Jumping from 7700k to 3950x; Anyone want me to benchmark anything?

[cassembler]

Jumped ship from an i7-7700k to a 3950x. (Idea is primarily to increase processing threads.)

I've run a few real-life, practical before/after benchmarks, results below.

TL;DR - it appears the 3950x effectively offers about 200% more multi-track processing performance over the 7700k (i.e. it's equivalent to three 7700k processors), even though the CPU benchmarks at roughly a 300% increase (equivalent to four 7700k processors).

(EDIT: Most (all) of my work is VSTi; I don't have a setup to test multi-channel tracking.)

[cassembler]

Results! (test notes and other details at bottom)

CPU-Z Benchmarks:
3950 is about 5% better single core, about 300% better multi-core. Pretty much aligns with the specs.

Lowest Latency, Single Guitar Tracking
Q: How much processing can you have on a single channel for tracking guitar at very low latency?

7700k: Guitar Rig 5 or Ampli Tube 4 with built-in FX @2.4ms; Mercuriall with FX at 3.9ms. Anything else causes artifacts, necessitating increase buffer size.
3950x: First effort was AmpliTube 4 @ 2.4 ms, and added Taupe (Acustica) for grins. Nope, not quite, even with the ZL version. So then I added AmpliTube 4, FabFilter Pro-R, Waves H Delay, Waves H-Reverb (long), Guitar Rig 5, Mercuriall Spark, Waves GTR, and SoftTube's Tape. All on same channel, all at 2.4 ms latency (each way). After that, started breaking apart.

Conclusion: The 3950x appears to significantly increase real-time processing for single-channel tracking; the only way I could break it was with something Acustica, but interestingly anything Acustica seems to break it at this latency.

Beefiest Single Channel
Q: How do these two processors handle an epic single channel?
- Note: Buffer size set to max (8192 samples, ~186 ms each way) from here out

7700k: Massive X + 3x Gold2 + 2x Taupe + 2x Azure2
3950x: Massive X + 4x Gold2 + 3x Taupe + 3x Azure2

(observation: this is where I start seeing the '4 thread channel cap' behavior)

Conclusion: The 3950x squeezes out one more instance of each Acustica plugin, call it 30% more total. Not sure what to think here - I suppose the limit is that a single channel can only use four threads... Thus, I would have expected about a 20% increase (5% better single-core performance). I'll offer up 10% may well be margin of error here, as my efforts are technically unscientific.

Number of High-End Tracks
Q: How many tracks can I run with a high-end FX chain?
(Massive X > Purple2P1 > Waves F6-RTA > Fab Filter Pro-R > TDR Kotelnikov > Taupe)

7700k: 10 Tracks
3950x: 31 Tracks

Conclusion: The 3950 bakes up about 210% more here. I thought this should be a bit higher - we're roughly expecting 300% increase in multi-core performance. All 32 threads were pegged more or less.

Number of Moderate Tracks
Q: How many tracks can I run with reasonably efficient FX chain?
(Arturia Wurlitzer V > Waves REQ 6 > Supercharger GT > SoftTube Tape > H-Delay Stereo)

7700k: 54 Tracks
3950x: 179 Tracks

Conclusion: Roughly 230% more tracks here. Again, was expecting a bit more, but we're at least getting to the neighborhood of expected results.

Project Load Time
Q: How fast does the CPU load a project?
(Same Kontakt/sample-heavy project, tons of Acustica libraries to open)

7700k: 2:00
3950x: 1:11

Conclusion: the 3950 helped load plugins and sample libraries about 40% faster. I didn't know what to expect here, since the storage medium is the same. For me, it's a matter of convenience, but I could see this directly impacting session work, particularly the rapid-fire "just need to fix a couple of things on each song" marathons.

Additional note: I picked this project because it pegged my 7700k at, effectively, 100%. With the 3950, the processor was running around 20-25%, which is directly aligned with the 300% theoretical increase I was expecting. I doubt anyone will actually realize 3x more muscle, but it's nice to see something close to it.

Project Render Time
Q: How fast does the CPU render a project?
(Same Kontakt/sample-heavy/processing-heavy project above)

7700k: 3:28
3950x: 1:38

Conclusion: the 3950 dropped render time by roughly 50%. I was expecting this to be more, and even though all threads are busy, there's some 50% headroom (again, we see the 300% theoretical improvement, but don't actually realize it). I'll chalk it up to a couple of 'uber' tracks that appear to be pegging the 4-core limit, as looks like eight or so cores are indeed pegged.


Other interesting Observations:

- It appears that Reaper, at best, allocates four threads to a single channel. I'm curious if anyone else can confirm this. This makes me want to test disabling CMT (hyperthreading)... stay tuned.
- This particular 3950 setup seems to peg the processor at about 4 gHz, according to Windows and CPU-Z. I suppose I can interpret this to mean AMD's CBP with an AIO cooler effectively results in a stock overclock of 500 mHz, for extended loads.
- It appears that each Acustica instance (and every other plug for that matter) claims new RAM space, instead of sharing RAM space for identical libraries. Not surprised, but now that running 30+ instances of Ivory is within reach, I'd say 32GB of ram is required here. I personally don't see needing 64, but I can imagine another user with heavier sample library work using that.
- BOO I SAY to copy protection, easily the single longest part of my rebuilds are tracking down legal licensing mechanisms across multiple plug-in vendors. Drives me crazy if I think about the various layers of irony.


TEST NOTES
- These tests are not scientific; they're designed to approximate real-world tests that answer very practical questions, in a reasonable amount of time and effort.
- Except for the 'low latency tracking' test, the ASIO buffer was set to max (8192 samples, ~186ms) for all other tests
- An L1 was strapped across the master for all tests, as I don't want to die yet
- I did not reboot the PC or restart REAPER between tests, and other small things - YMMV as always


RIG NOTES
- Audio Driver: Audient ASIO 3.2
BOTH: Stock speeds, Audient id14, 32GB RAM, Win1064 1909, NVMe drive, Open air flow case
7700k: Gigabyte G1 GA-Z170N-Gaming 5, 120mm air cooling tower
3950x: Asus ROG Crosshair VIII Impact, AIO 240mm water cooler

[SmajjL]

I would not be sad if you did a MIDI chord loop (un-scientific fast) thingie with U-he Diva (some heavy patch) on your AMD and how many duplicate tracks you can go before it crackles.
I'm on the 7700K myself and AMD could actually get me on the next upgrade, finally!
Very curious.

[Pinknoise]

I'd be interested in a guitar amp, IR loader + fx chain in realtime see how much you can throw in there with very low latency.

And/or an intensive vocal chain.

[cassembler]

Originally Posted by SmajjL

I would not be sad if you did a MIDI chord loop (un-scientific fast) thingie with U-he Diva (some heavy patch) on your AMD and how many duplicate tracks you can go before it crackles.

Absolutely, but I'm not rockin' Diva (have looked at it for years but could never talk myself into it). However, this is roughly what I'm planing for multi-track test, stay tuned. Thinking Massive through some Acustica stuff?

Originally Posted by Pinknoise

I'd be interested in a guitar amp, IR loader + fx chain in realtime see how much you can throw in there with very low latency.

I plan on snagging some latency tests, but I haven't had much success with a low buffer setting and anything beyond Guitar Rig 5 or Amplitube, at 2.4ms (which I believe is 4.8ms round trip, lowest I can get from my iD14). With Mercuriall stuff, up to 3.9ms (7.8ms). Beyond that, it's too much gap for me, so any tracking I do is pretty dry (DI guitar only, and terribly played at that).

I'll run the same test after the 3950x, but in short, a single instance of Mercuriall with some basic fx is about all I can eek out of the 7700k for tracking (again on the iD14).

I'll say that Reaper does appear to scatter even a single channel across multiple threads, which makes me optimistic.

What would your personal "low latency" value be, and I can just target that?

[SmajjL]

Does not really matter, but if it is a Hog enough so you don't need to go 10.000 tracks would be good.
Curious for the result vs a 7700K, hope it kicks ass, and slaps..

[airon]

I would expect the performance to be slightly better on heavy tracks, since the IPC is a bit better on the new Ryzens.

You have to wonder if Reaper can dedicate a CPU core to live input. Maybe that live fx processing setting in preferences/buffering is key to this.

This is about the same build I'm likely to go for this year. Btw, why did you pick the Asus ROG Crosshair VIII Impact ? I'm going to ask the audio interface manufacturer about compatibility before I buy the mainboard. Not sure if it'll work on USB 3 ports as well as it works on the USB2 port I have it running on right now.

[EvilDragon]

You can download Diva demo for benchmark purposes. You don't have to own a license, it will run in demo mode with occasional crackles.

[SmajjL]

Yup, and I don't recall U-he installs *crap* in a wierd way so you need to take a shower afterwards.. you're the boss.
And I read that AMD loOoves and prefeer fast Hz RAM (more than Intel)
Not sure if it is in general or a gameing thing, but I would not go below 3200Hz certified sticks.
Just a heads up if your finger is triggering shopping sites.

[cassembler]

Originally Posted by airon

Btw, why did you pick the Asus ROG Crosshair VIII Impact ?

Good question, it's an odd board (mini DTX form factor, appears geared toward overclocking, which I don't typically do...)

A few things factored into the decision:
- The form factor is perfect for the chassis (Thermaltake Core P1)
- The general OC support, specifically the power rails, were about as high end as I could find on mITX (or mDTX in this case). It's designed for OC, but also great for general stability and durability at stock. Very arguably a waste at stock, but I defend the philosophy.
- Two M.2 devices on one insert card, I don't have to pull out the board to get underneath to swap a drive
- Physical layout of the board/aesthetics (again, open-air case)
- Wanted X570 for future potential, particularly PCIe 4/USB 3.2

It's twice more than I've ever payed for a mobo, and I hope I don't regret it. BUT, my last AMD experience was driven by budget, and was a mediocre experience - this one is driven by power... I didn't want to skimp.

A close second was the Gigabyte X570 Aorus Pro mITX version, many other good options too.

Originally Posted by EvilDragon

You can download Diva demo for benchmark purposes. You don't have to own a license, it will run in demo mode with occasional crackles.

You and your logic Will kick the tires on it on the new rig.

Originally Posted by SmajjL

And I read that AMD loOoves and prefeer fast Hz RAM (more than Intel)
Not sure if it is in general or a gameing thing, but I would not go below 3200Hz certified sticks.

Already using a 32GB 3200 Corsair kit rated at 3200. There's better RAM out there, but that's another day

[EvilDragon]

Originally Posted by cassembler

You and your logic

_\\// Live long and prosper!

[SmajjL]

Originally Posted by cassembler

Good question, it's an odd board (mini DTX form factor, appears geared toward overclocking, which I don't typically do...)

--

Already using a 32GB 3200 Corsair kit rated at 3200. There's better RAM out there, but that's another day

Good Man
And will also eyeball small/compact next time, but I guess a House is portable if one can lift it, just saying..
I just only now went all SSD/M.2. and could throw the old-school spinning thing out the window (err, in the closet for just-in-case backup), just need to backup and react on the coffee first, good computer times!

Ps, I recall a slight difference with Diva when going Hyper-Threading ON/OFF, some things react good, others might not and maby AMD has a similar thing, dunno.

[Pinknoise]

Originally Posted by cassembler

Absolutely, but I'm not rockin' Diva (have looked at it for years but could never talk myself into it). However, this is roughly what I'm planing for multi-track test, stay tuned. Thinking Massive through some Acustica stuff?



I plan on snagging some latency tests, but I haven't had much success with a low buffer setting and anything beyond Guitar Rig 5 or Amplitube, at 2.4ms (which I believe is 4.8ms round trip, lowest I can get from my iD14). With Mercuriall stuff, up to 3.9ms (7.8ms). Beyond that, it's too much gap for me, so any tracking I do is pretty dry (DI guitar only, and terribly played at that).

I'll run the same test after the 3950x, but in short, a single instance of Mercuriall with some basic fx is about all I can eek out of the 7700k for tracking (again on the iD14).

I'll say that Reaper does appear to scatter even a single channel across multiple threads, which makes me optimistic.

What would your personal "low latency" value be, and I can just target that?

For me 3.5/5.0 ms roundtrip seems to be working fine picking guitar around 170 BPM. I would expect song over 200 bpm to need even lower latency though.

I mostly use either Kuassa or Kazrog amp. With Kazrog I need to keep things at minimum quality and can't add much FX but I rarely track with those.
I have a 3770k Ivy Bridge 16 gig Ram.

Planning for Ryzen my next build. To say I 'need' it is a bit exaggerated I just want it. The extra will be appreciated for rendering but for everyday song building and just fooling around with my toys my computer is still capable wich makes me hesitate to take the jump. But I like the idea of having a backup computer I probably developed a dependency by now !

[cassembler]

Posted updated results.

As a side note, this was a particularly challenging build, as something around the way I was changing BIOS settings and/or adding/removing hard drives would cause Win10 boot to fail - I'd have to repair startup errors, at which point the previous batch of software installations would have to be redone (i.e. installing plugins). Very strange symptom.

Otherwise, I have to give +1 to Cockos, again, for how easy it is to manage the software. Seriously, <120 seconds download to install to license and a sip of coffee to boot. (In contrast, Wavelab was roughly 45 minutes, including downloading a downloader, troubleshooting the downloader, troubleshooting the E-Licenser donlge that caused me to have to tweak BIOS settings. Boo! I don't miss Cubase!)

[SmajjL]

So, it is safe to say it goes well as a DAW-station, specially if you're a VST person?
Thank you for sharing your tests cassembler.

[enroe]

Originally Posted by cassembler

Results!
7700k: 10 Tracks
3950x: 31 Tracks

Conclusion: The 3950 bakes up about 210% more here. I thought this should be a bit higher - we're roughly expecting 300% increase in multi-core performance. All 32 threads were pegged more or less.

Mmmhhh ... ?

When I consider 10 tracks being 100%, then 31 tracks are 310%. So the
AMD actually excels the expectations.

[cassembler]

Originally Posted by enroe

Mmmhhh ... ?

When I consider 10 tracks being 100%, then 31 tracks are 310%. So the
AMD actually excels the expectations.

But in this case, it's only 210% _more_, where there are 300% _more_ cores.

I expected something closer to 35-38 tracks or so.

[cassembler]

Originally Posted by SmajjL

So, it is safe to say it goes well as a DAW-station, specially if you're a VST person?
Thank you for sharing your tests cassembler.

Yes, as far as I can tell this thing is a beast, and Intel's offering (10980XE) just isn't the same value, even at half price... I've always been an Intel guy, but I'm super glad to see AMD back.

And of course, I'm happy to contribute; this community has helped me a ton