NewsTechnology and InnovationThe Macroscope

Chiplets and the Extension of Moore’s Law

What happened?

AMD has announced its second generation EPYC server processor with 64 Zen2 cores, code-named Rome, which is expected to launch in 2019. Similar to Apple’s A12 chip, which was shipped with the latest iPhone, AMD’s processor makes use of TSMC’s latest 7nm manufacturing technology. More interestingly, the processor applies a so-called chiplet design approach, in which computer chips can be designed in a more modular way resulting in more customization and a better cost/performance tradeoff. In the case of the EPYC Rome 8 chiplets were used, whereas with last year’s version (Epyc Naples) AMD used 4 chiplets.

What does this mean?

As previously discussed, the International Technology Roadmap for Semiconductors (ITRS) expects Moore’s law to become obsolete by 2022. By that time, chips will have reached a node size of 5nm, after which further miniaturization will result in overheating and malfunction due to quantum effects. As expected, we can already see that the semiconductor industry is in addition to miniaturization looking for new ways to increase the performance of processors. According to a few outlets (Wired, Anandtech) it is expected that the chiplet approach will become an important part in future chip design. The importance of the chiplet approach is underlined by its inclusion in DARPA’s Electronics Resurgence Initiative (ERI), a 5-year, $1.5 billion program to advance computing performance. Moreover, Intel is aiming to drive chiplet development through their standard ‘Advanced Interface Bus (AIB)’.

What’s next?

In the first place, the use of the chiplet approach could open up new possibilities for the semiconductor industry to meet the demand for ever-increasing processing performance and specialized computing tasks (e.g. AI, real-time graphics rendering, simulations). Some speculate that these chiplets will be stacked in 2.5D and 3D packages, thus unlocking further performance increases. Furthermore, it could open up the possibility for multiple chipmakers to collaborate on a solution as exemplified by Intel/AMD’s mobile solution. Lastly, chiplets also seem to enable chipmakers to manufacture faster and at lower costs.