NVIDIA Maxwell Speculation Thread

[A1xLLcqAgt0qc2RyMz0y]

And typically when Nvidia has nothing much to brag about in the present, they talk up their future stuff even more than usual.

A mid-range Kepler (GTX680) more than put the hurt on the high end 7970 Tahiti.

If AMD had actually produced a real high-end GPU the Big Kepler (GK110) would have been brought forward to battle it but because AMD dropped the ball all that was needed was a mid-range GPU from Nvidia and then AMD had to castrate prices/overclock/and bundle games just to compete with it.

So if Maxwell crushes Kepler in energy efficiency AMD will once again be trailing in the GPU race.

 

[A1xLLcqAgt0qc2RyMz0y]

More than usual?
They have been giving presentations on their future projects for years, including Maxwell and beyond.

Ever hear of road maps?

I believe even Intel and AMD have those.

 

[rpg.314]

So if Maxwell crushes Kepler in energy efficiency AMD will once again be trailing in the GPU race.

Why? What will prevent AMD from matching Maxwell?

 

[jlippo]

Ever hear of road maps?

I believe even Intel and AMD have those.

I might have.

In addition to the classic 'we have invented names for our future hardware', they have shown presentations on their future looking projects. (some true GPUs some internal projects.)
Including why they will have L0 caches and so on.

I would love to see information on future hardware of AMD and Intel as well.

 

[Ailuros]

Why? What will prevent AMD from matching Maxwell?

The GPU won't be named Nescafe

 

[tviceman]

Unified virtual memory support and the integrated ARM cores both sound very interesting. Could this be a potential answer to the new consoles HSA abilities?

I am guessing only the big die maxwell chip will have arm integrated. I don't think the need yet justifies adding the down space to their <=300mm^2 chips yet. I could be wrong though.

 

[lanek]

The GPU won't be named Nescafe

I was ask me if im the only one who heard the song of the publicity of Maxwell Coffee each time i read a news about Nvidia Maxwell .... ( "Maxwell qualité filtres" )..

 

[Blazkowicz]

Totally forgot about Maxwell dehydrated café.. that song was at least 20 years ago.

I instead really thought about electromagnetism and Maxwell equations.

 

[keldor314]

The roadmap claims that Maxwell will have "virtual unified memory". Obviously them mean something more than zero-copy, which has been around since at least gt200. Could this end up being similar to virtual memory on the CPU, where you have a page fault which causes that memory page to be transferred over to the device? This would be VERY useful for complex structures, since you'd just pass a pointer and let the structure get copied over on demand.

 

[lanek]

The roadmap claims that Maxwell will have "virtual unified memory". Obviously them mean something more than zero-copy, which has been around since at least gt200. Could this end up being similar to virtual memory on the CPU, where you have a page fault which causes that memory page to be transferred over to the device? This would be VERY useful for complex structures, since you'd just pass a pointer and let the structure get copied over on demand.

Its the exact same features of AMD.. virtual unified space memory ( the name have a bit change now ) .. as it is too developped as a features for ARM by HSA fundation for permit a better integration of ARM system inside an x86 system and vice versa . i think you can find maybe more information about it by looking on AMD GCN roadmap. In reality it have been half integrate ( for PRT ) , but it was miss memory coherency. This will enter in function on both AMD and Nvidia next generation ( Maxwell for Nvidia as they need it with ARM )

 

[ninelven]

What a great discussion on Maxwell...

 

[AlexV]

It is, isn't it? Now, I've cleared most of the...ugh...something about SKUs and whose is bigger into a separate thread: http://forum.beyond3d.com/showthread.php?t=63404. Please keep it there (sorry Mintmaster, your post was on-topicish but also had the interesting property of making for a reasonable opener, so it took the mantle). I've moved some of the (neverending) armchair analysis of AMD's corporate fate into the gloom thread (please keep it there). An architecture speculation thread is intended for speculating about the architecture, and just that.

 

[Gipsel]

The roadmap claims that Maxwell will have "virtual unified memory". Obviously them mean something more than zero-copy, which has been around since at least gt200.

Zero-copy works with unified memory for CPU and GPU as one has with APUs. What you think of is pinned memory (which exists even longer iirc).

Could this end up being similar to virtual memory on the CPU, where you have a page fault which causes that memory page to be transferred over to the device? This would be VERY useful for complex structures, since you'd just pass a pointer and let the structure get copied over on demand.

That's a feature AMD proposed for its future architectures I think (and the groundwork of it is obviously already in GCN). Doing it with a unified virtual memory system (different from virtual unified memory?) without driver intervention would necessitate that the GPU also plugs into the page table system on the CPU side so that the access works always and not only if one has pinned a memory region for direct access by the GPU.

 

[Blazkowicz]

With Maxwell things can get interesting, or complicated. If you have ARM CPU cores on the GPU die.. Then you have to be careful about what you are talking about exactly. Instead of main x86 CPU(s) + GPU, you have x86 CPU, ARM CPUs and the GPU, hopefully all sharing a single address space but maybe only the ARM CPU has the tightest integration (and operates on the same physical memory as the GPU does)

That's quite heterogeneous, isn't it
Then if you have Intel CPU, Intel GPU, nvidia CPU, nvidia GPU or AMD CPU, AMD GPU, nvidia CPU, nvidia GPU : how does stuff work together?
It probably doesn't but I hope you can have two integrated solutions not stepping on each other's foot.

 

[Ailuros]

Honest question: will Maxwell actually have Denver CPU cores integrated after all?

 

[LiXiangyang]

I think Maxwell's ARM is just for thread controls, basically a more advanced type of giga thread engine or whatever name nvidia called it.

 

[Blazkowicz]

Well they are general purpose CPUs, and they feel like they would be the main CPU in a Tegra. Let's say the biggest Maxwell has 20 SMXX (see what I did there) and 8 cores, a Tegra could have 4+1 or 2+1 cores, and 1 SMXX.

But certainly, if you're only running some generic DirectX or OpenGL code or other, and not expressly running ARM CPU code, then nvidia might still be doing stuff with them through the driver.

Maxwell GPU without CPU cores? We don't know but I agree it's entirely possible.
I've maybe had a false sense that every Maxwell has Denver CPUs, since the Project Denver announcement two years ago.
We can imagine a full line of Maxwell GPUs get launched, then a big Maxwell + Denver chip for Tesla (and 'self-sufficient' Tesla) a bit later, similarly to GK10x then GK110.

 

[tviceman]

I'm thinking GM104, GM106, and GM107 won't have ARM cores and will all come 9 months or so before GM100/GM110 (whatever the designation is for the big GPU) and that GPU will have the ARM cores. I don't think it will be Denver cores though. I think Tegra 6 will be the first chip to feature Denver ARM cores from Nvidia.

 

[spworley]

Honest question: will Maxwell actually have Denver CPU cores integrated after all?

I expected it, but now I think ARM integration is an option. We may see Maxwell dies both with and without ARM.

I base this entirely on a single slide from Jen-hsun's presentation at this year's GTC.
The Parker mobile set (Tegra 6?) will have a Maxwell GPU and "Denver CPU".
If Maxwell inseperably includes the Denver ARM CPU, why would they split out the description of Parker that way?

 

[Ailuros]

I expected it, but now I think ARM integration is an option. We may see Maxwell dies both with and without ARM.

I base this entirely on a single slide from Jen-hsun's presentation at this year's GTC.
The Parker mobile set (Tegra 6?) will have a Maxwell GPU and "Denver CPU".
If Maxwell inseperably includes the Denver ARM CPU, why would they split out the description of Parker that way?

It's exactly that slide that leads me to ask my former question. Since Maxwell is supposed to launch for desktop in 2014, why would Logan skip Denver CPUs and the latter appear only Tegra SoCs in 2015? Isn't TSMC going for FINFET only at 16nm?

So far it was my understanding that Project Denver CPUs will appear under TSMC 20nm. Given the slide above it sounds suspiciously more and more that there might not be any Denver cores at all in Maxwell and the entire Denver introduction has been postponed by one more generation.