Enlight me, because every thread I participated and every experiment I have seen on this forum with mixed RAM capacity didn’t work both at full speed and in dual channel mode.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
2020 iMac Performance Problem?
- Thread starter mahaha
- Start date
- Sort by reaction score
I thought someone figure it out, but maybe you’re right. I didn’t think a bout the dual channel. Speed was the main concern and I know that can be addressed. In any case, I actually agree with you. I don’t like mixing memory. I plan to run the 64GB I bought and discard the 8GB that will come with the machine. I prefer all modules match across the board, including capacity. That’s probably just OCD thought.
At the risk of becoming collateral damage
:I agree with you on the memory. There is no practical situation where 72GB in single channel (or at lower clock speed) would be better than 64GB in dual channel. Maybe one could artificially contrive some such situation to prove a point, but not in realistic use. It would basically degrade performance every single day.
Don't need to take my word for it, see 7:02 in this video: (I know you know this already, but other readers may not):
Regarding editing RAW photos - I do a lot of this, and had 2GB VRAM for many years with no problem. So 4GB should be fine, except perhaps for some extreme cases that I do not have experience. The 8GB VRAM with the 5500XT I have now is really wasted for photo editing purposes - but nice for other things (video, games). Batch processing of photos is more CPU intensive, so a better CPU can shorten the time to do so, but actually this is not a frequent task (at least for me).
Hope this helps.
Last edited:
I’m in the (apparently pretty common) boat of having one of these beasts and debating whether to just leave the 8GB that shipped with it sitting on the desk. My starting point was this set of BareFeats tests. Essentially if you have dual channel with either two (properly placed) or four identical sticks, you get about 23GB/s. If you have mismatched dual-channel (properly placed but differently-sized sticks), you get about 19GB/s. If you install four sticks in the wrong order so you’re in single channel but have both banks available, you get about 13GB/s. Wrong order (single channel) with only two sticks is even lower, down around 10-12GB/s depending on brand.
So obviously matched sticks are best, but you take a much bigger hit from running single channel than mismatched dual-channel; ~20% vs ~45%.
In any case, it seems pretty obvious that unless your use case involves needing exactly 70GB of physical RAM to avoid swapping you will on average come out ahead nearly all the time with the faster 64GB, but I guess I’m still left wondering, from a speculative standpoint, how much real-world impact a 20% hit in RAM throughput will really have in day-to-day “general prosumer” computing.
So obviously matched sticks are best, but you take a much bigger hit from running single channel than mismatched dual-channel; ~20% vs ~45%.
I’m curious about the lower clock speed claims; I haven’t tested yet, but the 8GB stock RAM I pulled says 2666 right on it, so it seems like it shouldn’t reduce the bus speed, and the Barefeats test linked above shows essentially the same performance with 4-32-4-32 as with 3rd party (all 2666) 16-32-16-32.
In any case, it seems pretty obvious that unless your use case involves needing exactly 70GB of physical RAM to avoid swapping you will on average come out ahead nearly all the time with the faster 64GB, but I guess I’m still left wondering, from a speculative standpoint, how much real-world impact a 20% hit in RAM throughput will really have in day-to-day “general prosumer” computing.
Last edited:
Sorry for the delay in getting back to you all, I was away on vacation. I'll try removing the apple ram to see if that does help.
Interesting about the cc background tasks, we haven't done anything with those, they are all fresh installs, not even a transfer from our old mac pros.
We aren't using a font management system (nor were we on the mac pro's) and while we do have quite a few fonts in there, I would have to look to see how many, it's not like we're running the adobe library.
Hopefully it's just the ram issue. We're going to be assembling a 15-20gb file soon and really want to work this out before then.
Thanks!
I don't game, but I opted for the 5700 XT 16GB with the hopes that it can be leveraged for ML later in the future. Do you think that's a possibility?
I feel like Beachballs are just a way of life now with Apple. I love their stuff but my iMac Pro stutters doing very basic editing tasks from time to time. I am fine with it when its heavy video editing, but it shouldn’t happen with lighter work and get it does.
I have never really been a big fan of Catalina and I do think a lot of the issues are software related. I never had as many beachballs with Mojave so I hope Big Sur corrects some of these problems.
I have never really been a big fan of Catalina and I do think a lot of the issues are software related. I never had as many beachballs with Mojave so I hope Big Sur corrects some of these problems.
You just can’t do anything on AMD GPU regarding machine learning. No library officially support AMD GPUs, even less on macOS.
I was under the impression Core ML can and will use AMD GPUs, and in fact since that's all Apple sells is presumably optimized for them; I've seen a few benchmarks confirming that (which also indicated that more GPU RAM may improve performance for large datasets).
That said, I'm not a Mac app developer, and I haven't done a lot with ML, so I could be misunderstanding something here.
For me, CoreML isn't a real ML library. A real GPU-accelerated deep learning lib is Tensorflow or Pytorch... These are real libs you can use to create, train, validate, and deploy a model. Good luck doing the same with CoreML. It might be possible but you'll be alone in your world.
Didn't know PlaidML thought. Thanks for the link. It might then be possible to compile Tensorflow using this backend. But this still look very experimental and not appropriate for production. Still prefer PyTorch + Nvidia + Linux Ubuntu.
Last edited:
The uncompressed image file is kept in system memory not VRAM, unless it's using OpenGL or Metal buffers (Photoshop has this option). Even then, it's measured in 100s of megabytes. The image in the Photoshop window (the pixels being drawn to the screen) are kept in VRAM but only use about 150MB for a 5K 10bit screen like the imac.
VRAM makes little difference to most people, outside of gaming or gpu rendering.