Aceshardware

On the eve of so many upcoming GPU and CPU announcements it is only appropriate to take a step back and re-evaluate the status quo with a critical eye.

The computer industry is on the cusp of yet another turn of Sutherland's Wheel of Reincarnation, with the graphics processor unit (GPU) cast as the heir apparent of the floating point co-processors of days long gone. Modern GPUs are ostensibly higher performance and more power efficient than CPUs for their target workload, and many companies and media outlets claim they are leaving CPUs in the dust. Is this really the case though?

This article explores the quantitative basis for these claims, with some surprising results. Once you consider the larger die area and power consumption for modern GPUs, the efficiency advantages aren't quite as tremendous as they seem at first glance.

http://www.realworldtech.com/page.cfm?A ... 0909050230


Enjoy,


David

One observation - the comparison as it stands is strongly affected by using double precision for the comparisons. A comparison using SP would yield far different results. GPU vendors have focused almost exclusively on SP so far and DP has really been an afterthought at this point that they all freely admit is very low performance in current implementations. It looks like they are all aiming for DP close to 0.5*SP performance, which will boost all the DP numbers dramatically and push all the GPUs far to the right of the charts. I guess AMD is already getting there with their latest parts.

I know you know this of course, and do mention it in the article, but I feel it does influence the conclusion quite heavily in a way that may not be quite fair if the takeaway is that relative performance is not that dissimilar. At this point it is more a measure of a mature vs an immature implementation than any real fundamental comparison between the two approaches.

Might also be nice to do the graphs as log/log plots to spread out the data more evenly.

Thanks for an interesting article David. Using the more appropriate Istanbul for comparison we see that even for SP the theoretical advantage of GPUs is less than x10. You forgot one very substantial factor in your list, namely that GPUs are massively parallel compared to CPUs. Another metric that is often important for massive float is memory bandwidth, and the difference there is still well below x10 I believe. Magny Cours with it's twelve cores and four DDR3 channels is probably not going to improve the metrics in GPU favour either. In sum this means that GPGPU is still a limited playing field. OpenCL might improve the situation by making the programming model much simpler, but I am still sceptic.

And using the more appropriate RV790 for comparison we see that for SP the theoretical advantage of GPUs is just above x10 ;)


Comparing RV790 to Istanbul it's about x10 too.


But RV870 will... And also will come earlier.

And your point is?

The article mentions AVX, but there is more coming like fused-multiply-add. I could not see in the article any mention of Larrabee either, which is going to consist of several in-order cores with a wide 512-bit vector unit.

SP performance is also useful to compare because it is used a lot in video programming.

Magny-Cours will actually be worse than Istanbul in FLOPS/mm², due to lower clock speeds. It should however be more power efficient compared to its highest-clocked 6-core counterparts.

In other words it will move up from Istanbul, but also left.

At the same time, RV870/Cypress should almost double FLOPS/mm² and FLOPS/W over RV770, perhaps a bit more if AMD has put more emphasis on DP, so far we only have SP figures.

http://www.brightsideofnews.com/news/20 ... gddr5.aspx

"GT300" is supposed to perform DP operations at half speed compared to SP operations. This would mean >700GFLOPS of scalar MAD DP performance.

More details here: http://www.nvidia.com/content/PDF/fermi ... epaper.pdf

Very good paper focusing on the performance/watt and performance/mm^2. It's very important to measure how the CPU and GPU relate to one another as we start moving into the realm of GPU and CPU homogeneity or heterogeneity.

What I think needs some attention, though, is the fact that different types of workloads would work differently on CPUs and GPUs. GPUs do excel in performance, but only for the embarrassingly parallel workloads and other things that graphics cards have always been good at. So in this comparison, all is fair if we take into account that the workloads are more "SIMD-like" and if we pay attention to the issues of DP and SP. As for branchy code containing plenty of control statements, I'd still consider the CPU superior.