Article: Parallelism at HotPar 2010
By: Mark Christiansen (aliasundercover.delete@this.nospam.net), July 30, 2010 6:36 am
Room: Moderated Discussions
David Kanter (dkanter@realworldtech.com) on 7/29/10 wrote:
>I see the GPU as a relatively new platform, one that holds a good deal of promise
>for certain highly structured and HPC-like workloads that are free from dependencies.
>It's fundamentally different from a CPU in that it's really a bandwidth optimized
>device, and there are certain trade-offs that implies which make it unsuitable for many workloads/algorithms.
>
>Ultimately, the right balance is a combination of the CPU and GPU. ...
>
>David
This seems like sense to me.
But then I remember the parade of vector accelerators and similar coprocessor like things intended to give high performance for jobs with heavy computation needs and regular data. They all came with a nice performance advantage and they all faded away.
It seems to me they died of software. Save the tasks with the most present need and the readiest access to custom software the main stream processors beat them soon enough it just wasn't worth the development effort. Since they were all different software written for them soon lost its value.
Software on CPUs has to survive and be upgraded by multiple generations of new hardware. It isn't enough it go on working, it has to get more performance with the new. Processor compatibility from generation to generation is vital.
Will GPUs get this? Or will GPU software have a life span of 2 years at best?
It seems blindingly obvious a computation engine built for tasks like graphics with large data sets and less random control can out perform a CPU which is built to optimize control and flexibility. All that flexibility is vital to performance on most general purpose jobs but it costs dear in silicon and power. A steel I-beam is stronger than a robot arm.
How much performance can the GPU give while allowing software to go on working and go on gaining performance with new generations for 15 years? Meet that spec and give 5x performance for the same power on problems people care about and I predict GPUs move in next to the CPUs and find a lasting home.
Fail the software lifespan test and the wheel turns again with GPUs forgotton but maybe some new scheme attempted in a new decade.
>I see the GPU as a relatively new platform, one that holds a good deal of promise
>for certain highly structured and HPC-like workloads that are free from dependencies.
>It's fundamentally different from a CPU in that it's really a bandwidth optimized
>device, and there are certain trade-offs that implies which make it unsuitable for many workloads/algorithms.
>
>Ultimately, the right balance is a combination of the CPU and GPU. ...
>
>David
This seems like sense to me.
But then I remember the parade of vector accelerators and similar coprocessor like things intended to give high performance for jobs with heavy computation needs and regular data. They all came with a nice performance advantage and they all faded away.
It seems to me they died of software. Save the tasks with the most present need and the readiest access to custom software the main stream processors beat them soon enough it just wasn't worth the development effort. Since they were all different software written for them soon lost its value.
Software on CPUs has to survive and be upgraded by multiple generations of new hardware. It isn't enough it go on working, it has to get more performance with the new. Processor compatibility from generation to generation is vital.
Will GPUs get this? Or will GPU software have a life span of 2 years at best?
It seems blindingly obvious a computation engine built for tasks like graphics with large data sets and less random control can out perform a CPU which is built to optimize control and flexibility. All that flexibility is vital to performance on most general purpose jobs but it costs dear in silicon and power. A steel I-beam is stronger than a robot arm.
How much performance can the GPU give while allowing software to go on working and go on gaining performance with new generations for 15 years? Meet that spec and give 5x performance for the same power on problems people care about and I predict GPUs move in next to the CPUs and find a lasting home.
Fail the software lifespan test and the wheel turns again with GPUs forgotton but maybe some new scheme attempted in a new decade.