Yes, we know that the title of this week’s post sounds a lot like two previous posts. We wanted to link together the two threads from those posts into a single message that we believe reflects what is happening right now in the world of complex chips. This is a short summary in line with the short week due to the Thanksgiving holiday here in the United States. The line of argument is straightforward:
- Large chips are adding embedded processors to implement complex functionality while retaining flexibility
- Single-processor chips are adding multiprocessor clusters to get better performance at a given process node
- Multiprocessor chips are using shared memory for effective data transfer and interprocess communication
- Neighbor-connected processor arrays are moving to shared memory to reduce cross-chip data latency
- Multiprocessor designs are adding caches to reduce memory access time and bypass memory bottlenecks
- Multiprocessors with caches require coherency in order to ensure that the right data is always accessed
While most of these statements are not universally true, they reflect a significant sea change that we see every day when discussing current and future projects with our customers.