AMD Aims to Drive On-Device Gen AI with Its New Ryzen Embedded 8000 Series Processors
Delivering a claimed 39 TOPS of total compute between CPU, GPU, and new NPU, these embedded chips pack a punch.
Chip design giant AMD has announced a new family of Ryzen Embedded processors featuring an integrated neural processing unit (NPU) accelerator delivering up to 16 tera-operations per second (TOPS) of compute for on-device edge artificial intelligence (edge AI) and generative AI (gen AI) workloads: the Ryzen Embedded 8000 Series.
"The Ryzen Embedded 8000 Series processors [are] the first AMD embedded devices to combine NPUs based on the AMD XDNA architecture with traditional CPU and GPU elements, optimized for workload versatility and adaptability targeting industrial AI applications," the company claims in its announcement, brought to our attention by CNX Software. "Embedded solution engineers and developers can harness the processing power and leadership features for a variety of industrial AI applications including machine vision, robotics, and industrial automation."
The chips themselves use AMD's Zen 4 architecture, with up to eight cores and 16 threads depending on model. There's an integrated graphics processor (iGPU) based on the RDNA 3 engine with up to six workgroup processors, again model-dependent, but the headline feature is the XDNA-based NPU — delivering, the company claims, up to 16 tera-operations per second (TOPS) of compute independently from the CPU and GPU cores.
"With these resources in play, the full processor performance can scale up to 39 TOPS," AMD claims, "with up to 16 TOPS offered by the NPU – a powerful profile targeted for demanding, diverse industrial computing workloads. The increased TOPS is achieved by the combination of AMD architectures, AMD XDNA in the NPU, AMD RDNA 3 in the GPU, and Zen 4 in the CPU, that work together to optimize overall system performance."
AMD has confirmed four models for launch: the Ryzen Embedded 8640U, 8645HS, 8840U, and 8845HS. The U-suffixed variants have a 28W target thermal envelope, with six cores running at 3.5GHz based and 4.9GHz boost and eight cores running at 3.3GHz base and 5.1GHz boost respectively. The HS-suffixed variants offer a boosted clock speed, at 4.3GHz base and 5GHz boost and 3.8GHz base and 5.1GHz boost respectively, at the cost of an increased target thermal envelope of 45W.
The company promises support for popular AI frameworks, including PyTorch and TensorFlow, through a software development kit — which, at launch, will be exclusive to Microsoft Windows — with pre-trained models distributed through Hugging Face.
As well as direct sales, the new chips will appear in industrial systems from original design manufacturers including Advantech, ASRock, and iBASE — though pricing and availability have not yet been confirmed.
More information is available on the AMD website.