The us developer told us about the architecture of the processors.
In a time when machine learning and technology based on neural networks are actively developing, progressing and hardware for these areas. But each developer approaches the question differently: manufacturers of GPUs rely on massive parallelism initially, their development, companies such as Google are developing a relatively simple chips, all of which is subject to a single task as quickly as possible, but companies like Intel are adapting existing processor architecture. Recently, “blue” has revealed some secrets concerning the series heirs of the Knights Landing processor, Knights Mill.
Of course, the company explores other areas, for example, solutions based on Altera FPGA, as well as developing coprocessors Lake Crest Knights Crest, which most likely will resemble Google TPU. However, the heir series, the Knights Landing chip Knights Mill — get changed due to the popularity of machine learning technologies architecture. The basis remains the same, but each unit (VPU) will receive instead of the two “big” FPU (32/64 bit) smaller unit floating-point unit and four Vector Neural Network Instruction (VNNI). The latter would support the floating point single precision, and will have the ability to work with mixed integer formats (16 bit input, 32 output). A similar kernel is a part of NVIDIA V100, but they support more traditional formats FP32/64.
The end result can be predicted quite accurately: Knights Mill will double to give the Knights Landing operations in double precision, but twice to beat the single precision operations. And the appearance of blocks VNNI-supported integer formats may raise the productivity of Knights Mill four times to tasks related to deep machine learning. Thus, the processor will be more specialized, but in their fields of specialization it will surpass its predecessor quite significantly. Chips Knights Mill will not suffer from a lack of PSP, since it will get a 16 GB MCDRAM complementary 6-channel DDR4 controller. The total number of cores remains the same — 72, subject to 256 operations in tact in blocks VPU, Knights Mill will be able to demonstrate more than 27 trillion calculations per second for frequencies of around 1.5 GHz. This is less than 92 trillion operations Google TPU, but the Intel looks more versatile and balanced. It will appear on the market in the fourth quarter of this year.
© 2017 – 2019, paradox. All rights reserved.
