from Tom's Hardware (tomshardware.com)
Following my last post on “Keep talking about Multicore” a few colleagues pointed out that discussing Many-Core is applicable too. I certainly don’t disagree since Many-Core processors are coming as TNBT (The Next Big Thing). I think we will see adoption into specialized markets sooner than others. In particular networking which is already comfortable with multicore will use the many core tiles to speed up specialized processing. I can see applications in other areas such as imaging and devices that need higher end processing of audio, video or graphics.
In general, as I stated in my previous post, the embedded market can be slow to adopt new technology in specific market verticals – not because they are old school but because the product timelines and planning are years in advance and product life cycles are much longer. The move to multicore is underway and many-core in the near-term will be more niche than commonplace. Why? I think because we are still coming to grips with the software complexity of multicore. Moreover, tools are still catching up to multicore, how will we debug a 100 core system? Many-core will likely be handled the way heterogenous multicore solutions are today (e.g. general purpose CPU plus DSP or GPU). Vendors might supply libraries that take advantage of many cores will leaving a small number for general purpose processing. I have heard customers say that, heterogenous systems are particulary hard to debug because they are usually supplied with two disparate tool chains. Interestingly, an article from 2007 was quite prescient about this very topic, supporting my premise that hardware has leapfrogged the software – of course, would I quote an article that didn’t?
In the near term, I think we will see 16 to 32 core chips which will be “quite-a-few-core” systems. In these cases, I think a hypervisor solution can bring sanity to the solution by allowing you to create several virtual targets in one. For example, you could create a four core controller plus 12 specialized processing engines (this is being done today for deep packet inspection). This is manageable because the application complexity is isolated to your four-core target and the specialized processing engines are identical and relatively simple.
Whatever way the many-core technology pans out, I still contend that software is behind the curve versus the hardware. And I still won’t stop talking about multicore…
