will they disable the "Tensor" units on gamer hardware

Is tensor core just another name for cuda cores, also known as shader cores or stream processors?

>>60321745
Who the fuck cares its another Pascal.

>>60321794
It FP16 X FP16 i think

It's not wasted die area. They "bin" the bad chips otherwise it'd just be trash

Like, they had all the R&D time in the world and all we got was another Pascal?

>>60321794

no, i think it's special hardware that does 2*FP16 MUL + FP32 ADD. assuming it will be useless for gaming.

File: 1455593384134.jpg (4KB, 250x250px) Image search: [Google]

4KB, 250x250px

>>60321828
>talk about Pascal
>Who the fuck cares its another Maxwell
>talk about Volta
>Who the fuck cares its another Pascal

>>60321828

not really. if you google for pascal block diagrams, you'll see that the SMs are different.

File: NVIDIA-Telsa-V100.jpg (1MB, 1920x1080px) Image search: [Google]

1MB, 1920x1080px

> 815 mm^2

holy shit! is it the largest CPU/GPU ever made? what's the die area of jewtel's largest xeon?

>>60321944
Consumer models will have no Tensor Memes so it's going to be another Pascal.
>>60321972
Intel had Poulson with ~700mm^2.

>>60321875

try again newfag ;)

https://devblogs.nvidia.com/parallelforall/inside-volta/

> Unlike Pascal GPUs, which could not execute FP32 and INT32 instructions simultaneously, the Volta GV100 SM includes separate FP32 and INT32 cores, allowing simultaneous execution of FP32 and INT32 operations at full throughput, while also increasing instruction issue throughput.

>>60321999

how big the the latest xeon phi? (knight's something)

>>60322050
Landing?

>>60321999
>Consumer models will have no Tensor Memes so it's going to be another Pascal.

No way they disable it, it almost half size in die. nvidia will make it work in game

>>60322074
How? It's intended for ML.

>>60322037
Wow they separated the shaders. It's still Pascal.

>>60321794
It's the same idea as
https://en.wikipedia.org/wiki/Tensor_processing_unit

>>60322074
Oh they WILL disable it, and sell models without it disabled for hundreds more.

File: image4.png (16KB, 1024x254px) Image search: [Google]

16KB, 1024x254px

>>60321794

Each Tensor Core provides a 4x4x4 matrix processing array which performs the operation \textbf{D} = \textbf{A} \times \textbf{B} + \textbf{C}, where \textbf{A}, \textbf{B}, \textbf{C}, and \textbf{D} are 4×4 matrices as Figure 7 shows. The matrix multiply inputs \textbf{A} and \textbf{B} are FP16 matrices, while the accumulation matrices \textbf{C} and \textbf{D} may be FP16 or FP32 matrices.

>>60322184

sorry, copypasta didn't work. read the original text here...

https://devblogs.nvidia.com/parallelforall/inside-volta/

>>60321828
This. It's a rebrand.

>>60322074
They probably won't even have those on die in consumer models.

Amd hype fp16, Nvidia now with all new tensor core. Amd step behind again

>>60322359
>We literally have another 20 days before we even see Vega for 3 minutes

>>60322330
And it'll be basically Pascal. Meh.

File: kepler_block.jpg (224KB, 1286x854px) Image search: [Google]

224KB, 1286x854px

>>60322301

it's got lots of those little green squares like Kepler! Volta is just a Kepler rebrand!

>>60322466
And Kepler is gutted Fermi.

>>60322489

and Skylake is Nehalem on steroids

and /g/tards don't know shit about microarchitecture

>>60322524
And Nehalem is drugged P6.

>implying Sandy Bridge wasn't just Pentium 5

>>60322301
3bilion usd for a rebrand, come troll. Nvidia not that stupid

It's waste of space on consumer chips. I'm sure they remove most of the 64 and 16 bit units.

>>60322586
And then it's Pascal.

>>60322077
Gameworks

>>60322184
This really doesn't seem that impressive, just basic 1/2 rate for high end cards.

>>60322657
>This really doesn't seem that impressive, just basic 1/2 rate for high end cards.
Because it's nothing impressive. Fuck it's only 15tflops FP32 on boost clocks with ~800mm^2 die. Weird.

>>60322674

it's not targeted at fp32 gaymers, it's targeted at HPC and ML

>>60322685
Which is fucking FP8/FP16 at best.

>>60322685
Who cares?
if you want to do 2xfp16 calculations this wont make it go any faster as output is bottlenecking, there is no real advantage of this shit, at least by the info nVidia gave.

>>60322698

hpc is all fp64 you dummy

>>60322726
HPC is extremely small and GV100 is presumably aimed at ML market. Which uses FP8/FP16. And ASICs.

>>60321972
what's the yield on this?
1.7%?

>>60323622
Ye. Rev up those memes.