Overclock.net › Forums › Industry News › Hardware News › [Anand] Tesla P100 PCIE format (no ports) coming in Q3/4 2016
New Posts  All Forums:Forum Nav:

[Anand] Tesla P100 PCIE format (no ports) coming in Q3/4 2016 - Page 4

post #31 of 43
Quote:
Originally Posted by epic1337 View Post

GP102 will be on HBM2, so probably not GDDR5X.

We can hope. Guess NV could go either way, depending on {size, yield, margins}.
But 4 x short cards sure would be nice - free up slot 7 motherboard connectors mad.gif
post #32 of 43
Quote:
Originally Posted by nnunn View Post

We can hope. Guess NV could go either way, depending on {size, yield, margins}.
But 4 x short cards sure would be nice - free up slot 7 motherboard connectors mad.gif

wouldn't work anyway, Nvidia had artificially capped their SLI to only two cards max.
post #33 of 43
Quote:
Originally Posted by epic1337 View Post

wouldn't work anyway, Nvidia had artificially capped their SLI to only two cards max.

Re: 4 x GPU, for OpenCL we use MPI not SLI, so all's good thumb.gif

Re: tri/quad SLI, you still can. Details here: 3way and 4way SLI
post #34 of 43
Quote:
Originally Posted by Oj010 View Post

200 series was Tesla and had display outputs tongue.gif

I was talking about this one :



Not the arch tongue.gif
post #35 of 43
Quote:
Originally Posted by Olivon View Post

I was talking about this one :



Not the arch tongue.gif

I'm fully aware, I made a funny tongue.gif
post #36 of 43
Thread Starter 
Quote:
Originally Posted by nnunn View Post

GPGPU folks, fear not. Think of the 1080 as replacing the 980... that thin 256bit bus means entry-level.

As others are speculating, GP100 stripped of DP and hooked up to 384bit GDDR5X (GP102?) is what 4K and VR really need. And such a hybrid will get us (budget-constrained, single precision) academics signing purchase orders again. Give it a few months. GTX1080's will drop back to $500 and NV's FP32 monster rains on Vega's parade.

PS: many thanks to AMD for Vega thumb.gif

The HBM bus isn't interchangeable and integral to a large portion of the chip. Much more profitable to make the GP102 with the 384 bit GQDR5 bus. There are integrated features in the P100 (no G man) that make it useless for 32bit compared to a 1080 right off the bat. "stripping" a P100 would require making white-out alterations to all sorts of components inside the chip. These things are printed using a series of photographic negatives that are created by a computer.

P100 is nV's 64bit monster. A "chopped down" would still cost close to $6000 and require HBM2. The connections for HBM2 and GQDR5 are not interchangeable. 4096 lanes compressed to 1024 channels high speed is NOT 256 channels. Don't mistake the P100 for a graphics card, it is designed and tuned to do precisely the opposite. It DOES consist of a large number of cuda shader blocks and is capable of making graphics but it's designed to decompose live video into individual objects in real-time. It is the current pinnacle of passive intelligence collection from image input at this time, it can separate hundreds of faces in a picture in a few milliseconds so they can be fed to other hardware for identification.
post #37 of 43
Quote:
Originally Posted by nnunn View Post

Re: 4 x GPU, for OpenCL we use MPI not SLI, so all's good thumb.gif

Re: tri/quad SLI, you still can. Details here: 3way and 4way SLI

That video is incorrect, old news.
post #38 of 43
Quote:
Originally Posted by prjindigo View Post

"The HBM bus isn't interchangeable and integral to a large portion of the chip."

Good point.

Quote:
Originally Posted by prjindigo View Post

"stripping" a P100 would require making white-out alterations to all sorts of components inside the chip..."

On the other hand, assuming there'll be lots of "64bit monsters" that fail validation, simply disabling the 64bit pipes gives you a whole lot of surface area (dark silicon) in precisely the right place to help dissipate the thermal load from all those 14nm fp32 cores.

Hence (as an optimist) I dream of a 2.0 GHz base clock for salvaged GP100's. Now THAT's a Titan rolleyes.gif

Quote:
Originally Posted by Oj010 View Post

That video is incorrect, old news.

Thanks, and sorry for not following this more closely. Of course GPGPU folks who actually need/use quad-gpu normally leave off the SLI straps.
Edited by nnunn - 6/22/16 at 10:19pm
post #39 of 43
even as a compute card, gaming is still possible with this in a VM IMO

virtual gaming with passthrough and streaming is doable
post #40 of 43
Quote:
Originally Posted by akromatic View Post

even as a compute card, gaming is still possible with this in a VM IMO

virtual gaming with passthrough and streaming is doable

Possible. Stupid, but possible.
Arcane
(33 items)
 
Sanctum
(9 items)
 
Fukurou
(15 items)
 
CPUMotherboardGraphicsRAM
Ryzen 7 1700 MSI B350M Mortar Arctic RX Vega 64 G.Skill TridentZ  
Hard DriveHard DriveOSPower
Samsung EVO 960 Consatellation Windows 10 Pro Corsair HX 1050 
Case
Raijintek Aeneas White Window 
CPUMotherboardRAMHard Drive
Core i5-661 GA-H55-USB3 12GB DDR3 1333 WD Caviar Black AALS 640GB 
Hard DriveHard DriveOptical DriveCooling
WD Caviar Green EARS 2TB Vertex Turbo 60GB (SSD) Samsung Blu-ray, Samsung DVD Burner H50 (With push/pull nocturas) 
OSMonitorKeyboardPower
Windows 10 Professional 64 LG 47inch LED LCD Saitek Eclipse Lite-touch Wireless Seasonic X750 Gold 
CaseAudio
CM690 II Passive bookshelfspeakers + amp. 
  hide details  
Reply
Arcane
(33 items)
 
Sanctum
(9 items)
 
Fukurou
(15 items)
 
CPUMotherboardGraphicsRAM
Ryzen 7 1700 MSI B350M Mortar Arctic RX Vega 64 G.Skill TridentZ  
Hard DriveHard DriveOSPower
Samsung EVO 960 Consatellation Windows 10 Pro Corsair HX 1050 
Case
Raijintek Aeneas White Window 
CPUMotherboardRAMHard Drive
Core i5-661 GA-H55-USB3 12GB DDR3 1333 WD Caviar Black AALS 640GB 
Hard DriveHard DriveOptical DriveCooling
WD Caviar Green EARS 2TB Vertex Turbo 60GB (SSD) Samsung Blu-ray, Samsung DVD Burner H50 (With push/pull nocturas) 
OSMonitorKeyboardPower
Windows 10 Professional 64 LG 47inch LED LCD Saitek Eclipse Lite-touch Wireless Seasonic X750 Gold 
CaseAudio
CM690 II Passive bookshelfspeakers + amp. 
  hide details  
Reply
New Posts  All Forums:Forum Nav:
  Return Home
  Back to Forum: Hardware News
Overclock.net › Forums › Industry News › Hardware News › [Anand] Tesla P100 PCIE format (no ports) coming in Q3/4 2016