Momentum builds for open-source processors (2001)

With silicon scaling laws getting walls, Intel's ability to have an efficient general purpose CPU for every workload is fading. The only way to provide better performance is to provide specialization, now.

Processing will specialize in the following:

* databases: SQL projections in special hardware data slicing CPU instructions. Hardware-accelerated hashing, sorting, filtering.

* dynamic languages: Special code caching for JITted code. Memory models that won't have to be emulated. Memory barriers and copy on write to help the GCs do the work.

* AI: We're already seeing this with Google's TPUs: Low precision tensor operations. We're going to see more specializations inside this segment too. Tensorflow today has CPU, GPU and TPU backends. In a stable high volume configuration such as vision for an autonomous car, why not provide a fourth backend for a Tensorflow pipeline: VHDL/Scala?

* IoT/Sensors: Extremely low power modes, on-chip LoRa.

* Network: this is already happening with microsoft using FPGAs to help do the routing. Once settled, some use cases will have their silicon edition.

----

This will lead to a thining of the margins of Intel, NVidia and ARM, to the fragmentation of the monopolies, and less money will go to the lawyers.

So let's say your team has what it takes to design and validate the HDL for a processor microarchitecture that implements some un-encoumered ISA with performance that is attractive and a sufficiently short errata list. Then published the HDL under some free-as-in-speech license. You still have to get gcc, llvm, and gdb to support you, at a minimum. And then you want the hardware debugger people to support you. They will support you if you get enough users, if you get enough support from tool makers you will get users. The trick is not running out of money before the ecosystem spins up and your competition releases a more attractive core.