hello everyone. I am using TVAI 4.0 version. Looking at the user benchmarks, I have a question. When using the 4090, there was no significant difference in the benchmark results for models such as Iris and Nyx, regardless of whether the CPU was 5950x or 7950x3D.
Most answers in the community say that VGA has a greater impact on work than the CPU, but in fact, VGA usage is limited throughout the work, and neither cuda nor the total share does not exceed 50%, while the CPU share is always over 85%.
According to general opinion, wouldn’t it be reasonable for CPU usage to be below or low to a certain level and for VGA usage to peak?
From what I deduce, You need a powerful CPU with 10 or more cores to get the most out of a powerful GPU.
RAM speed can also increases processing speed.
Thank you for answer. I am mainly working on 4k upscaling using the Iris MQ model and Nyx model. I want to maximize the VGA usage, but most models have low usage.
I am working on one instance by adding Nyx model as second to Iris MQ model. I tried increasing the number of instances to 2 or 3, but the work time for individual tasks only increased, and the increase in VGA market share was minimal.
This is what I do, and it does increase the aggregate throughput a bit. For me by about 40%, but does not saturate the compute hardware. So it is indeed one way to increase the throughput.
Yes, I also observe scaling limits.
My current assumption (hypothesis yet to be proven) is that I’m simply memory bandwidth limited. When I find the time I’ll pull out AMDs uProf (their variant of Intel’s vtune) and see what it can tell me about memory utilization, cache use and scheduling.
Not necessarily. It entirely depends on how Topaz coded the software. Not all tasks can be performed on GPU. I’m sure they offload as many tasks as possible to the GPUs, but some problems simply can’t be solved on GPUs. As such they need to be run on CPUs.
From what I’m seeing on my hardware utilization, for 4x upscale, it seems a significant part of the problem they’re trying to solve can’t be performed on the GPU, that’s why my CPU usage is so high (relatively higher than the GPU). Do keep in mind however that the GPU has insanely more capacity than the CPU, so even if Topaz is able to solve 95% of the problem of upscaling on the GPU, that the GPU will just blaze through all that work, while the CPU has to really struggle with that remaining 5%.
Thank you for your sincere reply. So, in 4k upscaling work, if I have a VGA of a certain level or higher (for example, 4070ti), the work speed can be seen to have a significant impact depending on the type of CPU used rather than the type of VGA used.
Well it’s a balancing act of the entire system and all its components.
For example, I have pretty slow memory and that seems to slow down everything for me. The GPU finishes its tasks very quickly, but have to wait for the CPU to do its share of the work before the GPU will be handed more work to do. The CPU on the other hand is partly idling because it’s waiting for slow memory to be loaded from RAM.
It’s like a traffic congestion
Optimally, one would have fast enough memory so that the CPU can be fed with data at minimal delay. That keeps the CPU busy. As a result, the CPU can then hand over new work quicker for the GPU to process. There will always be some bottleneck somewhere, but a good system is a well balanced one, meaning no single piece of hardware is unduly waiting for some other part. Or phrased a bit differently; all parts are fully utilized all the time.
That’s why there are special hardware for specific tasks, and using general purpose computing, we can never aspire to such efficiency. For me my goals tend to just be to ensure the CPU and GPU can be utilized around 80% which had been the general “intel architecture” limitation for decades. Unfortunately my machine can’t at present due to the slow memory. Even more unfortunate, there is nothing I can do about this situation, since the memory controller in AMD’s Ryzen 9 CPU generation can’t handle the speed required (the ordinary rated speed of mid-tier DDR5 memory) while also handling the amount of RAM I need (128GB). So I’ll just have to write this hardware generation off as a gigantic fail on AMDs part, and go with Intel next time.
Edit: Here’s good rule of thumb; You want your system to be bottlenecked on the most expensive part of it.
