The above would indicate that AMD is very competitive with Intel and good value when running DSP and memory-intensive VI (i.e. Library Sample Playback) across the typical latency settings, but I’m not completely convinced that those benchmarks would translate closely to how well VCV performs, for multiple reasons, so…
…I’m interested in hearing about any experiences along the following lines:
VCV performance on Zen2 Ryzen relative to Intel
VCV performance scaling above 8c
VCV performance advantages when AVX-512 is available
On Ryzen 5 laptop, vcv rack has been stunning. I prefer it to the intel and mac. Drivers, elektron digitak overbridge, are solid. on aside note - the Touch screen on my laptop is going well. I’m glad I went back to amd.
I have a PC with a first generation Ryzen 1700x and it performes flawless,
mostley I need 2 to 3 cores only.
but a good graphics card is for good VCV performance important too.
at least as much as a strong cpu.
I would suggest you save some money on cpu and invest it in the graphics card.
I’m curious about the topic. I’ve read online that AMD cpus translate in a very high roundtrip latency because the 32cores are actually two 16core chips paired together. Can someone explain it better to me ?
i just made this choice myself. i actually upgraded my pc with a new motherboard, cpu, and ram last friday the 13th (and it was a full moon too)!
because vcv rack does need good single core performance, in a time-critical manner, i chose to go with intel. yes, we now can use multiple cores, but if you have a resource-hungry module (like rainbow or portland weather) that still needs to be processed by one core, as far as i understand. and more cores also introduce more overhead.
so i think the i7 is the best architecture for our application. i chose the i7-9700kf (current latest gen) as best bang for the buck processor in my (medium-high) budget. i already had a geforce gtx-1060 video card for the graphics side.
Hey Ben! Sounds like your new PC is similar to the one I brought home yesterday: i7-9700kf cpu, 48GB ram, 6GB Nvidia 1060 gpu. Loaded VCV Rack and opened Omri Cohen’s Braids QPSK patch. It plays flawlessly @ 128ms latency, with instant screen zoom. About US$1200 for this new box, but it should last 10 years, as the box it replaces did! Now to spend the weekend reloading all my Native Instrument and Arturia libraries…!
Paul “Uncle Chrome” Artola
Ellicott City, Maryland USA
Right now I’m using a trusty, but long in the tooth i7-3770K
Clock is 4.2GHz, VCV framerate is 50Hz, Sample rate 48KHz, ASIO buffer 256, VCV threads enabled 1
Patch is VCO1 -> Portland -> Audio-8
Core usage on Portland is hovering between 45-46%. A second Portland gets slightly glitchy
Enabling 3 cores in VCV, I can get 4 Portlands running smoothly, but I can’t quite get 5 Portlands into all 4 cores.
My expectation for 3900X is to comfortably get 2 Portlands running on one core, and hopefully 3. Fingers crossed that then holds true for all 12 cores…
For GPU I went for RTX 2070 Super (I like to play a game, now and again!)
My interface is RME PCIe Raydat, which I’ll be bringing over to the new build.
Not sure if PC vendor talk is off limits or not. I’ll just say my order is with one of the major UK providers, popular with audio people, and is sold as an audio-specific workstation.
do-it-yourself is the way to go. you can select the exact parts you want, from the brands you trust, for the best prices. pcpartpicker is a great tool.
I agree about diy computer building, but I am currently thinking about replacing a dsektop and nettop with two laptops. I live in a fire prone area and the idea of the computers being grab and go is appealing. Also with power outages having the built in “ups” of a latop batter would be handy.
Thanks for pcpartpicker site though, that looks very handy.
Hi jhunolt! No dark secrets here. I was planning to build a Franken-PC, adding new mb (Gigabyte Designaire), cpu (i7-9700kf), gpu (6GB Nvidia 1060), 32GB ram, and 1TB of M2 system drive. However, my local Micro Center had a PowerSpec B742 (ASRock Z390 / i7-9700kf / 16GB ram / 500GB M2 drive) for US$800. I added the Geforce 1060 card and 32GB more ram for another $330. I have installed my two SSD drives for a total of 2.2TB.
So far, this new machine is a beast and worth every penny. VCV zooms in and out like a bad acid trip! Haven’t reloaded my video editing software to really bring the CPU to its knees, but audio renders are much faster.
My audio interface is mostly just the HDMI codec in my Onkyo NR-646 AV Receiver. I also have a Steinberg UR22, if I need audio in or 5-pin MIIDI. I also have a Resident Audio T2 Thunderbolt interface for my laptop. The ASRock motherboard in my system has Thunderbolt headers, but they want $100 for the expansion port interface, so I’ll pass for now.
In the end, I compromised a little by buying the prebuilt system, but I have built dozens of PCs over the last 20 years, so the thrill is gone for that!
My “monster” 3VCO / 3VCF 8-voice polysynth patch on the old machine was idling on a minimum of 3 threads at 60% total CPU usage
On the new machine it’s running on 1 thread comfortably - with 6% total usage
Plus the AMD is running this at a core speed of 3.6GHz (max turbo is 4.6GHz)
Plus the AMD is feeding an ASIO4ALL motherboard sound driver instead of my RME HDSPe Raydat
This is just an initial run, no optimizations yet, straight out of the box. I didn’t expect this thing to be 10X faster than my old box!
I’ve set up an all-core max speed of 4.2GHz on the 3900X
Adding Portland Weather to a patch as before, with the same basic settings (48KHz, 256 buffer), I’m not seeing the ability to add any further copies of the module beyond a thread count of 8. Here’s a set of figures, showing threads switched on and stable quantity of Portlands added to the patch:
Threads
Portland Weather Instances
1
4
2
8
3
12
4
13
5
15
6
17
7
17
8+
18
By comparison - here’s a set of figures using a much less expensive module - VCO-1
The instances - except for a single copy hooked up to output for monitoring - were not connected to anything, which doesn’t appear to affect the CPU usage for this module.
Threads
VCO-1 Instances
1
176
2
239
3
322
4
???
Curiously, when switching up to 4 threads and beyond, I had to reduce the number of instances to prevent glitching!
The figures above indicated a problem scaling a patch above a small amount of cores, but further experiments confirm that there’s a massive performance hit when Rack utilises cores belonging to multiple CCXs
With the current iteration of my polysynth patch, and using Process Lasso to control core affinity, the patch can idle comfortably when tied strictly to cores belonging to one CCX. Changing the affinity to the same amount of cores but each belonging to a different CCX causes it to instantly crap out at 100% usage, nowhere near playable.
3900X is a 12 core processor with 4 CCXs (3 cores per CCX) - so it looks like any single VCV patch can only use 3 cores effectively on this setup.
The likely cause for the performance hit is the L3 cache being separate for each CCX. And I bet this scenario doesn’t crop up on Intel - which has shared L3 for all cores, it seems.
Not the end of the world, really - within those 3 cores I can fit a lot… And I can run DAW and other platforms on other CCXs of course - but if I’m trying to get something xtra-large-scale going purely on Rack then I’ll have to think carefully about how to set it up (running multiple copies of Rack simultaneously - if that’s even possible?)
Kinda infuriating though, that the 3700X (8 cores, 2 CCXs), all else being equal, could potentially run bigger patches than its bigger brother
