• Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.

Volta to bring 1 TB/s memory bandwidth to GPUs?

cadaveca

My name is Dave
Joined
Apr 10, 2006
Messages
17,248 (2.45/day)
Nvidia's GTC is on now.

volta.jpg



More here:

http://www.ustream.tv/recorded/30095793
 
That robot in the intro is scary awesome!
 
1TB/sec sounds great, but that's ~4x the bandwidth of current high end GPUs. If bandwidth doubles each generation and Volta is two generations away, then it's normal progress.
 
So Volta is their codename for the generation after Maxwell? Interesting...
 
Nvidia's GTC is on now.

The keynote was 12hrs ago. :slap:

Nothing too exciting just a bunch of graphs and the funniest part was when they were introducing the director for upcoming Fantastic Four and how GRID would help out. The crowd just didnt seam interested and JJ-H was try'n to get them interested. Even the director was like eh..

And not one mention of Project Shield just a reference to it when hes showing off Ira which had the responses scripted since you can tell it resets for each responce on the video.
 
Tegra roadmap:


tegra roadmap.jpg



The keynote was 12hrs ago. :slap:

I had school today, so missed this, some interesting stuff but at the same time seemed like too much pandering. Will be interesting to hear some of the other presentations that'll be given over the next few days, so yeah, GTC is still ongoing. :p Maybe more interesting stuff will come in the next few days.
 
I know the technology is completely different, but I remember Nvidia embedding the VRAM to the RSX GPU substrate for PS3 back in 2006, I think it didn't offer much of a performance advantage back then due to the 128bit bus.

Embedding some sort of cache in the same substrate as the GPU die and giving it a 1024bit direct bus to the GPU could provide the necessary increase in bandwidth.

Since all the traces between the GPU and this cache memory would be relatively short and not an actual part of the PCB design, this approach would be feasible from an economic point of view.

I bet cards would still require more VRAM on the PCB with a more conventional bus in order to reduce design complexity and cost, and allow for more memory availability due to VRAM density not being limited by packaging constrains.
 
Last edited:
Here is what I have to say to Nvidia now after watching a lot of the videos from today.

shut-up-and-take-my-money.jpg
 

Thats another thing that felt flat.

Tegra roadmap needs to be push 2yrs forward to be competative.

Logan should have been what Tegra 4 "Wayne" Tegra4i "Grey" is. Thats why Qualcomm S600 is stealing all the flagship this year. Not one new phone that has been anounced will be using Tegra4i on there flagship product. The only one commited to carring Tegra4i is ZTE.

Parker is going to be a year or more late as thing stand. Dell has been testing AppliedMicro 64bit ARM since late last year and AMD will be producing a hybrid by 2014. By the time Parker rolls out 2nd gen 64bit ARM SoCs will be coming out.
 
1TB/sec sounds great, but that's ~4x the bandwidth of current high end GPUs. If bandwidth doubles each generation and Volta is two generations away, then it's normal progress.

Bandwidth doesn't double with each generation. They've been slacking with bandwidth improvements for awhile now. It's been a pretty slow craw from 100 to 200 GB/s.
 
The question here is, is memory bandwidth really that important? My answer is that GPU clock, throughput, shading perfomance and everything around it is far more important than high raw memory bandwidth.
 
The question here is, is memory bandwidth really that important? My answer is that GPU clock, throughput, shading perfomance and everything around it is far more important than high raw memory bandwidth.

That is what I was wondering. Its a terabyte a second of bandwidth... At the very least it will make memory not a bottleneck for a while.
 
Those have already shown up in recent drivers

NVIDIA_DEV.0FC4 = "NVIDIA D14P1-15"
NVIDIA_DEV.0FC5 = "NVIDIA D15M2-03"
NVIDIA_DEV.0FDA = "NVIDIA GK107-ES-A1"
NVIDIA_DEV.0FE3 = "NVIDIA GeForce GT 745M"
NVIDIA_DEV.0FE5 = "NVIDIA GeForce K340 USM"
NVIDIA_DEV.0FE6 = "NVIDIA NVS K1 USM"
NVIDIA_DEV.0FE7 = "NVIDIA Generic K1 USM"
NVIDIA_DEV.0FEF = "NVIDIA GRID K340"
NVIDIA_DEV.0FF0 = "NVIDIA NB1Q"
NVIDIA_DEV.0FF1 = "NVIDIA NVS 1000"
NVIDIA_DEV.0FF2 = "NVIDIA GRID K1"
NVIDIA_DEV.0FF5 = "NVIDIA Tesla K1 USM"
NVIDIA_DEV.0FF7 = "NVIDIA Quadro K1 USM"
NVIDIA_DEV.1001 = "NVIDIA D15U-60"
NVIDIA_DEV.1003 = "NVIDIA GK110 DT Bringup"
NVIDIA_DEV.1004 = "NVIDIA D15U-50"
NVIDIA_DEV.1182 = "NVIDIA GK104"
NVIDIA_DEV.1184 = "NVIDIA GK104 "
NVIDIA_DEV.118A = "NVIDIA GRID K520"
NVIDIA_DEV.118B = "NVIDIA GeForce K2 USM"
NVIDIA_DEV.118C = "NVIDIA NVS K2 USM"
NVIDIA_DEV.118D = "NVIDIA Generic K2 USM"
NVIDIA_DEV.1192 = "NVIDIA GeForce GK104"
NVIDIA_DEV.119F = "NVIDIA GeForce GTX 780M"
NVIDIA_DEV.11D1 = "NVIDIA GK106-INT343"
NVIDIA_DEV.11D2 = "NVIDIA GK106-INT232"
NVIDIA_DEV.11D3 = "NVIDIA GK106-ES"
NVIDIA_DEV.11E2 = "NVIDIA GeForce GTX 765M"
NVIDIA_DEV.11E3 = "NVIDIA GeForce GTX 760M"
NVIDIA_DEV.11FF = "NVIDIA NB1Q "
NVIDIA_DEV.11B0 = "NVIDIA Quadro K2 USM"
NVIDIA_DEV.11B1 = "NVIDIA Tesla K2 USM"
NVIDIA_DEV.11B6 = "NVIDIA N15E-Q1"
NVIDIA_DEV.11B7 = "NVIDIA N15E-Q3"
NVIDIA_DEV.11B8 = "NVIDIA N15E-Q5"
NVIDIA_DEV.11BB = "NVIDIA Quadro 4100"
NVIDIA_DEV.11C1 = "NVIDIA D14P2-30"
NVIDIA_DEV.11C2 = "NVIDIA GK106"
NVIDIA_DEV.1281 = "NVIDIA D15M2-05"
NVIDIA_DEV.1282 = "NVIDIA D15M2-20"
NVIDIA_DEV.1283 = "NVIDIA D15M2-10"
NVIDIA_DEV.1285 = "NVIDIA GK208-100"
NVIDIA_DEV.1290 = "NVIDIA GeForce GT 730M "
NVIDIA_DEV.1291 = "NVIDIA GeForce GT 735M"
NVIDIA_DEV.1293 = "NVIDIA GeForce GT 730M "
NVIDIA_DEV.1294 = "NVIDIA GeForce GT 740M "
NVIDIA_DEV.12A0 = "NVIDIA GK208"
NVIDIA_DEV.12AD = "NVIDIA GK208-ES"
NVIDIA_DEV.12AE = "NVIDIA GK208-CS1-C"
NVIDIA_DEV.12AF = "NVIDIA GK208-INT"
NVIDIA_DEV.12B0 = "NVIDIA GK208-CS-Q"
NVIDIA_DEV.12B1 = "NVIDIA GK208 INT"
 
Last edited:
Back
Top