I’ve had negative views of NPUs for over a decade and I can’t shuffle them off. A decade ago when I was supposed to recommend an NPU vendor to acquire I couldn’t do it. I put my vote on QED which was providing general purpose RM5k and RM7k MIPS CPUs. I posted negatively on NPUs again just last week. This viewpoint kinda bugs me because “programmability of hardware” is a key seminconductor feature that system vendors are looking for. Why can’t NPUs cut it?
The key issue for NPU vendors is “positioning”. Where do they win? Unfortunately they are the first choice only in the worst segment.
- low-performance & lower-volume applications can deploy “standard embedded CPUs” that FreeScale, Intel, ARM, and MIPS licensees make for higher volume markets. NPUs are too expensive for this segment, don’t have the breadth of product, and don’t have the software base. There is no biz case for semi’s directed at these segments because they can be serviced by products developed for high-volume markets with a little “tweaking”.
- low-performance & high-volume applications can deploy “targetted SoCs”. Consumer markets drive SoCs for cellphones, PDAs, home routers, iPhone, etc. These are usually based on ARM CPUs, higher performance applications tend to use MIPS. In reality these could be called “NPUs”, but they are not because they used “standard” processors. The lower volume segments simply leverage the solutions for higher volumes applications.
- High Performance & Low Volume applications — This is the land of the NPU. Standard embedded CPUs can’t cut it and the market isn’t big enough to justify an ASIC/SoC. The NPU vendor hopes that there are enough of these “smaller” market segments to justify the chip development costs. But there is often a large software component required for each targetted segment which increases the development cost.
- High Volume & High-Performance applications – This is the land of the ASIC and today these ASICS can deploy CPU cores for flexibility as well. They’re just big SoC’s. This is a segment that NPU vendors are hoping to get, but the higher-volume applications are serviced by “targetted SoCs” like L2/L2+ Ethernet Silicon provided by Broadcom (BRCM), storage processors provided by PMCS-Sierra (PMCS), or security silicon provided by Cavium (CAVM)
The bottom line — if there is a known volume market then there will be a “targetted SoC” developed for that space. The lower volume markets will leverage these high-volume solutions as much as possible. If they can’t deploy those solutions then they move to NPU’s 😦 This what I described before in CommSemi Nightmares.
There are those that hope that biz cases of these targetted SoCs will begin to look bad and make NPUs look better. But Marc Andreesen’s great post on “VC cash as an Asset class” indicates that there will be loads of cash for startups to develop these SoCs.