(_mm256_loadu_si256((__m256i
*)&mat[i-1][j]))), _mm256_set1_ps(0.5f)),
_mm256_cvtepi32_ps(_mm256_loadu_si256((__m256i *)&mat[i-1][j])))));
}
}

However, when mandated to programmers to code a game for maximum
performance, the AVX code runs 6.5 times faster! Simple coding with
slower performance, or complicated coding with faster performance. The
tradeoff is more cost in coding work, debugging, and optimizing hence
more time to achieve faster performance. Considering have video games
have upped the number of moving objects, physics modeling, and moving
texture change, some video games have insane requirements compared to
games dated over a decade ago.

Video games are real use of AVX. It's not just making benchmarks look
better. Guess Linus doesn't have bleeding edge hosts (in technology and
to his pocket) on which to run the most demanding video games. Is Linus
even a gamer? Oh wait, yeah, not that big a selection for Linux.

On 15/07/2020 19:42, VanguardLH wrote:
Is Linus even a gamer? Oh wait, yeah, not that big a selection for Linux.

He isn't anyway.

--
Brian Gregory (in England).

Yousuf Khan wrote:

Linus Torvalds' comments came from this article: https://is.gd/6zpZRL

How much did he make off of Linux?
(I will look, but seems like an amusing fact.)

On 7/15/2020 2:42 PM, VanguardLH wrote:
Not all CPUs are waiting to do something for end users. Some are
involved in highly complex computing, like animated computer graphics.
You think Zootopia was composed on a home computer? So, you think Intel
(or AMD) are going to tool up for a completely separate production line
for consumer vs high-graphics design platforms? There is an economy in
production by reusing existing manufacturing processes. Do consumer
platforms utilize AVX? Rarely. Why didn't Linus bitch when Intel added
Streaming SIMD Extensions (SSE)? How about all those non-gaming users
that don't care even about the old SSE extensions? Oh my God, the CPU
has something they don't need.

Well, no, the SSE extensions were a big improvement over the old
stack-based FPU model. Directly accessible FP registers rather than
pushing and popping indirectly off of a stack. Even AMD's 3DNow achieved
this, requiring even less changes to the hardware (it just fixed the
existing FPU stack model), although AMD did not yet have sufficient
marketshare to push it widely onto the market. I think the point Linus
is making is that AVX takes FPU's to a state that no one asked for. When
the first version of AVX came out, and no one used it, well okay just a
mistake, then the second version came out, hoping that it would correct
the deficiencies of the first one, still kind of understandable. When
even that one wasn't used, and now we're at like version 3 or 4, none of
which are being used, then that's obviously gone too far.

Yousuf Khan

On 2020-07-15 11:42, VanguardLH wrote:
Is Linus
even a gamer? Oh wait, yeah, not that big a selection for Linux.

Linux is not tied with Windows for gaming. Take
a gander at:

Fedora 31 | Features, Gaming, and New Daily Driver
https://www.youtube.com/watch?v=1P8oBlOTBho

On Wed, 15 Jul 2020 at 13:42:37, VanguardLH wrote:
Yousuf Khan wrote:

Linus Torvalds' comments came from this article: https://is.gd/6zpZRL

Full URL:
https://www.pcgamer.com/linux-founde...inventing-magi
c-instructions-and-start-fixing-real-problems/#referrer=https%3A%2F%2Fww
w.google.com&amp_tf=From%20%251%24s&ampshare=http s%3A%2F%2Fwww.pcgamer.c
om%2Flinux-founder-tells-intel-to-stop-inventing-magic-instructions-and-
start-fixing-real-problems%2F

I'm a little surprised at VLH for the above: surely it's rather _more_
than a Full URL: I think you could truncate it before the # sign. What
follows are "referrer" and "From", with another couple of URLs in there
(with the "://"s and subsequent "/"s turned into their hex equivalents).

Linus is known for publishing his tirades on Windows, and even on Linux
variants. He lambasts everyone.

Tweaking hardware to look good in benchmarks is news to you? Video chip
makers have been doing this forever, making their hardware or firmware
[]
Not exclusive to computing hardware of course! The last _big_ one I can
remember is Volkswagen getting _caught_ detecting when their engines
were undergoing the annual emission tests (as required by most
countries) and running accordingly, but I'm sure there are myriad
examples. (Note: not myriad _of_.)
--
J. P. Gilliver. UMRA: 1960/1985 MB++G()AL-IS-Ch++(p)Ar@T+H+Sh0!:`)DNAf

Everyone learns from science. It all depends how you use the knowledge. - "Gil
Grissom" (CSI).

Yousuf Khan wrote:

On 7/15/2020 2:42 PM, VanguardLH wrote:
Not all CPUs are waiting to do something for end users. Some are
involved in highly complex computing, like animated computer graphics.
You think Zootopia was composed on a home computer? So, you think Intel
(or AMD) are going to tool up for a completely separate production line
for consumer vs high-graphics design platforms? There is an economy in
production by reusing existing manufacturing processes. Do consumer
platforms utilize AVX? Rarely. Why didn't Linus bitch when Intel added
Streaming SIMD Extensions (SSE)? How about all those non-gaming users
that don't care even about the old SSE extensions? Oh my God, the CPU
has something they don't need.

Well, no, the SSE extensions were a big improvement over the old
stack-based FPU model. Directly accessible FP registers rather than
pushing and popping indirectly off of a stack. Even AMD's 3DNow achieved
this, requiring even less changes to the hardware (it just fixed the
existing FPU stack model), although AMD did not yet have sufficient
marketshare to push it widely onto the market. I think the point Linus
is making is that AVX takes FPU's to a state that no one asked for. When
the first version of AVX came out, and no one used it, well okay just a
mistake, then the second version came out, hoping that it would correct
the deficiencies of the first one, still kind of understandable. When
even that one wasn't used, and now we're at like version 3 or 4, none of
which are being used, then that's obviously gone too far.

Yousuf Khan

Already pointed out: your "none of which are being used" is wrong. It
is being used. Video games use it, and those are not rare on Windows
platforms. Any game using DirectX 12 are utilizing AVX2. Scientific,
statistical, financial, encryption, and other programs can use it. Any
program using .NET Framework can use AVX. The latest versions of
Prime95 are optimized to use AVX. While it is used to stress test, that
was not its original or current intent which was to discover prime
numbers. Is prime hunting something that home users do? Of course not,
but it illustrates AVX *is* used.

https://www.tomshardware.com/reviews...de,5461-2.html
"By default, Prime95 automatically selects the newest instruction set
extension, such as AVX, AVX2, or even AVX-512."

Your claim AVX is not used is false.

To test, go into the BIOS settings and change the AVX offset, and then
monitor the core frequencies, like with MSI's Afterburner. Surprise, a
lot of video games use AVX. You'll see the core frequencies go down
relative to the AVX offset when running an AVX-enabled program. I don't
play many new games (I still wish the Thief series keep evolving since
stealth is so poorly done in newer games), but have read SofTR,
Darksiders 3, Monster Hunter Word, AC: Odyssey, and Overwatch use AVX.
Overclockers trying to maintain the highest but stable clock rates whine
when core frequencies drop due to AVX, and have to change the AVX offset
to up the freqs.

https://www.google.com/search?q=over...20avx%20offset

If the games weren't using AVX, overclockers wouldn't be stymied over
the reduction in core freqs (and possible instability from vcore
reduction).

https://www.youtube.com/watch?v=BXBSVT9lpGw

Notice the AVX mode has higher frame rates. Also, it seems the right
side (for AVX) seems sharper overall. Timemark 2:23 starts the charts.
Later the author shows AVX doesn't improve performance in all games that
implement AVX. Sometimes AVX helps, sometimes not (but it's not worse).
I'm not into game programming, so I'll let someone else expert in that
note why AVX doesn't do better than SSE.

"J. P. Gilliver (John)" wrote:

On Wed, 15 Jul 2020 at 13:42:37, VanguardLH wrote:
Yousuf Khan wrote:

Linus Torvalds' comments came from this article: https://is.gd/6zpZRL

Full URL:
https://www.pcgamer.com/linux-founde...inventing-magi
c-instructions-and-start-fixing-real-problems/#referrer=https%3A%2F%2Fww
w.google.com&amp_tf=From%20%251%24s&ampshare=htt ps%3A%2F%2Fwww.pcgamer.c
om%2Flinux-founder-tells-intel-to-stop-inventing-magic-instructions-and-
start-fixing-real-problems%2F

I'm a little surprised at VLH for the above: surely it's rather _more_
than a Full URL: I think you could truncate it before the # sign. What
follows are "referrer" and "From", with another couple of URLs in there
(with the "://"s and subsequent "/"s turned into their hex equivalents).

is.gd, the URL shortening service that the OP used, does not provide a
preview mode. With TinyURL, you can add the "preview" hostname to see
where shortened URL points.

Well, is.gd does have a preview mode, but it's clumsy. You go to:

https://is.gd/previews.php

click on the "... see preview page ...", leave the web browser open, and
then click on the shortened URL the OP provided. Their page then shows
the full original URL and, yep, it has all that crap in it. Or you can
use on of the URL lengthener sites to reveal the original URL.

I gave the full URL that the *OP* provided with the shortened version.
Complain to the OP about not truncating URLs to their minimum. If he
had, he would not have needed the URL shortening service. The full URL:

https://www.pcgamer.com/linux-founde...real-problems/

is perhaps longer than the typical line length viewed in NNTP clients,
but slicing up URLs that are longer than the logical (viewed) line
length by injecting newlines (slicing URLs into multiple physical lines)
is a defect of the sender's client. Physical lines can be up to 998
characters long (that's the old-time recommendation). Maybe some NNTP
clients have problems when viewing physical line lengths longer than
their viewable line length making the URL not clickable, and why I've
seen some posters enclose the long URL within angle brackets, like
URL, as a workaround for deficient clients.

In any case, I showed the original (full) URL of what the OP used when
they generated the shortened version (well, a short redirection URL). I
showed the original URL. I didn't edit what the OP supplied.

Tweaking hardware to look good in benchmarks is news to you? Video
chip makers have been doing this forever, making their hardware or
firmware

Not exclusive to computing hardware of course! The last _big_ one I can
remember is Volkswagen getting _caught_ detecting when their engines
were undergoing the annual emission tests (as required by most
countries) and running accordingly, but I'm sure there are myriad
examples. (Note: not myriad _of_.)

https://www.bbc.com/news/business-34324772
Dated 10-Dec-2015

https://www.businessinsider.com.au/c...candal-2015-12
Dated 15-Dec-2015
Notice the chart showing the huge drop in sales.

I wonder how a car knows a gas sniffer is poking up its ahole. Oooh,
warm that up first before sticking it in. I suppose the car's computer
could notice the car wheels weren't rotating when the engine got revved
up and the steering wheel wasn't turning.

My state dropped emissions testing (for consumer vehicles which the
owner had to pay an $8 fee before they could get tabs) a long time ago.
I had a '92 bought used in '94 and kept for 24 years that never required
emission testing. I still have a '02 bought used in '04 that has never
required emissions testing. Emissions testing in my state ended back in
Nov 1999. Six other states don't have emissions testing, either. Our
requirement ceased after the levels of CO, ozone, and other pollutants
fell below the specs for the federal Clean Air Act; however, some states
are lobbying for stricter emissions control (exceeding EPA guidelines
and becoming more green-centric), so my state might go back to vehicle
testing despite our state has a green light. Must've been in sub-EPA or
more green-centric states where VW got busted for cheating.

On 2020-07-16 11:35, T wrote:
isÂ*not

"is now"

Stinking typos