NVIDIA Maxwell Speculation Thread

[silent_guy]

I don't think that nowadays the main driving force behind CL is AMD (if they ever were).

Replace 'CL' with 'compute' then.

My point being: for AMD, GPU computing seems to be little more than a checkbox. They have almost no libraries. No obvious major investments in tools. In applications etc. Or if they are, they're definitely not trying to make it known.

I understand that they simply don't have the money to fund these kinds of things, and for a hobbyist tinkerer that doesn't matter much, but if you're doing any of this kind of stuff in a more serious capacity, proprietary vs open matters less than getting the job done.

You have 2 choices for GPU computing: NVidia with CUDA or AMD with OCL. Nvidia doesn't not to give OCL a lot of attention, so the openness of OCL doesn't buy you anything from a practical point of view. And thus it comes down to which solution has the most support.

 

[silent_guy]

Yes, that's possible, but hard to say without some quantitative data about the cost of fast DP, which I don't have. Still, it's the only explanation that makes sense to me.

The fun part is that AMD has the same constraints as well. But their starting point is an already a significantly larger die that doesn't perform as well. With HBM, I don't think they can justify not adding DP though. Really bummed that we'll have to wait another 4 months to see what they came up with.

 

[Jawed]

AMD can't afford to build compute GPUs and graphics GPUs like NVidia, so it will pay a far heavier price for DP, afflicting power/performance of graphics only work at the "ultra enthusiast" level. At least AMD's DP is worth having, whereas NVidia's isn't, for those who absolutely must have DP.

 

[3dilettante]

My point being: for AMD, GPU computing seems to be little more than a checkbox. They have almost no libraries. No obvious major investments in tools. In applications etc. Or if they are, they're definitely not trying to make it known.

For OpenCL, not exactly. The more pervasive compute investment AMD has is the currently provisional HSA. That is impactful enough that conformance to it has had real impact on the design of its chips.

At least as originally described, HSA would run below the OpenCL runtime, although as time goes on and the feature set supported by OpenCL expands, it is not clear if this lowers the ceiling that the secondary runtime must run under.
The expansion of Vulcan's representation to support compute and graphics in the same code might mean something for whatever is left for Mantle, since HSA may in theory hit a case where it won't fully support the runtime above it.

The fun part is that AMD has the same constraints as well. But their starting point is an already a significantly larger die that doesn't perform as well. With HBM, I don't think they can justify not adding DP though. Really bummed that we'll have to wait another 4 months to see what they came up with.

It might be longer. The big consumers for DP that like Hawaii's shift to 1:2 DP rate would be at the FirePro level, and those apps love their memory capacity.

 

[Alexko]

The fun part is that AMD has the same constraints as well. But their starting point is an already a significantly larger die that doesn't perform as well. With HBM, I don't think they can justify not adding DP though. Really bummed that we'll have to wait another 4 months to see what they came up with.

If HBM1.0 is really restricted to 1GB per stack, and AMD does indeed use it for Fiji, then it will likely be limited to 4GB or perhaps 6GB—at most 8GB. In that case, making a Fiji-based FirePro at all might not be viable, which would make fast DP pointless.

Why do you think we'll have to wait another 4 months?

 

[Pressure]

Pretty sure Apple are the ones spearheading OpenCL these days.

NVIDIA barely supports OpenCL 1.2, which also means the only Mac computer with an NVIDIA GPU today is the MacBook Pro with retina display.

 

[Jawed]

If AMD has architected its HBM properly, then in 12 months or so after the first chip arrives with it, it'll be possible to interface that chip to second generation HBM with higher capacity for the FirePro card (since it seems to take AMD about a year to get round to the FirePro variant).

I'm still not convinced the chip arriving this spring is HBM equipped, although Alexko made a fairly persuasive argument for it.

 

[lanek]

If HBM1.0 is really restricted to 1GB per stack, and AMD does indeed use it for Fiji, then it will likely be limited to 4GB or perhaps 6GB—at most 8GB. In that case, making a Fiji-based FirePro at all might not be viable, which would make fast DP pointless.

Why do you think we'll have to wait another 4 months?

I dont know, HBM2.0 should not be far , but when you see that 4GB HBM have allready ~ 640 GB/s of bandwith dont you think 8GB HBM should be pretty enough for Firepro ?... Anyway we will not see the Firepro quickly on the market, AMD have still the upper hand with his actual lineup.

Im pretty sure in the worst case, AMD could segment the Firepro for workstation / raytracing ( without releasing a server / HPC version ) first, with high TFLOPS SP/ FP64 but with less memory, keeping the 16GB Firepro on the on the server HPC one for the moment beeing.
Professionnal hardware timeline and renewal is not the same than gaming product . ( Professional, at contrario of gamers, and if there's no special contract will not run to the shop the next day a gpu is released )

But you have a pretty good point.

 

[AlexV]

Pretty sure Apple are the ones spearheading OpenCL these days.

That's not a great certainty to have as it's utterly incorrect.

 

[silent_guy]

If HBM1.0 is really restricted to 1GB per stack, and AMD does indeed use it for Fiji, then it will likely be limited to 4GB or perhaps 6GB—at most 8GB. In that case, making a Fiji-based FirePro at all might not be viable, which would make fast DP pointless.

Not totally pointless: there must be a decent market for fast DP with 'only' 4 or 8GB. After all, that's what most solutions are right now.

Why do you think we'll have to wait another 4 months?

If we are to believe that kitguru article...

 

[Pressure]

That's not a great certainty to have as it's utterly incorrect.

What other OS uses it as extensively as OS X?

 

[silent_guy]

If AMD has architected its HBM properly, then in 12 months or so after the first chip arrives with it, it'll be possible to interface that chip to second generation HBM with higher capacity for the FirePro card (since it seems to take AMD about a year to get round to the FirePro variant).

Probably. It'd be an immoral waste of BW though!

 

[silent_guy]

What other OS uses it as extensively as OS X?

I can't find any references of OCL 2.0 availability for OSX, but if it doesn't, then Apple is definitely not spearheading.

 

[Pressure]

I can't find any references of OCL 2.0 availability for OSX, but if it doesn't, then Apple is definitely not spearheading.

Perhaps my choice of word comes back and haunt me but OS X and iOS use OpenCL extensively for everyday stuff, albeit only OpenCL 1.2 as of now.

 

[Alexko]

Not totally pointless: there must be a decent market for fast DP with 'only' 4 or 8GB. After all, that's what most solutions are right now.


If we are to believe that kitguru article...

Kitguru talked of a Computex lauch, and Computex is less than 3 months away (which still seems too late to me).

 

[lanek]

Kitguru talked of a Computex lauch, and Computex is less than 3 months away (which still seems too late to me).

Does really AMD will make their launch at Computex or during the computex ? they allways do their own launch session, then the words was the familly, does partners want their cards at the Computex ? yes, thats for sure ( and im pretty sure the rumor is coming from them. )..

I dont remember the last time Nvidia and AMD have release gpu's in a so close timeline, the marketing teams will have some good fun. I allready buy the popcorn.

 

[RecessionCone]

Perhaps my choice of word comes back and haunt me but OS X and iOS use OpenCL extensively for everyday stuff, albeit only OpenCL 1.2 as of now.

Pretty sure iOS does not use OpenCL at all.

 

[sebbbi]

Pretty sure iOS does not use OpenCL at all.

iOS does not have OpenGL ES 3.1 either. Metal is the only way to write compute shaders on iOS.

 

[silent_guy]

iOS does not have OpenGL ES 3.1 either. Metal is the only way to write compute shaders on iOS.

Anyone wants to set up a betting pool about a Metal for OSX announcement at the next WWDC?

 

[lanek]

Anyone wants to set up a betting pool about a Metal for OSX announcement at the next WWDC?

Im pretty sure it will... But what will be the point, at least for gaming and softwares who are not available only on OSX ? ... Dont you think it will be better for a developper to develop / port a game to "OpenGL Vulkan", who will work with the hardware used by Apple ( AMD or Nvidia ).. than again port it to Apple Metal ? ... ( i mostly imagine a case where game is developped with DX12, but ported to Vulkan then, and then will need again ported to Metal OSX.. )

As for iOS games, thats another story ( why not seen compatibility between iOS games and OSX ? )