There’s more to the rise in HPC and AI of domain-specific architectures (DSAs), starting with GPUs, than just the superior throughput they can deliver. Yes, the deceleration of Moore’s Law and Dennard scaling are the primary reasons for the CPU’s decline in dominance. But other factors are playing a strong, enabling role for DSAs.
Consulting company McKinsey & Co. has issued a report, “Domain-Specific Architectures and the Future of Compute,” that delves into the trends and new technologies that have “increase(d) the incentive for architectural innovations….”
To begin, the McKinsey report states what has been widely observed. Regarding Moore’s Law (transistor density doubles every two years), “transistor scaling has meaningfully slowed in past years and is behind where Moore’s Law would have predicted by a factor of about 10.” As for Dennard scaling (power consumption remains constant as transistor density increases), it “is also failing, leading to an increasing need for complex cooling solutions in large data centers and other high-performance compute environments.”
Citing GPUs and tensor processing units (TPUs), the firm said workload-specific speedups of 15 to 50 times over CPUs are common.
“As we see DSAs expand to other application domains, we estimate that DSAs will account for roughly $90 billion in revenue (or about 10 to 15 percent of the global semiconductor market) by 2026—up from approximately $40 billion in 2022,” McKinsey stated, adding that $18 billion in venture capital has been invested in approximately 150 DSA startups over the past decade, a significant increase over the preceding 10 years.
McKinsey said there are five major factors enabling the rise of DSAs:
- DSA vendor access to the advanced chip manufacturing capabilities of foundries, “which can aggregate demand and achieve the efficiencies of scale needed to offset the escalating cost of producing modern semiconductors”;
- Partnerships between DSA vendors and cloud services providers with compute-as-a-service offerings, enabling DSA companies to sidestep development of their own go-to-market capabilities;
- Access to libraries of open-source and licensed IP (Arm, x86, RISC-V, for examples) to accelerate DSA designs;
- The advent of chiplets resulting from 2-D and 3-D semiconductor packaging for heterogenous integration, “allowing for connectivity with other compute, communication, memory and analog components with extremely high bandwidth and low latency”;
- The promise of next-generation technologies (quantum, photonics, neuromorphic) – “As these physical-layer solutions mature, they will open up new classes of DSAs,” McKinsey said.
Domain-specific architectures complement and substitute general-purpose compute by offering workload- and application-specific features. (source: McKinsey)
McKinsey also said semiconductor industry companies and end users should prepare for DSA-driven disruptions. The report offers advice for materials providers, front-end tool manufacturers foundries, chip design firms, EDA and hardware IP firms and compute consumers, including cloud services providers, enterprise customers and domain-specific OEMs.
“Moore’s Law has propelled the compute industry with incredible longevity, driving decades of performance improvements in general-purpose computing that largely negated the need for investments in workload specialization,” McKinsey concluded. “With transistor scaling slowing down, DSAs will increasingly gain a use-case-specific performance edge and drive significant disruptions for value chain participants and their customers.”
Good article