- Saturday, 28 November 2009
- INQ Mobile
- RSS
Intel tries to backstop its own roadmaps - Bob Colwell, former Intel chief architect
Intel man lets rip on GPGPU concept
IN A BOG POST sure to get temperatures rising in Satan Clara, Anwar Ghuloum at Intel has ripped on the concept of general purpose GPU use and laid out its shortcomings for all to ogle at.
You can read it here.
"Like a shimmering, unreachable oasis, today's GPUs offer the promise of all the performance you require, but achieving that goal for all but a few applications... is elusive."
Fighting talk indeed. Ghuloum suggests that the three problems preventing adoption of GPGPU are the programming model, the system bottlenecks, and the architecture - with an added bonus point for the outrageous form factors and cooling solutions.
In terms of architecture, Anwar reckons that achieving anything like peak performance out of a GPU is impossible, since the CPU is required to coordinate much of the action and the CPU-GPU link is a relatively slow one. Programming is hard because the way graphics cards are designed makes them great for predictable data sets and behaviours, and less good for often-used visual computing applications. On the upside, programming seems to be getting easier thanks to the implemetation of middleware, but it's still far from close to optimal, he thinks.
What's the solution to the problem? Well, our man from Chipzilla seems to think that better hardware architectures are needed to keep pace with the demands of the markets requiring this kind of high performance processing. Are we about to see the 'Zilla move into this market with some kind of product besides a quad-core Xeon or Itanic? ยต
Share this:
ShareBusiness Intelligence Software
Customer Relationship Management (CRM)
Enterprise Accounting Software
Enterprise Resource Planning (ERP)
Enterprise Systems Management
Isn't that what larrabee is about??
CPU to GPGPU link is slow?
AMD plans to have a HTX socket for the GPGPUs connected with low-latency / high-bandwidth HyperTransport to the CPU(s). Lack of dedicated ultra-high speed memory for the GPGPU would be a bigger problem for achieving peak performance, but this is solvable too.

Programming model? There are many emerging solutions to that problem. They will mature in a few years, but even before that you have many applications which would benefit greatly. It is not the number of applications that matters, it is the overall impact. Think about:
1. Server side anti-virus software (mail / content filtering)
2. Server side encryption (secure web servers, file servers etc.)
3. Server side RDBMS - SQL / Oracle (search / indexing / re-indexing)
4. Server side search - Exchange Server etc.

The benefits for the end-users? Faster and better Speach Recognition / OCR, picture search, desktop publishing software acceleration etc.

This is just well forgotten history - think abou the FPU co-processors back in the i286 / i386 era? Who would need FPU? Mostly scientists...;-))))
But he is right, I as a GPGPU developer can confirm that.
Does Wily actually read any of the INQ articles, otherwise he would know about Intel Larrabee and its 16 core mini AI. Intel is suppose to demo Larrabee next year.

creating a GPU solely for ray tracting would be great, but lets not forget 3D applications,they runs better under OpenGL. Hopefully Intel has thought about that also.
So if larrabee is supposed to have really wide vectors doesn't that mean it will have the same difficult time the GPUs currently do with those pesky unpredictable branches and irregular memory accesses?

I'd take it that Mr. Ghuloum stands more in favor of gni than lni.