In the age of artificial intelligence, the demand for GPUs like the H100 has skyrocketed, making it difficult for the average consumer to get one. However, A.J Reddit user He came up with a unique solution, converting an AMD Zen 2-based Ryzen 5 4600G “Renoir” APU into a 16GB GPU and then using it for AI workloads on Linux. You know, as they say, modern problems require modern solutions.
Legacy AMD Ryzen APUs provide decent performance on AI workloads after conversion to GPU
Before we get into how to convert an APU, let’s take a quick look at the AMD Ryzen 5 4600G. The Ryzen 5 4600G is known to be one of the best APUs in the market after it was eventually replaced by its Cezanne counterpart. It features a 6C/12T configuration with Radeon Vega iGPU with seven CUs (Computing Units). To explain how to achieve the 16GB VRAM mark, it’s necessary to note that APUs support “shared memory,” where you can allocate 50% of your RAM capacity to your APU. In this case, this Reddit user had 32GB of DDR4 memory, giving half of it to the processor.
The next big hurdle is actually running AI workloads on a Ryzen APU. If you have a desktop GPU, you can use AMD’s ROCm (Radeon Open Compute) platform to run AI applications on Linux. However, in the case of iGPU, third-party packages allow ROCm to run on APUs, which was also used here. Using ROCm solves most of your problems as you can now run every type of AI application from Tensorflow to PyTorch.
In a detailed video, a Reddit user shared his interesting experience, claiming that the Ryzen 5 4600G can handle all kinds of AI workloads. However, it only showed the Stable Diffusion test, and to our surprise, the APU managed to generate a 512 x 512 image in about 1 minute 50 seconds. This is a critical event for the APU, and we believe it could be a hack if used properly.
While you can’t compare the Ryzen 5 4600G to an AI GPU, like the H100, this experiment shows the untapped potential in Ryzen APUs. Imagine that instead of the Ryzen 5 4600G, the latest APUs from Phoenix are used here, which are much better in performance. However, offering such AI performance for $95 is undoubtedly something to be appreciated, and hopefully provides the “average consumer” with a way to meet their AI needs.
news source: tom devices
“Certified food guru. Internet maven. Bacon junkie. Tv enthusiast. Avid writer. Gamer. Beeraholic.”