Tesla unveiled its new dedicated D1 processor. The solution is designed for machine learning. The processor is manufactured at 7 nm and contains 50 billion transistors. The area is 645 mm 2 , which is quite a lot, although less than that of the same GPU Nvidia GA100. Tesla D1 is equipped with 354 learning nodes based on a 64-bit superscalar processor with four cores. The solution supports FP32, BFP16, CFP8, INT32, INT16, and INT8 instructions.
Single-precision computing performance (FP32) is 22.6 TFLOPS, and in the case of BF16 / CFP8 mode, it is 362 TFLOPS. This is achieved with a TDP of 400W. Because scalability is important to machine learning, Tesla has developed a custom interface with 10 TB / s bandwidth. The I / O ring contains 576 lanes, each of which has a bandwidth of 112 Gb / s.
As a result, Tesla can assemble D1 chips into special tiles with 25 processors each. And such tiles can also be interconnected. The company demonstrated such a tile with processors operating at 2 GHz. The performance of this solution was 9 PFLOPS (BF16 / CFP8).
The company also has plans to create a supercomputer based on D1 processors. The ExaPOD system will be based on 120 tiles with 3000 processors. The final performance will reach 1.1 ExaFLOPS (FP16 / CFP8). Once built, the system will be the most powerful supercomputer for AI training. Compared to current Tesla supercomputers based on Nvidia GPUs, such a system will offer four times the performance and 1.3 times the performance per watt with a fivefold reduction in footprint.
0 Comments