Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

crazy dave

macrumors 65816
Sep 9, 2010
1,454
1,230
Intel could be interesting for the upcoming MAC Pro where heat and power efficiency is not an issue, and you don’t have to deal with software compatibility issues. It just works.

For laptops and tablets, ARM is better in all benchmarks but you got to do deal with software headaches.

There are all sorts of rumors that Apple will release one last Intel Mac Pro. However, even in a desktop and workstation heat and power efficiency of individual cores still matter as they dictate how many cores and what frequency they can be run at. You'll notice that for both Intel and AMD the base clocks are far, far lower the higher the core count. The M1 won't need to do this because each core is only using about 5-6W. This also makes it easier to calculate what the MT scores of a Max Duo (20-thread) will be and the ST benchmarks will be identical to what we have now. Basically double everything and shave a touch off for interconnect between the dies. First order approximation of course, but this would make for a very powerful desktop. In most applications that will be faster than a 24-thread ADL CPU and quieter to boot. ;) For a workstation, 40-thread class chip, were it to be based on M1 which is unclear at this point as there are competing rumors, again pretty easy to estimate its performance. Because each core draws so little heat - basically quadruple the benchmark scores and shave a bit off. While granted efficiency is even more important in a mobile or physically thermally constrained environment ... put enough cores into an x86 chip and you are still going to be thermally constrained. ;)
 

jeanlain

macrumors 68020
Mar 14, 2009
2,463
958
In contrast the M1 is ~5W/10W for core/package in ST SPEC give or take (a touch higher in SPEC POV ray as that's a heavy test). I think @leman took 71/5 ~= 14, though it should be 55/5 ~= 11 as that's the core to core. Still 11x ... however, that's not what Linux reports for the same test and what Windows is reporting here seems bonkers even for Intel core. So ... I'm not sure.
And the M1 max is 23% slower than de 12900k in that test.
I can't find data for power consumption running povray on linux. Isn't povray notable for high power consumption, like prime95?
 

crazy dave

macrumors 65816
Sep 9, 2010
1,454
1,230
And the M1 max is 23% slower than de 12900k in that test.
I can't find data for power consumption running povray on linux. Isn't povray notable for high power consumption, like prime95?
It's in the quote I put in italics - ST 25-30W in Linux which is in line with the rest of Spec tests - POV Ray (CPU ray tracing program) is on the higher end of power usage for all CPUs including the M1 but nothing crazy (as you noted for Alder Lake they said the average was 25W in SPEC tests). Looking at the rest of the chart, every subsequent P-core added seems to add 20-30W even in Windows, so that "70W" of package power for the single core must include a lot of overhead - they put the overhead at 20% but I'm guessing that it must be actually half. Again in the quote I shared, they admitted they couldn't get the per core measurements on Windows working correctly, only package power. They didn't quote package power for Linux though so I can't compare.
 

crazy dave

macrumors 65816
Sep 9, 2010
1,454
1,230
Right. That third line is suspicious.

Yeah it's all down to the compiler - they compiled with clang-llvm for the M1 Max and ICC for the Intel processor which according to Intel gives a 48% boost over standard LLVM in Spec Int 2017.


That works out pretty perfectly as Apple's M1 Max was measured by Anandtech to be 37% faster than the 11980HK when the compiler was the same between both. I'm not sure how @Andropov in his talkedabout post measured the difference in performance the graph but assuming his 8% faster for 11980HK is right that 45% swing almost exactly in line with the boost Intel claims from using ICC.

Basically Intel is giving themselves a free ~45% boost by changing the compiler. As noted here:


Intel has an illustrious history of doing this.
 

cmaier

Suspended
Jul 25, 2007
25,405
33,474
California
Yeah it's all down to the compiler - they compiled with clang-llvm for the M1 Max and ICC for the Intel processor which according to Intel gives a 48% boost over standard LLVM in Spec Int 2017.


That works out pretty perfectly as Apple's M1 Max was measured by Anandtech to be 37% faster than the 11980HK when the compiler was the same between both. I'm not sure how @Andropov in his talkedabout post measured the difference in performance the graph but assuming his 8% faster for 11980HK is right that 45% swing almost exactly in line with the boost Intel claims from using ICC.

Basically Intel is giving themselves a free ~45% boost by changing the compiler. As noted here:


Intel has an illustrious history of doing this.
Exactly. They do this all the time. Not just cherry-picking tests, but actually comparing different things.
 
  • Like
Reactions: psychicist

leman

macrumors Core
Original poster
Oct 14, 2008
19,522
19,679
Yeah it's all down to the compiler - they compiled with clang-llvm for the M1 Max and ICC for the Intel processor which according to Intel gives a 48% boost over standard LLVM in Spec Int 2017.

That kind of improvements just because of a different compiler can mean one of three things:

1) the previous compiler sucks at optimization
2) the new compiler performs aggressive autovectorization
3) the new compiler performs aggressive benchmark-specific optimizations

Since we can exclude 1) - clang is a mature compiler that produces code of excellent quality, it’s probably the combination of two and three. So again, benchmark shenanigans.
 

Stratus Fear

macrumors 6502a
Jan 21, 2008
696
433
Atlanta, GA
That kind of improvements just because of a different compiler can mean one of three things:

1) the previous compiler sucks at optimization
2) the new compiler performs aggressive autovectorization
3) the new compiler performs aggressive benchmark-specific optimizations

Since we can exclude 1) - clang is a mature compiler that produces code of excellent quality, it’s probably the combination of two and three. So again, benchmark shenanigans.
Intel’s compiler and benchmark shenanigans go back over a decade; It’s definitely not a new thing
 
  • Like
Reactions: psychicist

Xiao_Xi

macrumors 68000
Oct 27, 2021
1,628
1,101
Since we can exclude 1) - clang is a mature compiler that produces code of excellent quality, it’s probably the combination of two and three. So again, benchmark shenanigans.
Do those compiler optimizations improve daily use programs?
 

leman

macrumors Core
Original poster
Oct 14, 2008
19,522
19,679
Do those compiler optimizations improve daily use programs?

That’s exactly the question. Is this a performance increase that’s specific to this one benchmark or does it apply to other workloads too? If it’s achieved via autovectorization does it also improve the performance on AMD CPUs or is it Intel only (and if the latter, why)? What do we know about these compilers?
 
  • Like
Reactions: JMacHack

Xiao_Xi

macrumors 68000
Oct 27, 2021
1,628
1,101
Is this a performance increase that’s specific to this one benchmark or does it apply to other workloads too?
Intel claims that:
  • Deep learning framework performance is accelerated up to 10 times over earlier versions with the latest Intel® Optimization for TensorFlow and Intel® Optimization for PyTorch.
  • New Intel® Extension for Scikit-learn speeds up machine learning algorithms more than 100 times on Intel CPUs over the stock open source version.
An Intel dev and a sklearn dev wrote this article: https://medium.com/intel-analytics-software/accelerate-your-scikit-learn-applications-a06cacf44912

Are those improvements due to the Intel compiler?
 

Taz Mangus

macrumors 604
Mar 10, 2011
7,815
3,504
Can’t wait to read all the reviews about the laptops using the 12th generation Intel processor, noting how quiet and cool they run. Plus the ability to run the most demanding workloads on battery alone. That is the holy grail for laptops. Can Intel rise to the challenge or fizzle out with a lot of hot air claims.

Intel has an opportunity, will they squander it? I am all for competition to drive better tech for the consumer. But that requires a company such Intel to bring something to the table that will cause other companies to improve. Unfortunately, or fortunately, Apple seems to be competing with themselves.
 
Last edited:

ingambe

macrumors 6502
Mar 22, 2020
320
355
Intel’s compiler and benchmark shenanigans go back over a decade; It’s definitely not a new thing
Intel has spent a lot of effort and money to build a powerful compiler and libraries for scientific computing. Why wouldn't they use it?

That's one thing I blame AMD for, they produce awesome hardware, but the software doesn't follow.
NVIDIA had the Cuda library for a decade. Thanks to their effort, all the deep learning community is using their hardware. Only recently AMD has started to catch up.
And it's the same with CPU. Intel MKL is very good and can provide a tremendous performance improvement, betting some AMD CPUs that, on paper, are faster.
 

Stratus Fear

macrumors 6502a
Jan 21, 2008
696
433
Atlanta, GA
Intel has spent a lot of effort and money to build a powerful compiler and libraries for scientific computing. Why wouldn't they use it?

That's one thing I blame AMD for, they produce awesome hardware, but the software doesn't follow.
NVIDIA had the Cuda library for a decade. Thanks to their effort, all the deep learning community is using their hardware. Only recently AMD has started to catch up.
And it's the same with CPU. Intel MKL is very good and can provide a tremendous performance improvement, betting some AMD CPUs that, on paper, are faster.
Did you even read the link I posted?
 

crazy dave

macrumors 65816
Sep 9, 2010
1,454
1,230
Do those compiler optimizations improve daily use programs?
Yes and no. ICC is rarely used outside of HPC. What Intel ICC is doing is basically auto applying compiler optimizations that would typically have to be done by hand with pragmas or through rearranging/rewriting code. The very smart compiler engineers at Intel have figured out how to do this better and more automatically than others (though sometimes people complain about ICC breaking things and even Intel states that some optimization aren't upstreamed to standard LLVM because they are "experimental"). But were these optimizations applied to the code itself or to the compilers, other compilers would likewise show similar improvements. Not every code gets improved in this way - even in HPC, some people see a benefit to adopting Intel libraries and ICC (on Intel chips), others not so much. Sometimes even regression, or as I mentioned earlier, problems. YMMV. The larger issue is that it doesn't showcase a true difference in the hardware - a great but old examples is in the link @Stratus Fear provided where in some egregious cases a few years the Intel ICC was deliberately not compiling faster code paths on AMD processors despite the fact that the AMD hardware was capable of running the faster code path. Thus making benchmarks compiled with ICC look like they were running faster on Intel than AMD. Given the stink I think, I hope they don't still this (also they want buy in for oneAPI, so I'm guessing they don't). But it's just an example of why closed source compilers aren't used by 3rd party benchmarks and you want to vary as little about the compilation process as possible between CPUs.

Intel has spent a lot of effort and money to build a powerful compiler and libraries for scientific computing. Why wouldn't they use it?

That's one thing I blame AMD for, they produce awesome hardware, but the software doesn't follow.
NVIDIA had the Cuda library for a decade. Thanks to their effort, all the deep learning community is using their hardware. Only recently AMD has started to catch up.
And it's the same with CPU. Intel MKL is very good and can provide a tremendous performance improvement, betting some AMD CPUs that, on paper, are faster.

The issue is *how* they use it. For instance, in the link in the post you quoted and from one of my earlier posts Intel ICC had a habit of 1) not compiling the faster code path for AMD processors even though they could've and 2) not disclosing that in addition to compiling benchmarks with ICC, they also switched out standard libraries with Intel accelerated ones. Now I don't know if they still do all of these things. But the net result was that their first party benchmarks would show these incredibly juicy scores relative to their competition but which wouldn't be reflected by anyone else's benchmarks or in the real world for actual applications. And this gets to a larger point:

Yes in HPC these libraries and compilers are used and can be useful (though GCC still predominates as far as I can tell). So if this were marketing aimed at getting HPC users to switch not just to Intel but the full Intel development ecosystem, then you could make an argument for this (well maybe not deliberately not optimizing code paths, but using Intel toolchains on Intel processors compared to everyone else with standard toolchains). However, this is marketing aimed at laptops for the general public and that makes this extra disingenuous beyond what I mentioned above because most standard software is not compiled with Intel-specific libraries and compilers. So people buying laptops even for productivity aren't going to see these kinds of relative speeds reflected in their software and workloads.
 

leman

macrumors Core
Original poster
Oct 14, 2008
19,522
19,679
@crazy dave offers an excellent reply to this above but here are my brief comments.

Intel has spent a lot of effort and money to build a powerful compiler and libraries for scientific computing. Why wouldn't they use it?

Because it’s highly misleading. I mean, I would be probably fine with it if they talked about a HPC CPU, but they are presenting general purpose laptop computing. Using a “magical” compiler with highly tweaked options and custom allocators/whatnot does not reflect the performance of the CPU - it reflects the performance of the compiler and the runtime. If I switch out the allocator and get a 20% improvement my CPU didn’t magically become faster. So why would I claim that it became fasstet?

Besides, normal users do not run software built with icpx.


NVIDIA had the Cuda library for a decade. Thanks to their effort, all the deep learning community is using their hardware. Only recently AMD has started to catch up.

You are thankful for Nvidia actively sabotaging and manipulating the market to make users dependent on their proprietary hardware? That’s some strong Stockholm syndrome going on here.


And it's the same with CPU. Intel MKL is very good and can provide a tremendous performance improvement, betting some AMD CPUs that, on paper, are faster.

Intel MKL is good, sure. But why does it run like crap on AMD CPUs that support all the same instructions? How is that user friendly to ship a high performance library that will deliberately sabotage performance on certain platforms? No matter how you spin it it’s a dick move.
 

EntropyQ3

macrumors 6502a
Mar 20, 2009
718
824
Intel’s compiler and benchmark shenanigans go back over a decade; It’s definitely not a new thing
Three decades at least.
I was around at the formation of SPEC, and unfortunately it never really succeeded in its stated goal of being an unbiased source of performance information, commercial interests crept in early.
However, I think the suite was hugely successful in putting the spotlight on the importance of compilers! Development and quality obviously, but also cheating, benchmark specific optimisations, and of course the allowed futzing around with settings for each subtest, et cetera - we've seen it all.
Intel also funds/have funded benchmark producers whose products miraculously favours Intel, and is nevertheless widely used for ... whatever reason.

Money talks.
 

JMacHack

Suspended
Mar 16, 2017
1,965
2,424
You are thankful for Nvidia actively sabotaging and manipulating the market to make users dependent on their proprietary hardware? That’s some strong Stockholm syndrome going on here.
Some people just can’t get off Jensens wang.

Examples of this?
Refusal to open source their Linux drivers, forcing game devs to use unoptimized code for AMD gpus to get partnered, (by not culling out of frame polys specifically on AMD hardware), gimping performance on previous gen hardware by removing optimizations that already existed, forcing PhysX to use the cpu on AMD cards, and then pushing partnered devs to use it more. GPP (which thankfully died), Gsync (which required an NVidia chip of course), as opposed to freesync which didn’t.

Off the top of my head.
 
  • Like
Reactions: Stratus Fear
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.