Nvidia BigK GK110 Kepler Speculation Thread

hmm? autocad uses GPGPU? I don't remember that - and even if it'd use it, you'd be hitting a wall since autocad uses all FP64 geometry...


Some parts can be accelerated by GPGPU computing for the 3D render for 3Dmax ( Iray with CUDA for the example )... I dont know why i write AUtocad, when i was think to Autodesk. ( plugins exist too, as for physic, OpenCL library exist for Maya etc ).
 
So what's the configuration then?
It uses "192-core SMX" but we know GK104 SMX was 192 FP32 cores and 8 FP64 cores, that's definately not enough FP64 for GK110
 
I'm watching the nVidia presentation.

Jeez, Dave Perry is one of those guys who talks like he finishes every sentence with a question mark.. how irritating!

Is it my impression or is this presentation more about services that can use cloud computing rather than Kepler itself?

What's this thing about virtual gpu? How is it better than a remote desktop?
 
I think the 192-core SMX was just in reference to GK104. GK110 will probably be 1/2 rate DP like Fermi. Anyway, it appears to be a huge die with many "cores."
 
6 prim's per clock? :oops:

i see 5 GPCs, upper mid one(wonder what is it? fully dedicated 1:1 DP units? as a plus other CCs run half speed or 1/4) is looking totally different.. it's in asymetric pattern kinda like GT200 unlike Fermi and GK104.. i try to figure die size by pcie area and i found 509mm2 but it's off as what JHH says biggest gpu we've ever built
 
i see 5 GPCs, upper mid one(wonder what is it? fully dedicated 1:1 DP units? as a plus other CCs run half speed or 1/4) is looking totally different.. it's in asymetric pattern kinda like GT200 unlike Fermi and GK104.. i try to figure die size by pcie area and i found 509mm2 but it's off as what JHH says biggest gpu we've ever built
But still 6 setup pipes means 48 fragments scan-out capacity, that nicely matches the ROP throughput.

Anyway:

mVFb6.jpg


:LOL:
 
So we're looking at 2880SPs, probably at 700MHz or so; perhaps more on GeForces—assuming they make GeForces with it.
 
Are current virtual machines servers like Vmware able to use GPUs for 3D rendering? Not talking about stuff like Parallels or Vmware Fusion, but the big iron stuff?
 
Sounds like BigK has the ability to launch kernels from within kernels (the HyperQ stuff mentioned in the keynote). Any more info on that?
 
SMX comparison:

JaXvZ.png


I think NV left the TMU configuration intact from GK104, which would mean 240 FP16 samples per clock!
 
currently you have to use an IOMMU that sits on the motherboard, Intel calls it Vt-d, and pass around the whole graphics card to a VM. hypervisor such as Xen or Vmware (ESXi variant) has to support it. same tech allows to use a real network card, storage controller or usb controller.

but doing that, it's one GPU per VM, and one VM per GPU, with implementation limitations that can kick in (it's tricky due to specific nature of the VGA BIOS, so for how many cards will this work? depends on hypervisor, graphics card and motherboard)
the IOMMU translates addressing, commands meant for the VM's address space into real addressing for the physical graphics card, and vice versa.
also if you want to beam framebuffer up through the network, you'll have to do it yourself it's not covered by the technique.

here, and I wondered before if I understood that well from the presentation, an MMU sits in the GPU and communicates with the hypervisor so the transition from virtual to real addressing is done there. so apparently you can use as many physical GPU as you want, and on top of that as many virtual (i.e. as seen by the VM) GPUs as you want.

it's in the scenario of VDI, i.e. one VM per user. I find this to be needless waste (unless Windows 7 pro or enterprise is allowed, saving a lot on licensing). why not have 20 users per VM. but maybe you can have multiple virtual GPUs per VM, I don't know.

[nice that you find something like Xen to be "big iron", I know of a nice installation of it on a pentium E2200 with 2GB ram, running 5 VMs with almost no down time for a few years :)]
 
15 SMXs is a weird number, not only is it not a power of 2 but it's the farthest one can get from a power of 2, in the sense that 15 = 2^3 + 2^2 + 2^1 + 1. I was expecting the possibility of a non-power of 2 after GT200 but I didn't know they could do this.

... first Tesla cards will be equip ed with 13/14 SMX parts.
Similar to GF100 in at least that respect then.

I'm curious, is there any reason not to expect 6 GHz memory on at least one of the first GPUs based on GK110?
 
Back
Top