For the sake of GPGPU, is it time for an AGP-style interface again?

If the rumors are true, then Skylake will feature 512-bit SIMD units. At 14 nm, octa-cores will be mainstream, so we're looking at 2 TFLOPS worth of fabulous homogeneous computing power. Why would any developer invest their time into dealing with the problems of GPGPU computing instead with little chance of widespread success?.

Indeed, if PC CPU SIMD capability continues to balloon at this rate then there would be little need for lower latency GPGPU interfaces IMO.

The big reason for the focus on low latency GPU for most people at the moment is I'm sure the fact that the consoles are likely to make big use of it and the question in everyone's minds (and in fact something stated as a specific consoles advantage that PC's may be unable to match) is how will PC's keep up.

APU's are too slow and GPGPU latency is too high so that only leaves regular old CPU's to compete with console GPGPU in games. That may be a problem for quad Sandy's/Ivy's (depending on just how much CU power is dedicated to GPGPU) but Haswells and especially 8 core Haswells (or better) should have little trouble in this regard. Assuming of course the games can take advantage of AVX2.
 
...new console may really change the desktop field.

In the end, all console games will make heavy use of compute shaders, given i.e. onion+.
And PC conversion may require as little as some shader 'tuneup'.
Yet, you cannot run fusion GPU shaders in dGPU due to PCIE lantency/bandwidth.

..so, as a paradox, I wonder if with new console games an APU could not offer an unexpected performance increase, given the rendering piepline running shaders in the discrete card and the GPGPU in the main processor.

...i am very curious to see what will happen and how game ports will be done.

By the way - APU are not slow: just let your compute shaders run in it... and you have enough speed for any thing.
It will be interesting, however, to see if AMD drivers will support this dual approach or not.
 
the way - APU are not slow: just let your compute shaders run in it... and you have enough speed for any thing.
It will be interesting, however, to see if AMD drivers will support this dual approach or not.

I just meant too slow on their own to deliver console parity. Whether the IGP in an APU could be used for GPGPU leaving the discrete card for graphics is another question. I'd certainly like to see it happen but from what I've heard on these forums there are quite a few barriers.
 
Back
Top