Amd polishing AI strategy: Xilinx, GPU and software plan emerge

After regaining its foothold in the data center in the past few years, amd now hopes to cover all aspects from the edge to the cloud by expanding its chipset and become an important participant in the AI computing field.

This is a very ambitious goal, especially now NVIDIA is dominant in this field with GPU and CUDA Programming Model, and there is increasingly fierce competition from Intel and several other manufacturers.

Recently, amd held the 2022 financial analyst day. At the event, senior executives said that AMD has suitable chips and software to expand its territory in the broader AI field.

Lisa Su, CEO of AMD, said in her opening remarks: "our vision is to provide a broad technology roadmap covering the training and reasoning fields involving cloud, edge and endpoint. We can do this because we have been involved in all these markets and have all relevant products."

She frankly said that AMD still "needs to do a lot of work" to catch up in the AI field, but she said that this market is "the biggest growth opportunity" for AMD.

Continue the advantages of CPU in AI reasoning field

Amd executives said at this event that they have begun to see the introduction of AMD products in the AI computing market, epyc server chips for reasoning applications, and the insight data center GPU for AI model training.

Dan McNamara, head of AMD epyc business, said that many cloud service providers have used AMD's software optimization function through AMD zendnn (Zen deep neural network) library to achieve "very excellent performance improvement" on the recommendation engine using AMD epyc CPU.

Zendnn integrates the mainstream tensorflow, pytorch framework and onnxrt, and is supported by the second and third generation epyc chips.

"I think it is very important to run most of the reasoning in the CPU, and we expect this trend to continue in the future," he said

Amd hopes to introduce more AI functions into the CPU at the hardware level in the near future, including the avx-512 vnni instruction. Later this year, AMD will introduce this instruction into the next generation epyc chip (code named Genoa) to accelerate neural network processing.

Since this function is implemented in the Genoa Zen 4 architecture, vnni will also appear in the AMD ryzen 7000 desktop chip, which will be available at the end of this year.

In addition, amd plans to use the AI engine technology of Xilinx, the FPGA design company acquired by US $49billion earlier this year, to expand the AI functions of AMD CPU in the future.