NVIDIA Maxwell Speculation Thread

[raghunathan]

So you agree with Alex, but he's "silly" for it. Good to know.

for all his complaining AlexV would do the same if he were in Roy Taylor's position.

 

[LiXiangyang]

Coming from some guy who currently have a HPC project running:

CUDA is certainly far and away from being dead, actually I think Open CL is the one to be replaced by MS's C++ AMP or even Open ACC in the future.

The problem with Open CL is, no major players seriously support it, in the world of heterogeneous computing, there are only two players, intel and nvidia, Intel's MIC use Open MP/TBB etc, Nvidia support CUDA.

Open CL can work on many hardwares but good at none, in the world of HPC, where the measure of your program is the achieved gflops/peak gflops of the hardware, to squeeze the performance of the hardware, you tend to use the hardware provider's recommended programming solutions.

As for consumer-level application, which is largely the world of windows-based platform, and which is not that sensitive to performance, then MS's C++ AMP or even Open ACC is likely to replace Open CL, since the former(s) are far easier to pick up than the latter for an average programmer with no experience in heterogeneous computing.

 

[rapso]

as always, if you cannot beat it, declare it will die.

I write programs for cuda and opencl, and both are in the majority of cases so alike, you don't really care what API/Language you use. you get your ptx assembly from opencl just like from the offline compiler of cuda.

on the other side, if you want to get performance, you sadly need to work hard for it. the API is unified, but the hardware is very different. When I've started with opencl, I've run it on an 8800GT, q9550 and the cell. getting a simple Julia fractal to work AT ALL was a day of work per platform (especially cell) and optimizing them was several days per platform.

on the other side, switching between cuda and opencl is mostly a compile flag for me. (sure, it's not always that simple, but 99% is that simple). if that AMD guy says Cuda will die, then so will OpenCL.

btw. OpenMP 4 shall add support to run your code on GPUs as well, smart move imho.
http://software.intel.com/en-us/blo...ant-solutions-for-targeting-and-vectorization

 

[Pete]

for all his complaining AlexV would do the same if he were in Roy Taylor's position.

The point being he doesn't think it's worthy of extended consideration in B3D. The predicate to that thought is an understanding of how PR works. This makes your ad-hom pedantic, thus doubly unnecessary. And debating how to keep Roy employed is preeetty tangential to this thread.

 

[boxleitnerb]

Apparently, Maxwell will come this year and still be 28nm:
http://www.xtremesystems.org/forums...ands-details&p=5203221&viewfull=1#post5203221

This guy had the correct info about the 220W FX and early info about Hawaii. So he could know something.

 

[ninelven]

There is a difference between Maxwell, and Maxwell based.

Is it possible we will see a consumer oriented Maxwell derivative on 28nm this year? Sure.

But the halo chip with 8 Denver cores? I seriously doubt it.

 

[boxleitnerb]

I doubt that too. But then Nvidia never said that every Maxwell-solution would have those ARM cores.

 

[ninelven]

No, but when people refer to the "xyz architecture" chip I generally infer that to mean the flagship. If it doesn't, then that should be made clear.

 

[Alexko]

NVIDIA's roadmap is fairly clear about Maxwell being a 2014 thing.

 

[boxleitnerb]

The professional versions, yes. DP GFLOPs were not exactly GK104-107's strong suit. But the first Maxwell GPU could be a January 2014 product, then it would be quite early and 2014.

 

[rapso]

isn't the address space already unified since fermi? (at least on 64bit) or is there some part missing (or emulated by the driver?).


@arm cores
maybe those lower end maxwell will also have less arm cores, but still be functionally compatible. would be useful if they also release cuda to use it. splitting the developer space isn't a smart move, imo.

 

[jlippo]

@arm cores
maybe those lower end maxwell will also have less arm cores, but still be functionally compatible. would be useful if they also release cuda to use it. splitting the developer space isn't a smart move, imo.

It certainly would be nice for them if they could use it trough driver on normal rendering tasks as well.

 

[rapso]

It certainly would be nice for them if they could use it trough driver on normal rendering tasks as well.

not sure if there is that much demand for 'normal rendering tasks' support. would be another man in the middle.

for compute on the other side, it would be quite nice to run your stuff purely on GPU. but I guess that's already been hinted by Kayla

 

[jaredpace]

Will Nvidia use die stacking with Maxwell? erps Nevermind looks like that is saved for Volta

 

[A1xLLcqAgt0qc2RyMz0y]

Why ask a question with an answer you already knew?

Will Nvidia use die stacking with Maxwell? erps Nevermind looks like that is saved for Volta

I don't see an EDIT.

Is there some hidden meaning in your post?

 

[iMacmatician]

In my experience, the "Last edited by" below the post doesn't appear if the edit was made within a minute or so after the original post was posted. On manually writing "EDIT" in the post, maybe jaredpace chose not to do so.

 

[UniversalTruth]

In my experience, the "Last edited by" below the post doesn't appear if the edit was made within a minute or so after the original post was posted. On manually writing "EDIT" in the post, maybe jaredpace chose not to do so.

Actually, it is 5 minutes

 

[LordEC911]

In my experience, the "Last edited by" below the post doesn't appear if the edit was made within a minute or so after the original post was posted. On manually writing "EDIT" in the post, maybe jaredpace chose not to do so.

aka ninja edits.

 

[iMacmatician]

Actually, it is 5 minutes

Okay. Thanks for letting me know.

 

[willardjuice]

Let's try to get back to GPUs.