Driven by the fierce competition of red (AMD), green (NVIDIA) and blue (Intel), artificial intelligence and data center technology made great progress last year Taking amd as an example, the company proposed a business strategy to focus on the research and development of high-performance computing (HPC) In addition, compared with 2020, the company hopes to improve the energy efficiency of data centers by 30 times in the next five years.
(Figure via wccftech)
According to the status information disclosed this week, after recognizing the industry development and trend demand of the next generation data center in the next few years, amd is making active preparations for this.
The efficiency of the company's AI and HPC acceleration platforms, including epyc series processors and insict computing cards, has been nearly seven times higher than a few years ago.
In previous years, the company has launched three generations of Xiaolong server CPU and two generations of cDNA architecture acceleration card. Taking the 64c / 128t epyc 7003 processor + four-way insict mi250 accelerator card platform as an example, its fp16 performance is almost 14 times that of four-way mi50 platform.
The insict mi250 accelerator card contains 13312 cDNA 2.0 stream processors (SPS), with a GPU frequency of 1.0 ~ 1.7 GHz and a power consumption of 500W. On this basis, amd also hopes to achieve the "30x25" goal set in 2020.
Under the guidance of this program, AMD will strive to expand the basic performance of its data center equipment, including improving performance efficiency per watt and application optimization, focusing on the product stack, and reducing power consumption while improving performance.
Compared with the hardware baseline established in 2020, the progress in the above areas will bring amd closer to its goal. Mark papermaster, chief technology officer of the company, said:
Although much remains to be done to achieve the 30x25 vision, engineers are very happy with the progress of their work and are encouraged by all the results so far. I sincerely invite you to actively contact AMD to obtain the progress report lasting for several years.
Wccftech pointed out that AMD uses the epyc 7742 server platform as the benchmark of 30x25. The CPU has 64c / 128t, the frequency is between 2.25 ~ 3.40 GHz, supplemented by 256MB cache capacity, and the power consumption of TDP is 225W.
In addition, the four-way insict mi50 computing card adopts the fifth generation GCN architecture, with 3840 stream processing, GPU frequency of 1450 ~ 1725 MHz and power consumption of 300W.
Initialize with triangle data (4K Taking matrix dgemm) as an example, the computing power of each mi50 accelerator card on the supercomputing platform is 5.26 tflops, the performance of fp16 is 21.6 tflops, and the power consumption is 1582w.
In the near future, AMD is expected to upgrade the HPC platform to a 96c / 192T Xiaolong processor that supports ddr5 memory (lower power consumption than DDR4) and PCIe 5.0.
Finally, combined with various optimizations, the performance of cDNA software and hardware of AI and HPC applications will also be continuously improved.