Microprocessor Report, a well-reputed chip-industry bulletin, reports that cloud computing workers like Amazon and few electronics titans like Huawei display remarkable outcomes compared to the graphics processing units and CPUs, parts that are ruling AI in the cloud.
David A. Patterson and John L. Hennessy, two legends of chip design, tells in an article published in the Communications of the ACM this month that generally, machine learning circuits denote roughly of a revolution in chip design. Last year, Hennessy and Patterson got the esteemed ACM’s A.M. Turing award for their years of effort on chip architecture design.
Linley Gwennap, the newsletter’s principal analyst in the Microprocessor Report editorial, calls the growth of custom application-specific ICs for the cloud with the idiom “when it rains, it pours.” Among the gush of chips is “Graviton” chip of Amazon, which can now be accessed in its AWS cloud service. One more is the “Kunpeng 920” chip of Huawei. Huawei wishes to utilize the chips both for its server computers and as a present for its cloud computing service.
For both Huawei and Amazon, the problem is the latch on inference by Xeon processor of Intel, and the latch on Nvidia’s GPUs’ cloud-based training.
“Cloud-service providers are concerned about Intel’s near-100% share of server processors and Nvidia’s dominance in AI accelerators,” Gwennap wrote. “ASICs offer a hedge against price increases or a product stumble from either vendor.”
Although it’s not easy for the ASICs to meet Xeon performance of Intel, “The strong performance of the Ascend 910 shows Nvidia is more vulnerable,” he wrote.
The article by Patterson and Hennessy takes a large view. The chip’s industry issue, they write, is the breakdown of the famous law of transistor scaling – Moore’s Law, and the breakdown of Dennard Scaling, which describes that in general chips become more energy-efficient. Simultaneously, a rule of thumb, Amdahl’s Law, says that the traffic jam in the performance of a processor is the number of sequential operations that must be measured, rather than parallel ones. All this implies that chips are roughly a calamity, however one that also offers opportunity.
They argued that mainly chip design has to transfer to specialization from general-purpose parts. The failure of Dennard Scaling and Moore’s Law “make it highly unlikely, in our view, that processor architects and designers can sustain significant rates of performance improvements in general-purpose processors.”
As an alternative, they perceive a nonstop step to “domain-specific architectures,” of which AI chips are the main case. The DSP chips can utilize many tricks that don’t operate for general-purpose processors like a compiler method for code called “very-long instruction-word,” or VLIW.
“VLIW processors are a poor match for general-purpose code15 but for limited domains can be much more efficient, since the control mechanisms are simpler,” they mentioned.
The authors forecast that DSAs will not only aid AI well, but they may be superior at securing code than general-purpose processors, escaping the fresh chip activities like Spectre and Meltdown.