• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

Nvidia Volta is 16nm, expected in May 2017

I

IgotElbows

Member
Will any of these be low any GPUs?

I have a 1060 3gb in the mail right now. (Bought for 175$ including tax)
Wondering if I should send it back when I get it or just keep it.
 
D

Deleted member 465307

Unconfirmed Member
IgotElbows said:
Will any of these be low any GPUs?

I have a 1060 3gb in the mail right now. (Bought for 175$ including tax)
Wondering if I should send it back when I get it or just keep it.
Click to expand...

There will almost certainly be low-end Volta chips, but they're probably at least 4 months, maybe more like 10 months, away. I think you should keep it unless you want to wait until late 2017/1H 2018.
 
A

AmyS

Member
Nvidia asked Square Enix to do something with Tesla V100 / Volta, so they did a real-time demo of Kingsglaive: Final Fantasy XV using the Luminous Engine.

jbBAMYf.jpg


rOuZZUh.jpg


https://www.cool3c.com/article/124078

https://www.kotaku.com.au/2017/05/n...live-blog-follow-all-the-news-as-it-happened/

The Tesla V100 is an enterprise GPU - but it still does nifty game graphics. We were shown a character model from the Square Enix game Kingsglaive: Final Fantasy XV which boasted some of the most photorealistic fabrics we've ever seen.
Click to expand...
 
I

Izuna

Banned
Servbot24 said:
This is the wrong attitude. For all things, but especially for tech. The want should be for your laptop to accomplish what you need it to do.
Click to expand...

I'm not super serious, but at the same time.. it's my very first time with a top spec machine. It's surreal
 
X

Xyber

Member
aajohnny said:
So glad I didn't pull the trigger on a 1080 ti
Click to expand...

Why exactly? You won't be buying this chip anyway and whatever consumer card Nvidia will put out (maybe towards the end of this year) will probably be at around the same performance as the 1080Ti.

If you are looking at having the best card that is not a Titan, then buying the Ti at launch is the best option if you are willing to spend that much money. Waiting another 6-8 months or so (from the Ti launch) to have a ~15% better GPU for about the same price isn't worth it if you are already thinking you need to upgrade.

I personally don't think the consumer card of Volta will be out this year, still think it will be very early next year. Perhaps November if it is this year, but Q3 still seems too close for them.
 
R

rtcn63

Member
Xyber said:
Why exactly? You won't be buying this chip anyway and whatever consumer card Nvidia will put out (maybe towards the end of this year) will probably be at around the same performance as the 1080Ti.

If you are looking at having the best card that is not a Titan, then buying the Ti at launch is the best option if you are willing to spend that much money. Waiting another 6-8 months or so (from the Ti launch) to have a ~15% better GPU for about the same price isn't worth it if you are already thinking you need to upgrade.

I personally don't think the consumer card of Volta will be out this year, still think it will be very early next year. Perhaps November if it is this year, but Q3 still seems too close for them.
Click to expand...

To be fair, if they made the next XX70 comparable to the current TI for ~$400, it could be worth if he doesn't absolutely need the card right away (just speculation as the 1070 is ~980ti). But yeah, He'll probably be waiting until 2018 at least. Not to mention the likely markup/supply/early QC issues at launch...
 
M

michaelius

Banned
X

Xyber

Member
rtcn63 said:
To be fair, if they made the next XX70 comparable to the current TI for ~$400, it could be worth if he doesn't absolutely need the card right away (just speculation as the 1070 is ~980ti). But yeah, He'll probably be waiting until 2018 at least. Not to mention the likely markup/supply/early QC issues at launch...
Click to expand...

Well yeah, that's the decision you have to make. Get the performance of the xx80Ti right now, or wait another half a year or more and get that performance for a better price.

That's why I say it's best to buy the Ti cards at launch if you are willing to spend that money, because the longer you wait the more it makes sense (money-wise) to just wait some more for the next gen cards. And we all know that waiting for a new GPU is awful. :p

I upgraded from a 1080 to a 1080Ti because I wanted to jump over to the Ti-upgrade path and now I'm just gonna ignore the 104 cards from here on out. It hurt a bit paying another 400€ after selling my 1080 for the performance boost I got now, but I still think it was worth it in the end.
 
I

Izuna

Banned
Durante said:
I don't want to disappoint you, but your laptop, being a laptop, was never "top spec".

:p
Click to expand...

I dunno what the terminology is, if top-spec means 1080 Ti or w/e, but a 6700HQ and 1070 gets all jobs done :)

I think the best game I've seen so far was Ghost Recon. I played 4 player among 980 Ti and Sli 970 PCs and this laptop held up.

~~

also since you're reading this. I love you durante
 
dr_rus

dr_rus

Member
Inside Volta: The World's Most Advanced Data Center GPU

SXM2-VoltaChipDetails.png


image7.png
Click to expand...

The new Volta SM is 50% more energy efficient than the previous generation Pascal design, enabling major boosts in FP32 and FP64 performance in the same power envelope. New Tensor Cores designed specifically for deep learning deliver up to 12x higher peak TFLOPs for training. With independent, parallel integer and floating point datapaths, the Volta SM is also much more efficient on workloads with a mix of computation and addressing calculations. Volta's new independent thread scheduling capability enables finer-grain synchronization and cooperation between parallel threads. Finally, a new combined L1 Data Cache and Shared Memory subsystem significantly improves performance while also simplifying programming.
Click to expand...

With 84 SMs, a full GV100 GPU has a total of 5376 FP32 cores, 5376 INT32 cores, 2688 FP64 cores, 672 Tensor Cores, and 336 texture units. Figure 4 shows a full GV100 GPU with 84 SMs (different products can use different configurations of GV100). The Tesla V100 accelerator uses 80 SMs.
Click to expand...

image3.png
Click to expand...

Similar to Pascal GP100, the GV100 SM incorporates 64 FP32 cores and 32 FP64 cores per SM. However, the GV100 SM uses a new partitioning method to improve SM utilization and overall performance. Recall the GP100 SM is partitioned into two processing blocks, each with 32 FP32 Cores, 16 FP64 Cores, an instruction buffer, one warp scheduler, two dispatch units, and a 128 KB Register File. The GV100 SM is partitioned into four processing blocks, each with 16 FP32 Cores, 8 FP64 Cores, 16 INT32 Cores, two of the new mixed-precision Tensor Cores for deep learning matrix arithmetic, a new L0 instruction cache, one warp scheduler, one dispatch unit, and a 64 KB Register File. Note that the new L0 instruction cache is now used in each partition to provide higher efficiency than the instruction buffers used in prior NVIDIA GPUs.
Click to expand...
This will most certainly be different on consumer Volta parts as such granularity is likely unnecessary for graphics. L0 I-cache will probably stay though.

Durante said:
Interesting to see that the configurable shared memory size is back.

Also, it appears like the divergence of gaming and HPC GPUs is progressing even further.
(Though you could argue that it has already peaked this gen by an HPC-only die existing)
Click to expand...

Yeah, I wonder how much of GV100 will even appear in GV10x gaming chips. Tensor cores are most certainly out, FP64 will probably be stripped down again, HBM2 won't happen. FP16x2 seems like a solid addition for gaming chips this time around. General flops number can also be a good indication of Volta lineup balancing - it's probably safe to expect 1.5x tflops from each respective Volta GPU compared to Pascal. But then we don't know the rest of gaming related changes yet, and so far Volta looks like a very big update comparable to Maxwell.

icecold1983 said:
So is consumer volta expected to just be wider maxwell/pascal?
Click to expand...
No.

icecold1983 said:
Architecture innovations again only for the hpc market?
Click to expand...
Again?
 
D

Durante

Member
Izuna said:
I dunno what the terminology is, if top-spec means 1080 Ti or w/e, but a 6700HQ and 1070 gets all jobs done :)
Click to expand...
Don't worry, I was just joking around.
(I think gaming laptops are not a very good investment, but even I have to admit they are a lot better these days than they used to be)
 
ethomaz

ethomaz

Banned
dr_rus said:
Yeah, I wonder how much of GV100 will even appear in GV10x gaming chips. Tensor cores are most certainly out, FP64 will probably be stripped down again, HBM2 won't happen. FP16x2 seems like a solid addition for gaming chips this time around. General flops number can also be a good indication of Volta lineup balancing - it's probably safe to expect 1.5x tflops from each respective Volta GPU compared to Pascal. But then we don't know the rest of gaming related changes yet, and so far Volta looks like a very big update comparable to Maxwell.
Click to expand...
Pretty much this for GV104.

- 5120 SPs @ 1800Mhz = ~18TFs FP32
- ~36TFs FP16 (not removed this time)
- 1/8 FP64 perf. (capped again)
- GDDR6 @ 14Gbps with 384bits = 676GB/s
- No Tensor Cores
- 128 ROPs
- 320 TMUs

Make it happen.
 
I

Izuna

Banned
Durante said:
Don't worry, I was just joking around.
(I think gaming laptops are not a very good investment, but even I have to admit they are a lot better these days than they used to be)
Click to expand...

Pascal line solidified by choice. The price I got this at, it was cheaper than a desktop (same spec) + laptop for school. Ofc, as I'm bringing up, I can't upgrade this thing so I have to hope I can get £800 or so back when I wanna upgrade or go desktop.

~~

As such, I'm going to be keeping an eye on mobile Volta GPUs, whether it's like Pascal or they go back to a gimped line.
 
R

rtcn63

Member
In a few years, you'll probably be able to pick up a 1080p/60fps gaming-capable laptop (even if just an older model one) for pennies. What. A. Time.
 
I

Izuna

Banned
rtcn63 said:
In a few years, you'll probably be able to pick up a 1080p/60fps gaming-capable laptop (even if just an older model one) for pennies. What. A. Time.
Click to expand...

Well, if we're talking 7th Gen games, the integrated graphics from intel can already do that since Haswell

>_<
 
K

Killer Queen

Banned
Izuna said:
Pascal line solidified by choice. The price I got this at, it was cheaper than a desktop (same spec) + laptop for school. Ofc, as I'm bringing up, I can't upgrade this thing so I have to hope I can get £800 or so back when I wanna upgrade or go desktop.

~~

As such, I'm going to be keeping an eye on mobile Volta GPUs, whether it's like Pascal or they go back to a gimped line.
Click to expand...
I don't see them going back to doing mobile GPUs. But I feel the same way you do about not being top spec. Ill just build a PC next year though as i was already planning.
 
D

Durante

Member
ethomaz said:
Pretty much this for GV104.

- 5120 SPs @ 1800Mhz = ~18TFs FP32
- ~36TFs FP16 (not removed this time)
- 1/8 FP64 perf. (capped again)
- GDDR6 @ 14Gbps with 384bits = 676GB/s
- No Tensor Cores
- 128 ROPs
- 320 TMUs

Make it happen.
Click to expand...
Far too big a die for an x04 I'd say.
 
dr_rus

dr_rus

Member
ethomaz said:
Pretty much this for GV104.

- 5120 SPs @ 1800Mhz = ~18TFs FP32
- ~36TFs FP16 (not removed this time)
- 1/8 FP64 perf. (capped again)
- GDDR6 @ 14Gbps with 384bits = 676GB/s
- No Tensor Cores
- 128 ROPs
- 320 TMUs

Make it happen.
Click to expand...

That's too much for GV104, I would expect 5376 FP32 SPs from GV102 and 2/3 that from GV104 (3584). Double rate FP16 should be in now since it shouldn't affect the DL inferencing markets this much with Tensor cores in GV100 + AMD will have it and chances are that games will start making use of it.

3584 SPs with 50% better power efficiency should be able to hit ~15TFlops of FP32 throughput pretty easily while maintaining GP104's ~180W power draw. Add SM improvements, scheduling improvements and possibly even higher clocks - and it seems like such GV104 should be able to beat 1080Ti by the same ~30% as GP104 did with GM200.
 
M

michaelius

Banned
ethomaz said:
Pretty much this for GV104.

- 5120 SPs @ 1800Mhz = ~18TFs FP32
- ~36TFs FP16 (not removed this time)
- 1/8 FP64 perf. (capped again)
- GDDR6 @ 14Gbps with 384bits = 676GB/s
- No Tensor Cores
- 128 ROPs
- 320 TMUs

Make it happen.
Click to expand...

That's what GV102 will look like more likely

Also 384 bits = 96 rops
 
I

icecold1983

Member
dr_rus said:
Inside Volta: The World's Most Advanced Data Center GPU










This will most certainly be different on consumer Volta parts as such granularity is likely unnecessary for graphics. L0 I-cache will probably stay though.



Yeah, I wonder how much of GV100 will even appear in GV10x gaming chips. Tensor cores are most certainly out, FP64 will probably be stripped down again, HBM2 won't happen. FP16x2 seems like a solid addition for gaming chips this time around. General flops number can also be a good indication of Volta lineup balancing - it's probably safe to expect 1.5x tflops from each respective Volta GPU compared to Pascal. But then we don't know the rest of gaming related changes yet, and so far Volta looks like a very big update comparable to Maxwell.


No.


Again?
Click to expand...

all the architecture innovation of pascal was limited to gp100
 
Locuza

Locuza

Member
icecold1983 said:
all the architecture innovation of pascal was limited to gp100
Click to expand...
Instruction/Pixel-Level-Preemption and Dynamic Load Balancing was available on all consumer Pascals.
HW page faulting probably too.

Also every Pascal under GP100 introduced at least better DCC, improved Multi-Projection capabilities and 4x int8.
 
O

Orayn

Member
I desperately want to know when desktop Volta cards are coming out before I blow way too much money on a 1080Ti. I just know the 2070 will be comparable.
 
I

icecold1983

Member
Locuza said:
Instruction/Pixel-Level-Preemption and Dynamic Load Balancing was available on all consumer Pascals.
HW page faulting probably too.

Also every Pascal under GP100 introduced at least better DCC, improved Multi-Projection capabilities and 4x int8.
Click to expand...

is this even of any use outside of deep learning?

back on topic, come on dude those are incredibly minor improvements. smaller than the smallest gcn revision
 
R

Renekton

Member
Orayn said:
I desperately want to know when desktop Volta cards are coming out before I blow way too much money on a 1080Ti. I just know the 2070 will be comparable.
Click to expand...
You won't know until one month before launch. We only guesstimate the window to be between 2017-09 to 2018-02.
 
Locuza

Locuza

Member
icecold1983 said:
is this even of any use outside of deep learning?

back on topic, come on dude those are incredibly minor improvements. smaller than the smallest gcn revision
Click to expand...
Probably not but you talked about that "all the hardware innovation" were limited to the GP100 which isn't true.
And of course the improvements were overall relatively minor but let's not act as if Polaris was so much more.

Also Pascal brought Conservative Rasterization Tier 2 for DX12.
 
W

wachie

Member
What a trainwreck of a conference. Not an Nvidia conference without a super awkward Jen Hsun being his super tyrant self.

Having said that, most of the technological parts were cool.
ethomaz said:
That is dam impressive specs on paper.

Probably 30TFs FP16.

Vega for HPC will probably born dead.
Click to expand...
AMD never had a chance in HPC, atleast in a serious way. They will do what they do best, play the pricing game. The big players are not strapped on cash so they will still go with Nvidia.
 
Compsiox

Compsiox

Banned
The 1080ti is definitely worth getting. It's a very powerful caRd. You guys really don't need to wait for that like 12% performance increase.
 
I

icecold1983

Member
Locuza said:
Probably not but you talked about that "all the hardware innovation" were limited to the GP100 which isn't true.
And of course the improvements were overall relatively minor but let's not act as if Polaris was so much more.

Also Pascal brought Conservative Rasterization Tier 2 for DX12.
Click to expand...

polaris improvements are absolutely more meaningful.
 
Locuza

Locuza

Member
icecold1983 said:
polaris improvements are absolutely more meaningful.
Click to expand...
7% on average vs. the previous gen and more if you are looking for use cases where the improved DCC and the Primitive Discard Accelerator really helps.
But you can also construct cases for Pascal with time critical VR applications which are using async timewarps or Multi-Projection for decreased workload.
Or looking for concurrent graphics and compute work which is now dynamically adjusted.
The DCC also got on Pascal better.

Do we really want to fight on this low level and use cases here and there?

All started with the notion if Volta only brings innovation to the GV100 and the consumer variants are just wider Maxwell/Pascals didn't it?
And consumer Volta definitely won't be.
 
I

icecold1983

Member
Locuza said:
7% on average vs. the previous gen and more if you are looking for use cases where the improved DCC and the Primitive Discard Accelerator really helps.
But you can also construct cases for Pascal with time critical VR applications which are using async timewarps or Multi-Projection for decreased workload.
Or looking for concurrent graphics and compute work which is now dynamically adjusted.
The DCC also got on Pascal better.

Do we really want to fight on this low level and use cases here and there?

All started with the notion if Volta only brings innovation to the GV100 and the consumer variants are just wider Maxwell/Pascals wasn't it?
And consumer Volta definitely won't be.
Click to expand...

gcn revisions are mostly improvements that help performance across the majority of typical workloads. pascal has specific corner case improvements
 
R

Renekton

Member
Any info on this TSMC 12nm FFN?

FinFET Next?

GlobalFlounder's "7nm or bust" decision makes me even more worried for AMD.
 
U

Unknown Soldier

Member
Anandtech has a detailed article about Volta now.
http://www.anandtech.com/show/11367...v100-gpu-and-tesla-v100-accelerator-announced

Basically: Volta is an entirely new architecture. The entire compute array has been replaced, the old Maxwell/Pascal compute cores are gone and now we have TENSOR CORES.

They are building GV100 to the limit of the TSMC process, the die literally cannot get any bigger on the 12nm FFN ("FinFet Nvidia") process. This process is a specially customized version of TSMC 12nm FF intended for Nvidia's high-performance demands.

In every way, this is a GPU that pushes the limits of GPUs today. The demand for deep learning compute is strong enough that Nvidia will probably have way more buyers than they will have working GV100 dies to sell. If Anandtech's speculation is correct and each GV100 will sell for $15k+, Nvidia won't need to sell a lot of them to make a whole lot of money. As with Pascal, the first GV100 cores will go into the new DGX-1V. If you've got $150,000 laying around, you can get Volta today.

There's no reason to assume there will be consumer Volta before 2018, especially if Vega underperforms.

Locuza said:
No, they only licensed 14nm LPE and LPP from Samsung.
For 7nm and other process variants they want to go there own way.
Click to expand...

That doesn't sound like a good idea, but I'm not GF so who knows.
 
You must log in or register to reply here.

Similar threads

NVidia announces a 21% decrease in revenue YoY
Business Hardware
28
3K
winjer replied
NVIDIA Ada Lovelace GPUs For GeForce RTX 40 Series To Bring The Same Generational Jump As Maxwell To Pascal
2
78
9K
Dream-Knife replied
Nvidia Announces Next-Generation Orin SOC on 7nm EUV
5
1K
Seko replied
NVIDIA’s Next-Gen GPU Specifications & Performance Leaks Out – Massive Die With 7936 CUDA Cores (8192 Full Die), Up To 48 GB HBM2e Memory
17
2K
kittoo replied
NVIDIA Will Be Announcing 7nm Ampere GPUs At GTC 2020 In March
News 2
51
5K
D.Final replied
Top Bottom