Interview
David Patterson: Computer Architecture and Data Storage | Lex Fridman Podcast #104
- Cryptocurrency development is expected to redefine the nature of money in the future, while error-corrected quantum computing is projected to become a reality by 2030, remaining confined to data centers for "big science" rather than consumer devices.
- A machine learning benchmark suite named MLPerf has begun development approximately two years ago to ensure fair performance measurement across hardware vendors.
- RISC-V is predicted to potentially become the most popular instruction set architecture due to its open nature and lack of licensing fees, though it faces a decade-long transition challenge in the cell phone market against ARM's dominance and binary compatibility requirements.
- Performance acceleration over the next decade is expected to occur primarily at the hardware, compiler, and domain-specific language levels, with the "golden age" of computer architecture returning as software design shifts from explicit code to data and hyperparameters ("software 2.0").
- As Moore's Law no longer delivers performance doubling every 18 months, the industry is moving toward domain-specific accelerators and specialized hardware to support machine learning, which is now the dominant approach for creating artificial intelligence.
- General-purpose microprocessors are forecast to improve by only a few percent annually, whereas specialized hardware accelerators may deliver significant gains if software embraces machine learning, contrasting with the slowing of traditional transistor density gains.
- Flash storage is expected to continue replacing magnetic hard disk drives in personal computers and embedded devices, while magnetic disks will likely remain the more economical choice for large-scale cloud storage.
- The transition to open-source hardware and instruction sets may enable individuals and organizations to design custom processors, potentially accelerating innovation in specialized domains like biological computing using simple core instruction sets.
- Strategic shifts in the industry include Intel canceling its Nirvana product line and acquiring Habana to integrate hardware and software stacks for MLPerf compliance, reflecting a risk for companies focusing solely on hardware without supporting software ecosystems.
- Risks include the possibility that the industry may fail to meet growing computing demand from AI and data if software writing or hardware design does not fundamentally adapt to the slowing of Moore's Law, and the trend of dropping individual transistor costs may eventually reverse due to increasing technology complexity.