You are here

SOS article: AMD vs Intel CPUs Tested

Page 1 of 1

SOS article: AMD vs Intel CPUs Tested

PostPosted: Tue Jun 30, 2020 1:02 pm
by Forum Admin
We're advance publishing this useful August issue Core Wars! SOS article:

https://www.soundonsound.com/sound-advice/core-wars-amd-intel-cpus-tested

It's unlocked and readable by anyone.

Re: SOS article: AMD vs Intel CPUs Tested

PostPosted: Wed Jul 08, 2020 1:16 am
by MaestroMikeT
Thanks! A great read!


Enviado do meu iPhone usando o Tapatalk

Re: SOS article: AMD vs Intel CPUs Tested

PostPosted: Sat Jul 11, 2020 3:06 am
by thevisi0nary
This is strange to me, how could it be that the 10900k performs worse than the 10900x?

Re: SOS article: AMD vs Intel CPUs Tested

PostPosted: Mon Jul 13, 2020 11:46 am
by Pete Kaine
It's a different micro-architecture and the X299 chipset is sitting somewhere between the server range than the consumer range, so whilst they are all closely related there are differences in the silicon. You've also got double the memory bandwidth to work with, one of the reasons we're going to look at RAM overclocking again and how spreading the load affects it.

Re: SOS article: AMD vs Intel CPUs Tested

PostPosted: Mon Jul 13, 2020 1:12 pm
by thevisi0nary
Pete Kaine wrote:It's a different micro-architecture and the X299 chipset is sitting somewhere between the server range than the consumer range, so whilst they are all closely related there are differences in the silicon. You've also got double the memory bandwidth to work with, one of the reasons we're going to look at RAM overclocking again and how spreading the load affects it.

Very interesting. One other question, would Ryzen still hold the lead like in these results when you have fx like Ozone on the masterbus, or recording live tracks? I only ask because of the difference in single thread performance.

Re: SOS article: AMD vs Intel CPUs Tested

PostPosted: Tue Jul 14, 2020 11:37 am
by Pete Kaine
With the masterbus effects chain, or indeed any channel effects chain I would be concerned for getting enough performance across all threads to support what needs to be done on each channel. It can be a problem when running low powered chips where threads might get spun down to the low 1.*GHz's and suddenly a heavy effect chain gets thrown at it for processing, but we're not working with any chip here that's going to be spinning anything down anywhere near that low.

The chips all generally have high headline single core turbo speeds, but those are not what we're looking for here. It's more about what the chip can do as a constant and all of them were holding their all core ratings which tend to sit between 4.2 and 4.7 depending on the chip in question.

The DSP test would be the closest we have here, but only when comparing the cores like for like. For instance the 10700k vs 3700X or 3800x has the 10700k coming off better at the 64 buffer (although, still neck and neck with surrounding models) than it tends to manage as the others scale up at higher buffer settings.

However, the single threaded performance question is dependent on the workload in question and how well the software is able to put all this into action. The general synthetic single thread benchmarks are fine in isolation, but they don't reflect how your software is going to be working with the chip in real-world terms. Running a test within a real-world application works better for our needs here, but there are still possible variables to be had between sequencers and even different plugins depending on how they've been optimized by the developer, so all we can do is a set a baseline and measure against it.

The big kicker for the single thread question is that none of these chip cores are running truly single thread only as they all feature SMT or HT so it muddies the water somewhat. You might see a clearer difference if we disabled those features, but then you'd also lose a chunk of performance overall.

Re: SOS article: AMD vs Intel CPUs Tested

PostPosted: Fri Jul 17, 2020 10:14 pm
by thevisi0nary
Pete Kaine wrote:With the masterbus effects chain, or indeed any channel effects chain I would be concerned for getting enough performance across all threads to support what needs to be done on each channel. It can be a problem when running low powered chips where threads might get spun down to the low 1.*GHz's and suddenly a heavy effect chain gets thrown at it for processing, but we're not working with any chip here that's going to be spinning anything down anywhere near that low.

The chips all generally have high headline single core turbo speeds, but those are not what we're looking for here. It's more about what the chip can do as a constant and all of them were holding their all core ratings which tend to sit between 4.2 and 4.7 depending on the chip in question.

The DSP test would be the closest we have here, but only when comparing the cores like for like. For instance the 10700k vs 3700X or 3800x has the 10700k coming off better at the 64 buffer (although, still neck and neck with surrounding models) than it tends to manage as the others scale up at higher buffer settings.

However, the single threaded performance question is dependent on the workload in question and how well the software is able to put all this into action. The general synthetic single thread benchmarks are fine in isolation, but they don't reflect how your software is going to be working with the chip in real-world terms. Running a test within a real-world application works better for our needs here, but there are still possible variables to be had between sequencers and even different plugins depending on how they've been optimized by the developer, so all we can do is a set a baseline and measure against it.

The big kicker for the single thread question is that none of these chip cores are running truly single thread only as they all feature SMT or HT so it muddies the water somewhat. You might see a clearer difference if we disabled those features, but then you'd also lose a chunk of performance overall.

Thank you that is very interesting indeed. Does this have anything to do with the core structure? I guess not because the 10980xe scales well and is monolithic like the 10900k which does not scale very well.

Another thing I've noticed is the difference in how gaming performance and low latency audio performance differ. Gaming seems to benefit from clocks, even with the lower ipc on intel, while audio performance does not seem to care about clock speed as much.

Re: SOS article: AMD vs Intel CPUs Tested

PostPosted: Fri Jul 17, 2020 11:02 pm
by Eddy Deegan
thevisi0nary wrote:Another thing I've noticed is the difference in how gaming performance and low latency audio performance differ. Gaming seems to benefit from clocks, even with the lower ipc on intel, while audio performance does not seem to care about clock speed as much.

If you consider the amount of data being shunted through the CPU by digital audio it's pretty low bandwidth compared to the vast amount of number crunching required for generating 3D scenes to render at decent framerates, even if the GPU is taking up much of the geometry calculations and rendering load.

In gaming a 3D scene has to be defined, a processing pipeline kept fed and framerate is paramount. Clock speed counts because the CPU is doing a lot of work: updating the scene every frame, keeping track of data structures representing graphical objects, handling gaming logic, I/O, networking, audio and so on, even though it's offloading much of the rendering to the GPU (which it has to, due to the phenomenal amount of matrix multiplications and other mathematics involved in rendering a scene).

Pure audio recording and playback is therefore lightweight and was even possible (albeit at lower sample rates and resolution) by the 32-bit 68030 running at 16Mhz on the Atari Falcon back in the day.

Of course modern resolutions and sample rates are higher but not so much that they bother any processor these days. Rather it's the overhead associated with plugins and the DAW itself that up the system requirements.

What audio performance does require more of is low latency, and this is more a function of the interface drivers and system design than it is of the CPU itself. After all, the CPU is effectively sitting there waiting for things to do, doing them, and pushing the result back out again. The surrounding systems responsible for receiving the incoming data stream from the AI driver, delivering it to the CPU in the shortest possible time and then pushing the results back out to the audio interface again have far more effect on the user experience than than CPU itself.

Thus for digital audio a slow CPU within a system that can shunt data to between the AI and the CPU quickly is preferable to a faster CPU in a system with less ingress/egress performance because in the latter case the CPU simply spends more time waiting for that data to arrive and the results take longer to reach the outside world again.

The CPU can only work with the data it is being fed and the means of delivery is outside of its control for the most part. With graphics, the bandwidth between the CPU and the GPU is high and it's local, thus the chip is working hard for much more of the time as opposed to waiting for stuff to come into it to process.

Pete can certainly talk more in detail about this than I can but that's the gist.

Re: SOS article: AMD vs Intel CPUs Tested

PostPosted: Sat Jul 18, 2020 10:09 am
by Matt Houghton
I use a lot of Acustica Audio Acqua plug-ins, and they certainly benefit from a higher CPU clock speed...

Re: SOS article: AMD vs Intel CPUs Tested

PostPosted: Sun Jul 19, 2020 6:15 pm
by thevisi0nary
Eddy Deegan wrote:
thevisi0nary wrote:Another thing I've noticed is the difference in how gaming performance and low latency audio performance differ. Gaming seems to benefit from clocks, even with the lower ipc on intel, while audio performance does not seem to care about clock speed as much.

If you consider the amount of data being shunted through the CPU by digital audio it's pretty low bandwidth compared to the vast amount of number crunching required for generating 3D scenes to render at decent framerates, even if the GPU is taking up much of the geometry calculations and rendering load.

In gaming a 3D scene has to be defined, a processing pipeline kept fed and framerate is paramount. Clock speed counts because the CPU is doing a lot of work: updating the scene every frame, keeping track of data structures representing graphical objects, handling gaming logic, I/O, networking, audio and so on, even though it's offloading much of the rendering to the GPU (which it has to, due to the phenomenal amount of matrix multiplications and other mathematics involved in rendering a scene).

Pure audio recording and playback is therefore lightweight and was even possible (albeit at lower sample rates and resolution) by the 32-bit 68030 running at 16Mhz on the Atari Falcon back in the day.

Of course modern resolutions and sample rates are higher but not so much that they bother any processor these days. Rather it's the overhead associated with plugins and the DAW itself that up the system requirements.

What audio performance does require more of is low latency, and this is more a function of the interface drivers and system design than it is of the CPU itself. After all, the CPU is effectively sitting there waiting for things to do, doing them, and pushing the result back out again. The surrounding systems responsible for receiving the incoming data stream from the AI driver, delivering it to the CPU in the shortest possible time and then pushing the results back out to the audio interface again have far more effect on the user experience than than CPU itself.

Thus for digital audio a slow CPU within a system that can shunt data to between the AI and the CPU quickly is preferable to a faster CPU in a system with less ingress/egress performance because in the latter case the CPU simply spends more time waiting for that data to arrive and the results take longer to reach the outside world again.

The CPU can only work with the data it is being fed and the means of delivery is outside of its control for the most part. With graphics, the bandwidth between the CPU and the GPU is high and it's local, thus the chip is working hard for much more of the time as opposed to waiting for stuff to come into it to process.

Pete can certainly talk more in detail about this than I can but that's the gist.

Thanks a lot. What I am mainly referring to is how despite having lower ipc, Intel is able to take the lead in gaming situations over AMD because of how fast they can clock on a single thread (per core). I was just curious as to how this advantage does now seem to apply to audio and cpu load at low latencies, as a higher raw throughput on each core would theoretically allow a larger load.

VERY Theoretical example - Suppose the 3900x with 12 cores could run 100 instances of "plugin x" per core at a 256 buffer, so 1200 instances of plugin x total. The 10900k on average has a 5%-10% single core advantage (let's say a conservative 5% for now). With two less cores, the 10900k would get around 1050 instances of plugin x. Thats 13% more instances on the 3900x, but the difference in the dawbench is far larger than that at about 25%.

Another example is how the 10900k beats the 10900x for gaming because it clocks faster, but in the dawbench the opposite is true for audio despite the 10900x having a slower single thread speed and despite having the same number of cores.

Re: SOS article: AMD vs Intel CPUs Tested

PostPosted: Mon Jul 20, 2020 7:14 pm
by Pete Kaine
In testing, the 10900K was holding it's all core average around 4.68, the 10900X was holding about 4.5, so your talking an extra couple of 100MHz per core. Is that enough to allow an extra plugin to be inserted on each channel?

We've got 10 cores running dual threaded and getting a 100 plugins (looking at the vi test it's around that), so if we averaged it out you could say that's load spreading around 10 plugins per core, or 5 plugins per thread. If each thread is 4500MHz and doing X amount, then looking at how the load is working already, I wouldn't expect another 200MHz to really give us anything above and beyond what we're what we're already seeing so in that regards it should be close.

I think the extra few percent over that to the X model is coming from the extra memory bandwidth on the quad channel arrangement. Its something I want to check further once I can get a chance to sit and go over them again. I recall running an earlier X299 in dual channel mode and seeing some drop in the performance previously, so if it did so here again even by 5% it would answer the question.

Re: SOS article: AMD vs Intel CPUs Tested

PostPosted: Tue Jul 21, 2020 7:36 pm
by thevisi0nary
Pete Kaine wrote:In testing, the 10900K was holding it's all core average around 4.68, the 10900X was holding about 4.5, so your talking an extra couple of 100MHz per core. Is that enough to allow an extra plugin to be inserted on each channel?

We've got 10 cores running dual threaded and getting a 100 plugins (looking at the vi test it's around that), so if we averaged it out you could say that's load spreading around 10 plugins per core, or 5 plugins per thread. If each thread is 4500MHz and doing X amount, then looking at how the load is working already, I wouldn't expect another 200MHz to really give us anything above and beyond what we're what we're already seeing so in that regards it should be close.

I think the extra few percent over that to the X model is coming from the extra memory bandwidth on the quad channel arrangement. Its something I want to check further once I can get a chance to sit and go over them again. I recall running an earlier X299 in dual channel mode and seeing some drop in the performance previously, so if it did so here again even by 5% it would answer the question.

It makes sense to me! The 4.6 all core on the 10900k is definitely lower than the advertised all core boost of 4.9, and id argue the main incentive of getting the K chips is the overclocking ability. Getting the 10 cores to say 5.1 would probably yield a marked improvement, and the thermals are much more manageable than on the 9900k or ks. People have been getting the 10700k to 5ghz on air cooling.

Then however, given how close 3900x is core for core, the 10900k with two less cores doesn't make much sense if there aren't any situations where it clearly comes out on top. That's why I was inquiring about things like master bus processing or tracking with FX at low latencies.

I'm super interested in the upcoming Zen 3 equivalent of the 3900x or 3950x. It's just that I've had a 4790k that has been incredibly solid for the last 6 years, and I guess I am trying to scrutinize the differences between these chips before I jump ship to a new platform I've had no experience with. Your dawbench tests have certainly been very helpful with that.

On an mostly unrelated note, something I have been speculating about is the rumor that the infinity fabric might be increased to 2000mhz on Zen 3 chips. It doesn't sound like an easy task getting 64gb of ram to 4000mhz to be at 1:1 with the F clock. :?