Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,678
1400 points more than Fury X. It didn't even scale with clocks due to the slower memory. And rumours has it that its already 225W at this point.
![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Radeon RX Vega Discussion Thread
- Thread starter DPI
- Start date
1400 points more than Fury X. It didn't even scale with clocks due to the slower memory. And rumours has it that its already 225W at this point.
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,678
So the Chinese on the NCU.
Because of the performance indicators we are now getting from IBM with Nvidia for their contracted with the 2017 supercomputers (live to staff 2018) involving Volta + Power9.
Pascal was a technical risk milestone, you do not bring in so many new techs and archs straight into high profile contracted supercomputer projects and proposed also to others.
Pascal allowed combining some high risks early such as node shrink with 600mm2 die+HBM2+mixed precision (1st phase)+NVLink 1.0 (2 is with Volta), with other arch improvements coming with Volta.
Helps it also sold really well, but technically it would be madness for Nvidia to try and do it all in one milestone just with Volta, also they needed to prove the technologies from themselves and IBM so optimisation/modelling/etc could be calculated and any changes or arch updates required completed ahead of schedule.
Anyway the earliest mention of Pascal was 2014 in their presentations, worth noting Volta was mentioned on its own in presentations by Nvidia in 2013.
Nvidia mention both designs in presentations together with their HPC timelines back in early 2014, schedule then was 2017 for Volta and aligned with the supercomputer contracts-projects involving Volta; think of the 2017 date in same way they selectively distributed P100 and possibly one consumer GPU.
Cheers
Does it have to move the bar that much? The Vega leaks make it look like it will struggle to match GTX 1080, and not come anywhere near GTX 1080 ti.
http://wccftech.com/amd-vega-10-3dm...c1-device-8gb-700mhz-vram-1200mhz-core-clock/
"Radeon RX Vega ( Vega 10 ) Is Set To Be A GTX 1080 Contender"
That is assuming they can get the clock speed up to 1500+ MHz. Which is a big assumption for AMD. To match a regular GTX 1080 it looks like AMD will need HBM and water cooling on top.
In all likelihood, NVidia doesn't even have to release Volta new to compete with Vega.
It's very sad, but that is the way it is looking.
Last edited:
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,678
I dont think anyone with common sense ever imaged Nvidia needing Volta for that. Volta however is needed to make the revenue momentum flow running. In short, the competition to Volta is Pascal.
They don't need much performance there either. Top end NVidia buyers have shown they will salivate over minuscule performance increases and fork over another $1200 for a video card. See Pascal Titan X/XP.
NVidia has milking these rubes down to a science.
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,678
Yes, a 699$ 1080ti is complete milking.
People dont upgrade without the reason to do so. And Volta will give a very compelling reason to do so like Pascal.
Vega? Maybe not so much. RTG used its last resources on AI/HPC it seems with graphics as a second thought.
Volta is being driven by the HPC space, plus potentially other upgrades one can expect for Grid/cloud/databasing/improved mixed-precision platform solutions/etc.
But it does not make sense for Nvidia to split the cycles between consumer/prosumer and enterprise/professionals with too large a large delay as the ROI is needed from all market segments (including Tegra) and R&D/manufacturing lines are mutually synchd.
Cheers
This is what I was getting at, people think Volta is the second coming for gaming but all the talk on Volta I have read, show it to be very HPC focused and maybe even a return to a big flexible hardware scheduler.
Unless Volta is on 7nm, it may not even be a perf leap over Pascal for gaming on 16nm, since Paxwell is so lean & efficient already for that purpose.
GP102 is also only 471mm2, if NV wanted to, they can blow it up to 600mm2 for a Paxwell 2.0 in Volta's HPC era to cater to the gaming market. Titan XXX, now 25% faster, $1499. There's no chance for Vega touching that so NV can price it whatever they want and fools will still buy it out.
cybereality
[H]F Junkie
- Joined
- Mar 22, 2008
- Messages
- 8,789
So, linking wccftech, I know. But apparently we will find out something on May 16th.
http://wccftech.com/amd-taking-the-covers-off-vega-navi-may-16th/
http://wccftech.com/amd-taking-the-covers-off-vega-navi-may-16th/
They do this all the timing. Analyst meeting, I don't think it's anything special but we may hear something about Vega given how silent they have been and they probably know they need to say something because we will have another day where people are more disappointed.
Singularity_Survivor
Limp Gawd
- Joined
- Apr 7, 2016
- Messages
- 212
Wait, isn't this a Vega thread?
It is. But seeing as AMD has been so tight lipped about Vega, people have to talk about something.
Anarchist4000
[H]ard|Gawd
- Joined
- Jun 10, 2001
- Messages
- 1,659
More likely to see upcoming products like Naples and the APUs which are relevant to analysts. Doubt anything technical comes out.
Titan X (either version) wasn't a 'miniscule' increase, they were a good 30% better than the next best contender in their lineup.
Titan Xp? Sure, that's milking, so was Titan Black, and probably Titan Z as well, Titan, GTX Titan X and Titan X are not what I call "miniscule milking"
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,678
I would say not quite, Pascal was also engineered for professional/enterprise and yet has massive performance for consumers.
The jump on Volta will be applicable to all segments, remember I said the R&D and manufacturing line is synched between all segments.
Regarding GP102
It has 12TFLOPs FP32 in enterprise/quadro uncut form in an efficient single GPU form.
Also notice this is a professional designed (in your context of why Volta is only professional not consumer performance gains) that came to consumers, so again emphasises my point that performance jumps are also for consumer segment.
Nvidia already has a 600mm2 die (with FP64) and from a cost perspective to scale the full architecture even further by using an FP32 focus such as GP102 as Pascal may not be possible until the Volta arch changes or it creates cost-sales headaches with competing models against the P100.
From a scaling perspective; Remember for awhile Nvidia has used 6 GPCs with 4 SM per GPC in the past and managed to scale a bit more by making it 5 SM per GPC with Pascal, at some point such scaling has diminishing returns without arch changes.
Just look at Vega that has not scaled its core architecture relative to Fiji.
Anyway as I mentioned Pascal was a stepping stone and we are probably likely to see an improved mix-precision architecture with the next 600mm2 die and what it can be used for (compared to P100) along with other arch changes through the range.
Cheers
Last edited:
The performance is just speculation on what has been seen so far so that may change a bit for actual games or when launched.
But interesting if they were told correctly about the availability being low due to HBM2 and costs (no one should be that surprised).
Cheers
Jesus, if that's true they're in horrible shape. The Fury X wasn't terrible, obviously not as fast as the 980Ti, but not that far off. Now they're coming back two years later with a slight improvement? Something seems off with all of these rumors, if they've really made that little amount of progress it's flat out embarrassing.
Fury X is already around 1070 level, +/- some depending on the games. They are all around 980Ti level as is.
AMD is going to spend years on R&D, to give Fury X +10%? With massively more transistors on a new node and new architecture?
Really man give it a break. Do you people not even think before you take these rumors in? If they wanted Fury X +10%, they could have shrank Fiji, enjoy the benefits of 14nm FF vs 28nm, and get a higher clocked Fiji while lowering power usage. No need to bother doing anything else...
I'm starting to see a trend with users here, when Vega rumors are great, they say Volta is coming soon. When Vega rumors are bad, they throw salt on as if it's real.
GP100 was engineered for HPC (512TB virtual address space, proper unified memory, instruction level preemption), the rest of the stack is just Maxwell on speed. Call it what it is.
Will Volta be HPC focused or a repeat of Paxwell, big Volta, aka GV100 will be 1:2 FP64, full on HPC features and smaller Volta, be lean gaming chips? On 16nm? I won't expect much. If it's on 7nm, it could be a major leap.
I'm thinking AMD has kept this thing under wraps so well no one has any idea what it's going to look like when they launch it, and speculation is just speculation. I've never seen this little amount of information leaked about a card before a release maybe ever though, it's amazing how tight lipped they've been. I don't know what angle they're going for, but being quiet just makes me assume the worst. In any case, I bought my Fury Nitro for $225 and i'm still thoroughly enjoying the card.
12nm and 10nm are also possible. There was some news of nvidia taking some wafers from samsung (probably testing for now) so any node provided by TSMC or samsung is a possibility.
It's just sad these tech sites are so clueless that they would even post such an obviously wrong rumor.
To even claim Vega 10 the flagship is on GTX 1070 level is absurd for the simple reason:
Compare what AMD has achieved with Polaris 10 on 14nm FF vs Hawaii on 28nm. Both chips are similar in performance (290X vs RX 480).
Hawaii is 438mm2, 512 bit bus and 250W.
Polaris 10 is 232mm2, 256 bit bus and 160W.
Now people are assuming AMD is repeating a Fury X with Vega 10.
From a 600mm2 on 28nm to 520mm2 on 14nm FF and it's the same class of performance. With 50% more transistors. AMD would have to massively regress in performance with their new architecture for that to even be possible, a huge drop in perf/mm2 and perf/transistor unlike what we've ever witnessed.
12nm last i checked is just a renaming of 16nm FF++ at TSMC. Their 10nm, is that low-power mobile focused? IIRC that was the talk from TSMC & mobile industries.
The high performance node was supposed to be 7nm. Is Volta on 7nm? It's possible, we don't know.
Well exactly, but the trolls are getting so good people are starting to believe them. Logic has no place in an argument here, in this thread, anyhow.
There's no information on Volta at all whatsoever, even less than Vega at this point. Not even really worth discussing; Nvidia doesn't have to make any moves right now, they dominate the market. As long as they're dominant there's no benefit to them displaying their hand at all.
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,678
"7nm" is a 2019 or later node. And its another renamed node like 12nm is.
Also Kepler->Maxwell I guess you remember. Same node, quite the difference. And that's a class example on why R&D matters.
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,678
Ask yourself 2 things. Whats the RTG R&D budget and whats the focus with Vega. Is it gaming or for AI/HPC and what budget is it made on. How many years is irrelevant as such.
Polaris was also impossible...until it was released.
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,678
Lets look at Polaris with perf/watt.
Also the transistors are used for the IN8 and FP16 performance. Not the gaming performance with Vega.
Vega is pretty much Fiji expect FP16 and IN8 and slightly higher clocks.
auntjemima
[H]ard DCOTM x2
- Joined
- Mar 1, 2014
- Messages
- 12,142
They already went to an entirely new node and new architecture and it performed subpar. Considering most of their new cards are competing with NVidias bottom end cards and even last gen is keeping up with it, I would say thats an issue. If you want to look at this through beer goggles, go ahead, but that doesn't mean people are just going to praise AMD's work when it doesn't warrant it.
A new generation should bring with it performance increases. AMD's have not. If we are comparing the rx580 with a 1060 in reviews and you say the Fury X (old tech) is on par with a 1070, does that not show an issue?
You were saying in response to my post that gains would not be there for consumer relative to HPC/enterprise because that is where it is focused.
But in reality Pascal was HPC presented as well!
P100 was the primary GPU presented and can be seen as the stepping stone towards the GV100 for the high profile HPC projects.
But we still had massive gains across the board all the way down.
So again there will be a jump in performance on consumer.
The arch in Pascal was more than just Maxwell on steroids BTW (unified memory/int8/Polymorph engine changes/Cuda core changes/SM structure within GPC/etc goes through the whole range,), and Volta as I keep saying has indicators that its performance is a jump again relative to Pascal.
In reality Pascal and Volta has to support a broad range of segments; consumer/Tegra/Quadro/HPC/databasing/modelling/etc, this applies to the architecture in terms of R&D.
As an example look at what Nvidia is pushing with Drive PX2 and moving to Xavier (Volta), there will be some synergy between consumer-Quadro-Tesla.
To achieve the higher clocks this is a combination of low level silicon-arch improvements with regards to voltage-frequency-performance envelope.
It is wrong to say Pascal is consumer but Volta is HPC focused, both have a lot to do with advances in compute utilisation and scale up/out.
And no way will it be on 7nm, wish those rumours died from WCCFT before they even started as they really do not fit.
Consider Kepler->Maxwell same node but very much matured with arch changes and Pascal->Volta again with a much more matured node (12nm that is the best 16nm offered by TSMC) and arch changes.
Cheers
Last edited:
OrangeKhrush
[H]ard|Gawd
- Joined
- Dec 15, 2016
- Messages
- 1,673
Ya, I don't claim to know enough about these nodes to make an informed guess, but there's more than just TSMC 16nm to use. It's just conceivable to me that nvidia could make volta on one of the new nodes and that might be enough to get an extra 200Mhz across the board or something.
Fury X is dependent upon game when saying at 1070/980ti levels, and one needs to use custom AIB 1070 and 980 ti as these are the primary GPUs for consumers and more real world while Fury X unfortunately was very restricted in these terms.
And also dependent on how measured, really needs 3rd independent such as PresentMon rather than relying upon internal tools that are 'skewed'.
Cheers
What do you mean subpar? AMD only has Polaris on 14nm FF and they target it for low-end and mainstream. They ceded the high-end to NV already, this was well known.
With 28nm, AMD was worse with Hawaii vs GM204. Really, go look up the basics.
AMD is now in a better position than they were with Polaris 10 vs GP106.
RX 480 vs 1060 are neck & neck, so AMD's Polaris is competing with NV's latest gen Pascal chip of similar size.
Suggest you look at this: http://www.techspot.com/review/1393-radeon-rx-580-vs-geforce-gtx-1060/
Over 27 recent games benchmarked. They are so close in performance, and very close in perf/w contrary to what Shintai is fudding about.
30W difference between RX 480 vs 1060. No, you cannot game with your GPU by itself sitting on your desk.
The 390X vs GTX 980 had a much bigger gap in power efficiency.
It looks to me like Polaris GCN iteration nearly caught up to NV's great efficiency in Pascal which uses Tile-based Rasterization for its big perf & perf/w leap.
Vega is bringing Tile-based Rasterization & improved Geometry (a big weakness for AMD GPUs) and suddenly it's a huge regression in performance? -_-
I don't frequent these forums much so I never paid attention to the names of individuals.. but just going through this thread, it's quite clear.
There's some here that loves to shit on AMD no matter what. Even so far as to deny all basic semiconductor logic to assert that Vega, a 14nm FF chip with 50% more transistors and new tech is a Fury X class.
Shintai
Supreme [H]ardness
- Joined
- Jul 1, 2016
- Messages
- 5,678
Even Techreport that's owned by an AMD employee shows this.
Not that its any surprise.