AMD Ryzen 3rd gen. ZEN 2 processors for audio PC

[tooneba]

Another downside is that (almost all of) the new X570 motherboards will have one of those annoying small chipset fans. They tend to be more unreliable than their larger counterparts. What if they break? And how loud will they be? Remains to be seen.

Edit: according to this test, the fan is annoying...

https://www.tomshardware.com/reviews/as ... 217-3.html

You don't need X570.

[Scotty]

When do you think we're going to see scanproaudio benchmarks?

Yes, the audio world awaits ScanAudio with baited breath and appreciation!

[poshook]

Please, guys, make sure you do not use huge sampled libraries and super low latencies as here new multiple dies Ryzens are far behind the Intel's single die CPUs.

It's worth emphasizing that only the 3900X has two compute dies, the 3700X and 3800X (and lower) are all single die processors, just like intel...

You are right, my fault. With Ryzen 3000 they changed the number of cores per die from 4 to 8 but still uses 4-core CCX within one CCD. In terms of inter CCX communication there is still one infinity fabric for both CCX inside the CCD. So it is not clean 8-Core die. We will see if 3990x and higher change the bad reputation of previous Ryzen.

Cores always need access to other cores and cache wherever they are architecturally located, and there's no one perfect solution.

AMD has infinity fabric, Intel used to have a ringbus system and now has a mesh. The systems are actually not very easily comparable because they work very differently. If you want a gross simplification, ryzen is fast inside one CCX, but becomes slow once you need the 5th core on the other CCX -- intel's mesh is between those two examples in speed. What I mean by speed here is less latency.

What makes the whole question of "which is better" complicated, is that everything is also very dependent on what kind of workload you are feeding the processor. There simply is no universal "VST" performance. Effects and instruments can be programmed in myriad ways.

Sample based VST instruments work the same. A midi event plays respective sample located in RAM and SSD. CPU is waiting on event then search for right sample and play it. When you use multigigabytes libraries, initial part of all samples are preloaded into memory and the tail is played directly from disc.

[Scotty]

It is all speculation until we get some specific audio benchmarks. AMD has made significant changes to their CPU designs and we can't just draw a line back to where we were with first gen Ryzens. We need some mixed usage scenarios as everyone's workflow could be different. Ultra low latency ASIO settings, high ASIO buffers, single threaded VSTs and deep mixes with loads of audio and instrument tracks will likely tax the CPU differently and we have to assess the processor from different vantages points in comparison to Intel counterpoints AND price matters for some of us. Even mixed usage like video editing may factor into this decision for some of us.

[pljones]

And if memory access is still an issue, my 70 odd tracks each with several gigabytes of samples I want to trigger in real time will be another rather different workload. My Intel CPU (Intel Core i5 8600... maybe I should have gone i7) struggles. Anyway, I have to wait a few years now before I can get away with a new PC.

[Passing Bye]

When do you think we're going to see scanproaudio benchmarks?

From GS: https://www.gearslutz.com/board/music-c ... ies-3.html

It's coming, delayed through and probably towards the end of the week.

Our 3600X sample cropped up with a X570 board last week, but nothing else has landed yet. No samples or stock on the 3700X and 3900X and no other boards yet for stock.

Physical stock was due with us last Friday week and I will pull and test it before the first systems get built. The 3600X I started on Friday and ran into the machine check BSOD's that a few other reviewers noted, so I'll say don't try migrating your old drive on this one (just yet at least), rather do a clean install and it seems to behave itself... obviously, I had to build a new test bench in the meantime!

Also to note, I spent last week reworking both tests in order to get something that wouldn't max out by 128 buffer on each of the new chips. It might be why Tech Hardware dropped it I guess (and nothing but a guess at this point), if the reviewer didn't feel like reinventing the wheel. The impression I've been given is that review chips went out late as it is, so chances are the deadline might have put a halt to it if any complications came up.

When I publish, I'll list the testing changes I enacted, should anyone feel like playing along at home

[poshook]

It is all speculation until we get some specific audio benchmarks. AMD has made significant changes to their CPU designs and we can't just draw a line back to where we were with first gen Ryzens. We need some mixed usage scenarios as everyone's workflow could be different. Ultra low latency ASIO settings, high ASIO buffers, single threaded VSTs and deep mixes with loads of audio and instrument tracks will likely tax the CPU differently and we have to assess the processor from different vantages points in comparison to Intel counterpoints AND price matters for some of us. Even mixed usage like video editing may factor into this decision for some of us.

Scotty, we all discuss about the new Zen 2 architecture. However that is not that different from the previous generations. First deep testing uncover similar issues like huge memory access delays and relation between RAM speed and CPU performance. I will be very happy to see great results in DAW bench as my i7-7820x TPD is way too high and I would gladly change the platform even the AMD prices are a lot steeper compared to those announced some months ago (12-Core and 16-Core variants).

[Liero]

Please, guys, make sure you do not use huge sampled libraries and super low latencies as here new multiple dies Ryzens are far behind the Intel's single die CPUs.

It's worth emphasizing that only the 3900X has two compute dies, the 3700X and 3800X (and lower) are all single die processors, just like intel...

You are right, my fault. With Ryzen 3000 they changed the number of cores per die from 4 to 8 but still uses 4-core CCX within one CCD. In terms of inter CCX communication there is still one infinity fabric for both CCX inside the CCD. So it is not clean 8-Core die. We will see if 3990x and higher change the bad reputation of previous Ryzen.

Cores always need access to other cores and cache wherever they are architecturally located, and there's no one perfect solution.

AMD has infinity fabric, Intel used to have a ringbus system and now has a mesh. The systems are actually not very easily comparable because they work very differently. If you want a gross simplification, ryzen is fast inside one CCX, but becomes slow once you need the 5th core on the other CCX -- intel's mesh is between those two examples in speed. What I mean by speed here is less latency.

What makes the whole question of "which is better" complicated, is that everything is also very dependent on what kind of workload you are feeding the processor. There simply is no universal "VST" performance. Effects and instruments can be programmed in myriad ways.

Sample based VST instruments work the same. A midi event plays respective sample located in RAM and SSD. CPU is waiting on event then search for right sample and play it. When you use multigigabytes libraries, initial part of all samples are preloaded into memory and the tail is played directly from disc.

You're right, except that "directly from disc" in reality means "load into memory, but flush it immediately after it is used" (to conserve memory space).

If you use a lot of disc streaming, or DFD as Native Intruments calls it, memory frequency, memory latency and storage device seek time are things you should value -- which is noteworthy because most other plugins, effect or instrument, don't need those at all. For example a really CPU-intensive oscillator or delay effect don't really benefit from those.

[poshook]

Please, guys, make sure you do not use huge sampled libraries and super low latencies as here new multiple dies Ryzens are far behind the Intel's single die CPUs.

It's worth emphasizing that only the 3900X has two compute dies, the 3700X and 3800X (and lower) are all single die processors, just like intel...

You are right, my fault. With Ryzen 3000 they changed the number of cores per die from 4 to 8 but still uses 4-core CCX within one CCD. In terms of inter CCX communication there is still one infinity fabric for both CCX inside the CCD. So it is not clean 8-Core die. We will see if 3990x and higher change the bad reputation of previous Ryzen.

Cores always need access to other cores and cache wherever they are architecturally located, and there's no one perfect solution.

AMD has infinity fabric, Intel used to have a ringbus system and now has a mesh. The systems are actually not very easily comparable because they work very differently. If you want a gross simplification, ryzen is fast inside one CCX, but becomes slow once you need the 5th core on the other CCX -- intel's mesh is between those two examples in speed. What I mean by speed here is less latency.

What makes the whole question of "which is better" complicated, is that everything is also very dependent on what kind of workload you are feeding the processor. There simply is no universal "VST" performance. Effects and instruments can be programmed in myriad ways.

Sample based VST instruments work the same. A midi event plays respective sample located in RAM and SSD. CPU is waiting on event then search for right sample and play it. When you use multigigabytes libraries, initial part of all samples are preloaded into memory and the tail is played directly from disc.

You're right, except that "directly from disc" in reality means "load into memory, but flush it immediately after it is used" (to conserve memory space).

If you use a lot of disc streaming, or DFD as Native Intruments calls it, memory frequency, memory latency and storage device seek time are things you should value -- which is noteworthy because most other plugins, effect or instrument, don't need those at all. For example a really CPU-intensive oscillator or delay effect don't really benefit from those.

You are completely right. This is the reason why DAW Bench uses two methods to measure CPU performance. First is DSP (math), where Ryzens are rather fine at reasonable latencies and VSTI Polyphony - tight memory-allocation process at low buffer settings, where Ryzens are rather weak.

[Scotty]

You are completely right. This is the reason why DAW Bench uses two methods to measure CPU performance. First is DSP (math), where Ryzens are rather fine at reasonable latencies and VSTI Polyphony - tight memory-allocation process at low buffer settings, where Ryzens are rather weak.

What does "rather weak" mean in the context of a given workflow? If you need to track a live band at 64 samples using 32 tracks max (each with a basic channel strip plugin) using a good audio card such as an RME... can a ZEN 2 handle it? That is enough low latency performance for me. It doesn't have to be better than Intel it just has to be good enough.

From there I build up my mix and can increase the buffer settings as I add layering. If I have to leverage zero latency monitoring to add a vocal dub later I can do it. Ryzen only has to be "good enough" at low latency to get my bed tracks recorded. From there multi threading for layering and having the CPU to add the high end processing at higher latencies is my typical process. One size doesn't fit all. If the new Ryzen's can do that, then give me the multicore/multithread overhead that I need to add expensive (in terms of CPU) processing then this is a good solution for me. On the other hand if it craps out tracking 32 channels live at 64 sample buffers then I really have no choice but to remain with Intel whilst sucking up the additional cost and accepting that on a per dollar basis I MAY get lower multicore performance.

I think we need to get away from the binary - "winner or loser" approach to this and assess it in terms of use-case scenarios. I am not sure if you have an argument with that or not and apologies if I missed the main thrust of your post. I do appreciate the comments in this thread. - Scotty

[Pictus]

Scotty, we all discuss about the new Zen 2 architecture. However that is not that different from the previous generations. First deep testing uncover similar issues like huge memory access delays and relation between RAM speed and CPU performance.(...)

The architecture got improvements/rebalance, at the same time it has slower
"memory something" it got better inter-core data latency, bigger cache, etc...
As already mentioned, is hard to tell how it behaves for audio workloads
by looking into other benchmarks.

[Scoox]

Scan Pro Audio, release the gold!!

[Obsolete236871]

Speaking of audio interfaces, drivers and latency - does anyone know of a good overview of the realtime audio performance (eg ASIO) of various audio interfaces?

I know that RME is often recommended, but I would like to know how the more affordable products perform (especially USB interfaces, eg the current interfaces of Focusrite, Presonus, Native Instruments or M-Audio to name a few common brands).

[Passing Bye]

Speaking of audio interfaces, drivers and latency - does anyone know of a good overview of the realtime audio performance (eg ASIO) of various audio interfaces?

I know that RME is often recommended, but I would like to know how the more affordable products perform (especially USB interfaces, eg the current interfaces of Focusrite, Presonus, Native Instruments or M-Audio to name a few common brands).

https://www.gearslutz.com/board/music-c ... -base.html

[pljones]

Are all those on the chart still currently made by their manufacturers..?