The B450 chipset is being overwhelmed by the GPU due to excessive workload or resource demand.
The B450 chipset is being overwhelmed by the GPU due to excessive workload or resource demand.
I developed my machine learning setup and encountered a curious issue. On the MSI B450 Tomahawk Max with a 5600G CPU, I added an Arc A770 16 GB to test Stable Diffusion in the main PCIe slot connected to the CPU – everything worked fine. To enable CUDA support, I inserted my older NVIDIA GTX 1650 Super (4GB) into the PCIe 2.0 x4 slot provided by the chipset. I also connected an SSD via its SATA port, which was fine. However, this configuration led to random stuttering during YouTube playback and occasional system freezes. I realized that connecting the OS drive to the SATA port directly linked to the CPU resolved the problem. It appears that when a GPU is placed in a PCIe slot offered by the chipset, other components can’t function properly. This issue appears across both Windows and Linux environments. When training neural networks with Keras on an NVIDIA GPU running Linux, the system became unresponsive, the GUI lagged, and restarts were necessary. Copying data between drives behind the chipset while using the GPU also caused delays – the transfer would stall for several seconds. Ultimately, the main challenge seems to stem from bandwidth limitations: the GPU uses PCIe 2.0 x4 lanes, while the chipset only provides PCIe 3.0 x4 lanes, leaving insufficient capacity for both the GPU and additional storage, despite the integrated sound card handling audio without issues.
It’s not really my thing to believe your concerns are about chipset bandwidth. It could very well be a power issue. Running two GPUs can push up to 150W onto your motherboard when everything is running at maximum, which might cause problems in similar setups. I remember this happened years ago with four GPUs on one system—so many miners added extra power connectors for PCIe slots to handle the load.
It's best to check your specific power supply model. The A770 prioritizes strong power delivery rather than PCIe support, so confirm with your PSU's specifications.
Have you attempted to run the setup without the A770 to check for similar issues?
The issue wasn't tied to the A770 model specifically; there were two GTX 1650 Super cards in the system. I tested Blender with BMW rendering on a GTX 1650 Super and Stable Diffusion 2.1 on an A770 while playing YouTube, but minor stutters appeared—mainly noticeable during scrolling. At the moment, the build works fine and hasn't caused any problems. I'm confident it's related to the chipset, or unlikely it was due to a faulty SATA cable since I used another one that was already connected properly.
I request only basic elimination methods. I’d evaluate each individual part unless you’ve already limited it to a consistent case where including the GTX 1650 Super in that position causes the issue.
The prolem appeared before, in Linux I relied on it to link displays to the card via a slower PCIe 2.0 x4 slot and executed ML on the card using a PCIe x16 slot. Previously, I had Intel Optane in M.2 for testing (on an AMD platform...) which shared with SATA, so all my drives were connected through the SATA chipset. Also, it wasn't a major issue—kernel 5.15 HWE generally handled both, but input stuttered for several seconds to minutes even after completion. I would shuffle hardware because of national holidays and moved the Optane out, which sparked the idea to connect the drive directly to the CPU and improve performance. Now I won't run more experiments since it seems to work, but I remain curious about the real cause, as I didn't search for anyone else reporting this. Looking at the MB diagram, the ALC892 isn't even routed through the chipset, which adds another layer of confusion...
5600G features a solid-state die, with part of it being the 'uncore' similar to the I/O die found on regular Ryzen models. It appears to have a much more restricted PCIe setup compared to other Ryzen CPUs. I'm curious about running it without the A770, as it might be possible to fully utilize the PCIe bandwidth to the CPU since both the north and south bridges have limited data transfer capabilities. However, if testing without the A770 still causes problems, adding active cooling to the chipset could be worth considering to see if it resolves the issue—possibly because the chipset is being overloaded. You're also investing a lot of time optimizing a 5600G, which seems excessive. This isn't a processor design, and the B450 south bridge isn't built for dual GPUs. Another concern is running APIs that need specific bandwidth from the GTX 1650S, which might not be fully supported at 2.0 with four lanes.
When all else is off, YouTube in Chrome runs smoothly while Blender renders on the GTX 1650S in that slot. Once rendering begins (CUDA or OptiX), a brief sound crack occurs as data moves to the GPU. Using HIP (5600G APU) or oneAPI (A770) eliminates this issue entirely. The only concern is excessive use of unified memory in Keras, which can slow performance. I once adjusted ai-benchmark to run with unified memory; previously, even with both cards active, the system froze—xorg and KDE nearly halted. However, having one card in PCIe 2.0 x4 had minimal impact on speed. >> AI-Benchmark-v.0.1.2 >> Let the AI Games commence... *TF Version: 2.3.0* *Platform: Linux-5.4.0-47-generic-x86_64-with-glibc2.29* *CPU: N/A* *CPU RAM: 31 GB* *GPU/0: GeForce GTX 1650 SUPER* *GPU RAM: 15.3 GB* *GPU/1: GeForce GTX 1650 SUPER* *GPU RAM: 15.3 GB* *CUDA Version: 10.1* *CUDA Build: V10.1.243 The benchmark is running... Tests may take up to 20 minutes Please don't interrupt the script ... Device Inference Score: 4907 Device Training Score: 3528 Device AI Score: 8435 For more details and results, visit http://ai-benchmark.com/alpha I’m okay with some sound distortion, but it mainly happens when I find ways to fix the issue and have a bit of time to relax. That’s why I became interested in this. Accepting that using the GPU in a PCIe 2.0 x4 slot requires nothing else connected—especially drives—to be active at once, seems odd.