AMD Radeon RX VEGA to Launch at SIGGRAPH 2017

[Taf the Ghost]

35 minutes ago, DXMember said:

competition is green with envy, having to produce 800mm² dies

Haha, good one.  Though that's really what is going on. The GV100 with those Tensor Cores (which are really cool), is a massive die. GPUs are already massively parallel processors, so being able to stack dies together should work just fine.  Once you're used to building that direction. AMD could also create something like the Tensor Cores themselves and put them as a separate die on the package. But that's for engineers & designers paid a lot of money to figure out.

[Belgarathian]

On 6/1/2017 at 11:51 PM, leadeater said:

Also the 5870, X1950XTX Ultra or w/e the hell it was called.

X1900XTX. I had one of these beasts, the fastest DX9 card of all time. 

 

[Princess Luna]

I like AMD CPUs but sorry GPUs are still nVidia department for me.

[leadeater]

5 hours ago, LAwLz said:

But it's not... Volta will be released soon, so there will only be a few months at most where Vega is competing against Pascal. Once Volta is out, it will be Vega vs Volta until AMD releases a new architecture, which won't be for like another year.

I don't think Volta will be that soon at all, it's barely even present in the HPC/server market yet. Then there is the competition aspect of it, if Vega isn't soundly beating Pascal then there is no reason to rush in to gaming Volta at all and just keep bringing a higher rate of return on Pascal consumer products and Volta server products.

 

Volta consumer will have no tensor engine on it at all along with other die modifications and we have no idea how much down the line Nvidia has actually worked on that yet.

[LAwLz]

1 hour ago, leadeater said:

I don't think Volta will be that soon at all, it's barely even present in the HPC/server market yet. Then there is the competition aspect of it, if Vega isn't soundly beating Pascal then there is no reason to rush in to gaming Volta at all and just keep bringing a higher rate of return on Pascal consumer products and Volta server products.

 

Volta consumer will have no tensor engine on it all all along with other die modifications and we have no idea how much down the line Nvidia has actually work on that yet.

Q3 release for Volta according to roadmaps. That means it will be released within the next 3 months. Of course that might chance but I think we should expect it. Pascal is over a year old at this point so it would be naive to assume that Nvidia hasn't been working on a successor and is readying it right now. It's not like Nvidia is run by idiots who will have no plan in case Vega turns out to be really good. 

 

Nvidia have had a very long time to prepare and make sure they can be one step ahead. Don't Sasuke they they have just been sitting on their butts and given AMD the opportunity to catch up. 

 

If it's needed, Nvidia will most likely be able to push out Volta shortly. 

[leadeater]

26 minutes ago, LAwLz said:

Q3 release for Volta according to roadmaps. That means it will be released within the next 3 months. Of course that might chance but I think we should expect it. Pascal is over a year old at this point so it would be naive to assume that Nvidia hasn't been working on a successor and is readying it right now. It's not like Nvidia is run by idiots who will have no plan in case Vega turns out to be really good. 

 

Nvidia have had a very long time to prepare and make sure they can be one step ahead. Don't Sasuke they they have just been sitting on their butts and given AMD the opportunity to catch up. 

 

If it's needed, Nvidia will most likely be able to push out Volta shortly. 

Q3 for Tesla V100 as far as I know, all major publications so far have said 2018 for stuff we are interested in.

 

Quote

NV is expecting the first GV100 products to start shipping in Q3 of this year and they won’t be cheap – on the order of $18K per GPU for a DGX system – however for those customers with deep pockets and who will essentially pay NVIDIA to eat the risk of producing such a large GPU, this will be the most powerful GPU released yet.

http://www.anandtech.com/show/11367/nvidia-volta-unveiled-gv100-gpu-and-tesla-v100-accelerator-announced

 

Edit:

And what I was saying is there is no reason to push forward the release date of consumer Volta from 2018 unless there is a need to. I'm not saying Nvidia aren't doing anything, they are working on server and workstation products for Volta. They are no doubt working on GeForce products but not as a higher priority than those server/workstation products.

 

Also a year old GPU architecture is not that old. Fermi was 2010, Kepler was 2012, Maxwell was 2014 and Pascal was 2016. Notice a pattern 

[Taf the Ghost]

18 minutes ago, leadeater said:

Q3 for Tesla V100 as far as I know, all major publications so far have said 2018 for stuff we are interested in.

 

Edit:

And what I was saying is there is no reason to push forward the release date of consumer Volta from 2018 unless there is a need to. I'm not saying Nvidia aren't doing anything, they are working on server and workstation products for Volta. They are no doubt working on GeForce products but not as a higher priority than those server/workstation products.

 

Also a year old GPU architecture is not that old.

It's generally been around 18 months between GPU cycles, compared to about 2 years for CPUs. Though, obviously, everything is bring strung out now as die shrinks are massively more expensive. (And they were never cheap in the first place.)  So Q2'2018 for consumer Volta is probably correct, completing a 2 year cycle this time. Volta is also a die-shrink from 14 nm to 12 nm, so we'll seen to see what that actually brings.  Though it can generally be assumed that 1080 Ti class goes to mid-tier and on down the line, with a new top-tier.  (Though with how big the dies are getting, expect the new top tier to be maybe even more spendy.  But, that might be why we're seeing 2 year cycles, so all of the costs can come down.)

[MoonSpot]

1 hour ago, leadeater said:

Also a year old GPU architecture is not that old. Fermi was 2010, Kepler was 2012, Maxwell was 2014 and Pascal was 2016. Notice a pattern 

Yes.  They're moving up the alphabet!  Now there's "Volta", next there's James Watt(I'd guess).  After that there isn't many more letters left; that'll be the end of GPU progress, the end of moore's law and when the hollow earth collapses in on itself because the marshmallow pillars can no longer keep the multiverse apart and the full weight of the 22 other dimensions come crashing down on the value of the dollar and bitcoin inherits the moon.

 

Can't believe the master plan was right there in front of our eyes all along.

 

(/me buys 4 gpus to sli in his bomb shelter)

[leadeater]

22 minutes ago, Taf the Ghost said:

It's generally been around 18 months between GPU cycles

Nvidia has been very consistent with 24 months released in April for a while now, see my later re-edit.

[zMeul]

4 hours ago, Belgarathian said:

X1900XTX. I had one of these beasts, the fastest DX9 card of all time. 

 

that was back when ATi made GPUs and not being controlled by the dipshits at AMD

I loved ATi, really did! 14y of only ATis

[ravenshrike]

5 hours ago, Belgarathian said:

X1900XTX. I had one of these beasts, the fastest DX9 card of all time. 

 

Aww it looks so tiny and cute compared to modern cards.

[Taf the Ghost]

48 minutes ago, leadeater said:

Nvidia has been very consistent with 24 months released in April for a while now, see my later re-edit.

It looks like GP100 cards landed in early Q2'2016, and we're looking at GV100 cards in Q4'2017. But, correct, as die shrinks have gotten more expensive, we're on full 2-year cycles for GPUs now.  Looking to be almost 3 years for CPUs.  I was commenting more in general over the last 30 years, but, well, things have changed.

 

Here's something to back up your Nvidia point: 

[leadeater]

1 minute ago, Taf the Ghost said:

It looks like GP100 cards landed in early Q2'2016, and we're looking at GV100 cards in Q4'2017. But, correct, as die shrinks have gotten more expensive, we're on full 2-year cycles for GPUs now.  Looking to be almost 3 years for CPUs.  I was commenting more in general over the last 30 years, but, well, things have changed.

 

Here's something to back up your Nvidia point: 

Yep the older generations of Nvidia products that released faster were the same architecture while donig node shrinks. Tesla (the architecture) before Fermi spanned across 90 nm, 80nm, 65nm, 55nm, and 40nm which were the GeForce 8, GeForce 9, GeForce 100, GeForce 200 and GeForce 300 product lines.

[Taf the Ghost]

4 minutes ago, leadeater said:

Yep the older generations of Nvidia products that released faster were the same architecture while donig node shrinks. Tesla (the architecture) before Fermi spanned across 90 nm, 80nm, 65nm, 55nm, and 40nm which were the GeForce 8, GeForce 9, GeForce 100, GeForce 200 and GeForce 300 product lines.

The fascinating part about Computer Tech is that we've got a generation coming up (everyone under age 12 right now) that simply isn't going to see the same type of gains we've been used to seeing for about 60 years now, all because of the grand problem: Physics!  Darn that physics, it'll always get ya!

 

Yet, somehow, webpages will still load slower in 2030 than they do in 2000.  Unless Flash is finally dead by then. 

[Drak3]

1 hour ago, Taf the Ghost said:

Unless Flash is finally dead by then.

Flash will never die completely. Though flash based sites and ads can die in the fire of a vengeful god and their creators can freeze eternally in the last ring of hell. Newgrounds and mah porn can stay.

[Taf the Ghost]

50 minutes ago, Drak3 said:

Flash will never die completely. Though flash based sites and ads can die in the fire of a vengeful god and their creators can freeze eternally in the last ring of hell. Newgrounds and mah porn can stay.

It was sad when I realized that Geocities sites made in HTML 4.0 with rotating .gif banners are normally better coded than most modern websites with 20 applets set to run from a baseline xml calling a CSS file. It used to be you had to wait for the page to download; now you have to wait for all of the instance of Flash to load. (And people wonder why AdBlockers have been massively adopted.)

[LAwLz]

8 hours ago, leadeater said:

Q3 for Tesla V100 as far as I know, all major publications so far have said 2018 for stuff we are interested in.

 

http://www.anandtech.com/show/11367/nvidia-volta-unveiled-gv100-gpu-and-tesla-v100-accelerator-announced

 

Edit:

And what I was saying is there is no reason to push forward the release date of consumer Volta from 2018 unless there is a need to. I'm not saying Nvidia aren't doing anything, they are working on server and workstation products for Volta. They are no doubt working on GeForce products but not as a higher priority than those server/workstation products.

 

Also a year old GPU architecture is not that old. Fermi was 2010, Kepler was 2012, Maxwell was 2014 and Pascal was 2016. Notice a pattern 

Hmm that might be correct. But if the 2 year cycle thing is true (which it isn't, we have gotten new generations of graphics cards quicker than that historically) then Vega will compete with Pascal for something like 6 months, and compete with Volta for another 18 months.

That still means it needs to be competitive against Volta. 

[tom_w141]

15 hours ago, LAwLz said:

Once Volta is out, it will be Vega vs Volta until AMD releases a new architecture, which won't be for like another year.

Navi and Volta are both marked for 2018 so just no... Though Volta will likely be 1st. I think Navi will be the show stopper not Vega to be honest. Navi is supposed to be 2 GPUs linked by IF and recognised as just 1 GPU by the system, that will be incredible performance and Nvidia don't have an equivalent technology (as far as we know). Process shrinks have become hard work now and what we are seeing with AMD is they are using their unique position of being a CPU and GPU manufacturer to share technologies between the 2. With the process shrink getting more difficult AMD is gearing up in both CPU & GPU to focus on modularity, which on the CPU front is already paying off for them.

15 hours ago, LAwLz said:

Vega will need to be as good or better than Volta, not Pascal.

Whilst AMD might be late to the party they aren't competing with Volta that's not how it works. That's like saying When Intel launch coffee lake that Ryzen will compete against coffee lake, it won't Zen 2 will. What you compete against is whatever is in the market at the time of release.

Edited July 3, 2017 by tom_w141

[Taf the Ghost]

4 minutes ago, tom_w141 said:

Navi and Volta are both marked for 2018 so just no... Though Volta will likely be 1st. I think Navi will be the show stopper not Vega to be honest. Navi is supposed to be 2 GPUs linked by IF and recognised as just 1 GPU by the system, that will be incredible performance and Nvidia don't have an equivalent technology (as far as we know)

Whilst AMD might be late to the party they aren't competing with Volta that's not how it works. That's like saying When Intel launch coffee lake that Ryzen will compete against coffee lake, it won't Zen 2 will. What you compete against is whatever is in the market at the time of release.

Well, NVlink does exist, much in the way Intel has EIMB. They're steps in that direction & useful tech, but they're not going all-in like AMD is with the parallelization approach of processor design.  Which is a little odd for Nvidia as GPUs are already completely parallelized processing units. (We when you have >4000 CUDA cores or Stream Processors, we've entered that territory.)  This also isn't too surprising given both companies' market leading positions. It's a lot of money & time investment to switch up approaches that much.

[tom_w141]

1 minute ago, Taf the Ghost said:

It's a lot of money & time investment to switch up approaches that much.

I did an edit to the quoted post that was missed by your quote, sorry.

[LAwLz]

13 minutes ago, tom_w141 said:

What you compete against is whatever is in the market at the time of release.

No, that's stupid. You compete against what consumers can buy at any given point in time. If Volta was released the day after Vega you can't just say "well Vega was released when Pascal was the latest so therefore Vega is not to compete against Volta". 

 

Vega will compete with Pascal at first, and then Volta, unless AMD releases another generation of GPUs before Volta is out. 

[tom_w141]

6 minutes ago, LAwLz said:

Volta was released the day after Vega

If that was the case you'd have an argument but 6+ months no that's not a competing product. That's the next offering from the competition which your next product has to compete against. That's like saying a bulldozer CPU is supposed to compete with a Kaby Lake CPU because both are available for purchase. NO.

6 minutes ago, LAwLz said:

Vega will compete with Pascal. at first, and then Volta, unless AMD releases another generation of GPUs before Volta is out.

Fixed

 

EDIT: Volta should drop around Q2 2018. Assuming its start of Q2 so April, that is 9 months after Vega... Definitely not meant to compete with Volta.

[Prysin]

Ah, seems like the queen of veasel words are back in action. Facinating argument, except its not an argument.

[leadeater]

22 minutes ago, LAwLz said:

Hmm that might be correct. But if the 2 year cycle thing is true (which it isn't, we have gotten new generations of graphics cards quicker than that historically)

You're quite welcome to go look up those release dates, I did to make sure that information was correct. Also remember not all products were released at the same time for new architectures so it's not even necessary the correct dates to be using if you just want to look at the base top offering i.e. non Ti.

 

GeForce: 8800 series (G80, Tesla): 8 November 2006

GeForce: 9800 series (G92, Tesla): 1 April 2008

GeForce 100 series rebrand, OEM only (G92b, Tesla): 10 March 2009

GeForce 280 (GT200-300-A2, Tesla): 17 June 2009

GeForce 300 series rebrand, OEM only (GT21x, Tesla): 27 November 2009

GeForce 480 (GF100, Fermi): 26 March 2010

GeForce 580 (GF110, Fermi): 9 November 2010

GeForce 680 (GK104-400-A2, Kepler): 22 March 2012

GeForce Titan (GK110, Kepler): 19 Feburary 2013

GeForce 780 (GK110, Kepler): 23 May 2013

GeForce 980 (GM204, Maxwell): 18 September 2014

GeForce 1080 (GP104-400-A1, Pascal): 27 May 2016

 

Since the discussion is mainly around architectures and when they get released I'm not aware of Nvidia ever having a shorter release span than 24 months so don't expect Volta any sooner than Jan 2018 otherwise you'll be more disappointed than Vega fans hoping it will be faster than the 1080Ti.

[leadeater]

55 minutes ago, LAwLz said:

Vega will compete with Pascal for something like 6 months, and compete with Volta for another 18 months.

That still means it needs to be competitive against Volta. 

Vega mostly likely will end up competing against Pascal and Volta, and in regards to Volta we know what kind of shit show that is going to be for Vega. Vega will have to be the fastest selling card they have ever made to recoup costs from development because as soon as Volta hits Vega sales will stop, or might as well say they have considering how low they will be.