We gathered a few musicians and tested many CPUs’ performance in Cubase. These are what we got. We haven’t made a video for English version yet but we can compare the graphs here. Details about the test platforms and how we set up the tests are also in the video above.
As for the Bilibili platform: If the website pops up a window asking for logging in, just close it and continue playing the video.
AFX in Reaper has a even higher default buffer (200ms) than ASIO Guard High (~100ms) in Cubase, and some big.LITTLE Intel CPUs are less problematic with DIVA in reaper:
All ASIO Guard OFF. Because we encountered many weird results, probably due to threading issue with Kon7 + ASIO Guard On, making them not directly comparable.
and we also tested K7 with Spitfire. Apple M-series get consistently low scores. Combined with other results we assume that RAM Latency matters for Kontakt sample engine.
High Consumption FX Chain (Serum2 → Amplitube + ASH + O7 +…) Track Layering Test:
Note: Performance of Intel U7-265K here is not very stable. Most of the time it is very powerful but sometimes it could even drop to Zero. Threading Problem.
RAM Bandwidth and RAM Latency Matters; CPU Cache alleviate a little.
There are also some more information in these posts that I wrote before:
Most of us in the project are composing amateurs, it is very time consuming to get done all these test. If you have any question or suggestion, please feel free to tell us.
Thanks for everyone in the test project, and the Steinberg forum, and you.
For example, in Kontakt 7 tests, the result number stands for the highest note count one device could handle before the sound starts to glitch / performance meter gets red. So the more powerful one device is, the more notes it could play without glitch, the longer the bar.
Is there a pattern and logic behind the different sample rates and buffer sizes?
To me it seems like the best practice would be to focus on settings that people find usable. In other words an instrument playing back at 1024 (hypothetically) would yield a really long delay that a musician would object to if playing live. A higher sample rate would help of course and increase the load. I don’t play virtual instruments live but I’ve seen people prefer buffers of 64-128, with outliers demanding 32 and some being ok with 256. In contrast Dolby Atmos is (was?) locked to 512 samples, so for a lot of people mixing Atmos that isn’t just acceptable but is (was?) required.
In other words, realtime VSTi use = lower buffer, mixing = higher buffer.
I think , after looking thorugh all the metrics/graphs, this shows that for Cubase a high end Windows machine still is the best performer, the M4 is good on single core stuff but overal the windows CPU’s are still better performers, and they didn’t test the 285 either which is the best performer at the moment. And, the M4 mac will be twice the price too..
For mobile use I still think an M4 MBP is one of the best machines as a portable DAW platform.
i dont think so.. i would think mac mini m4 max would clean up the pc’s because in a daw multi core can only be as good as the best single core - like i know you know all your plugins in a chain would be on one core.. for me m4 max is winner
Those look like they’re the ‘regular’ M4 as well (except for the Max of course).
Aside from it only have 4 performance cores and 6 ‘efficiency’ cores, the base M4 has a FSB speed of half what the Pro CPU does (and the SSD read/write as well). The Pro chip has 8 Performance/4 Efficiency cores and a FSB of 273GB/s vs 120.. I would think the Pro/Max/Ultra would eat the base CPU alive in most cases..
You’re right. So actually it says “DDR4 3200C16 (Simulated)”. It is actually single channel DDR5 6400C32 to reach about the same latency and bandwidth with DDR4 3200C16. We know there are other differences but here we just wanted to stress how memory bandwidth is important to run a large amount of plugins.
We can speculate on M4 Pro Multi-core, though. Given that M4 Pro shares the same memory bandwidth limit with M4 Max (yes, M4 Max CPU clusters cannot make full use of those 8 channel LPDDR5X, they are limited to ~270GB/s which is the same to M4 Pro CPU clusters. Only GPU cluster can use all the ~540GB/s bandwidth), the P-core (Everest V3) is about 3 times the performance of E-core (Sawtooth V3) and multicore performance increases almost linearly when there aren’t too many cores, we can just do simple maths.
Just assume 3 E-cores = 1 P-core, for example. Then M4 6e+4p = 6p(45%), M4 Pro 4e+8p = 9.33p(70%), M4 Pro / M4 Max 4e+10p = 11.33p(85%), M4 Max 4e+12p = 13.33p(100%).
We tried this method on M4(6e+4p) and M4Max(4e+12p) It matched our observation with some extent of accuracy in tests that are not bandwidth bottlenecked. But don’t take this as a solid reference, anyway.
I would be curious to know how well some of these performed on a HEDT system. I was thinking of going back to an older threadripper platform. The RAM bandwidth would be there for sure, but high clocks and die latency would not.
To all those are confused about the graphs: We are gradually working on translating the whole report into English so all the meanings should be clear for you then…But I can’t guarantee when could we get this done, because all people involved in the test are busy in the day, and we are not English native speakers. My point is that you can see the relative value to make a general impression about one cpu’s performance. Sorry about all the inconvenience.
Similar here, on 44.1Khz 24bit source material yesterday; my deliverables were well received and approved by artist/director/producer. Last week too, on some live concert recorded (multitrack) material… (show was back in late July).
Does that make me a bad boy.? A nobody.?
Diving deeper, there are more opinions out there - this google search returns some helpful responses:-
Is 44.1Khz 24 bit audio ok for editing and mixing.?
