The Silicon Renaissance

“A lot of people much smarter than I am say that the entire data centre is a computer,” Victor Peng CEO of Xilinx told a Silicon Valley think forum recently, “there’s going to be computing in the CPU, in the storage, memory in flight in the network adapter, even in a switch.”

That’s why, thinks Peng, there’s going to be a Silicon Renaissance.

“I think a renaissance is going to be driven by the fact that there will be new architectures,” explained Peng, “and when I say architectures I mean the technology as a whole: at the technology level in transistor, device, packaging, advanced packaging, integration, micro-architecture and instruction set architecture and even architecture at the scale of entire data centres.”

Peng, a designer of CPUs, GPUs and ASICs in stints at AMD, ATI, MIPS and DEC, devised Xilinx’s ACAP architecture for dynamically adapting a server to a workload. His starting point was that GPUs and ASICs are no longer enough.

“GPUs are OK for heavy lifting because they don’t care about latency – that’s why they can sell top-end GPUs using 300W,” he says “but burning several hundred Watts is not a good thing for a datacentre. If  you go to ASIC, it’s even narrower than GPU, so you’re not going to have a broad solution.”

 “There are other ways you can have the same impact of what would be a customized piece of silicon now optimised for specific algorithms: the thing with ASIC is you’re you’re literally solving a very narrow band of a problem,” said Peng, “what you need for an ASIC or fixed function chip, is some degree of stasis. The problem you are trying to solve has to be stable enough; if you are interfacing with standards those standards have to be stable – not least because it takes a while to implement and execute. You also need quite a bit of capital and really expert people in silicon level validation software and two to three years, end-to-end,” said Peng.

 “You don’t really want to do a lot of specialised processors,” said Peng, “that may have been the state of the art one point in time, but it’s an exponentially changing world. What you really want is a different level of virtualisation at the hardware level and you want to be able to compute wherever it makes sense: near the storage, near the network, or near the processor.”