AMD will collaborate with Rapt AI, a GPU workload management software maker, to improve AI model inference and training on AMD’s Instinct processors.
On Wednesday, the companies said they will work together to integrate Rapt AI’s workload automation platform with AMD’s Instinct MI300X, MI325X and upcoming MI350 series GPUs. The Rapt AI software will help enterprises with resource allocation and GPU management, and assist them in fixing performance bottlenecks, the companies said. In the future, Rapt AI will also focus on memory utilization.
Enterprise IT departments struggle to determine whether data center resources are fully utilized at all times, according to Jack Gold, principal analyst at J.Gold Associates. The first step to reaching maximum use of resources is to monitor them, which is what Rapt AI software does with GPUs.
The company’s technology tracks workloads on every GPU and makes sure that each runs at full capacity. Getting the most out of each chip while monitoring performance means that companies will buy more GPUs only when they’re needed, he said.
“It’s really about cost savings,” Gold said. “If you’re overbuying compute, then you’re paying for something you’re not using.”
Rapt AI software supports AMD and Nvidia GPUs in a private or cloud data center. It also supports tensor processing units available on Google Cloud and AWS’ Trainium AI accelerator. A TPU is an application-specific integrated circuit, or ASIC, designed explicitly for AI applications’ high-volume mathematical and logical processing tasks.
Rapt AI’s collaboration with AMD means that its software will ensure that the software maker’s technology will perform at the highest level on Instinct, Gold said.
“They’re going to work a little more closely to make sure everything is optimized,” he said.
AMD’s Instinct GPUs compete with Nvidia H100, H200 and Blackwell GPUs for model inference, training and fine-tuning. In the fourth quarter that ended last December, slightly more than half of AMD’s revenue came from sales of its Instinct GPUs and Epyc CPUs, both data center chips.
AMD plans to start shipping the Instinct MI350 in volume in midyear. The company designed the GPU for AI workloads in cloud providers’ data centers and large enterprises, such as pharmaceutical companies, financial institutions and retail giants.
AMD plans to ship the MI350’s successor, the MI400, in 2026. The MI400’s CDNA 4 microarchitecture tightly integrates networking, CPU and GPU capabilities at the silicon level. AMD’s programming platform for Instinct processors is called Radeon Open Compute, or ROCm.
Source: https://www.techtarget.com/
