AI optimization is indeed a pending issue for all hardware and software companies. As we’ve seen several times before, most recently with NVIDIA, which increased its performance by up to 5 times, Intel doesn’t want to be left behind. They have announced a significant performance improvement, achieving a 170% increase with their GPUs in Stable Diffusion, especially with the Arc A770. This is all thanks to the latest driver and Microsoft Olive, which is positively impacting hardware enabled with specific AI units.
Intel is enhancing its drivers everywhere. It’s the weakest aspect of computing for gaming and AI, and therefore, after stabilizing their working and development teams, improvements are being introduced frequently for both games and AI software. In this case, it’s also true, although there’s a “catch”.
Microsoft Olive, the toolchain for optimizing models with hardware recognition
It’s a significant part of what Intel has gained with its driver. Olive’s model aims to unify any hardware with specific AI units to devise effective compression techniques, compile executable models, and optimize them.
It’s a solution for optimizing a model effortlessly and quite effectively, as stated by Microsoft themselves:
By providing a configuration file that specifies your model and scenario-specific information, Olive adjusts the optimization techniques to generate the optimal models on the “Pareto Frontier” based on the metric objectives you set. When working with the configuration file, you typically need to provide information about the input model, including input names, shapes, and the location where the model is stored.
Additionally, you specify your performance preferences, such as desired latency, accuracy, or other relevant factors. Along with this information, you can choose from a variety of optimizations provided by Olive that you want to apply to your specific hardware target. You also have the option to define the target hardware and use any additional features offered by Olive. By using the configuration file, all you need to do is run a simple command, eliminating the need for any Python code.
The Xe Matrix Extensions make the “miracle” of Intel Arc and Stable Diffusion possible
Of course, not everything depends on Microsoft Olive. We already saw how with ONNX and driver 31.0.101.4900, performance increased significantly, where after switching to Olive as an optimized model with the same driver, it doubled.
However, this latest Intel driver for Arc, 31.0.101.4952, has boosted performance from the starting point to a +170%, or if we want to compare driver to driver in Olive, a 70% increase in AI performance.
Naturally, Olive is more effectively utilizing the XMX instructions thanks to the driver, which specializes in executing scalar products and accumulated operations in 2D systolic arrays. Therefore, Intel has a considerable stake in the impressive improvement in Stable Diffusion with just a couple of drivers, once again following in NVIDIA’s footsteps and maintaining their pace and momentum.
The post “Intel increases the performance of its Arc GPUs in Stable Diffusion AI by 170%” first appeared on El Chapuzas Informático.