Vega/Navi Rumors (Updated)

[ozzy702]

Well your 1080ti's are GP102 and regular 1080 is GP104 so no surprise there.

Yeah, I would really be surprised if 1080ti level performance doesn't comes in a single 8 pin with Volta. Hopefully AMD has the cash to throw at graphics from here forward.

Threadripper sure is looking tempting but I think I'll wait for V2 so clocks will be in the 4ghz + range.

 

[happy medium]

Not sure how it's less efficient. Vega RX 64 is a 295 watt TDP card. Fury X was 275 watt. So Vega RX 64 uses 7% more power (if we go by TDP) but offers 20-35% more performance depending on resolution.

This means Vega RX is more efficient not less.

Well that's not fair. The LC Vega is a 350 watt card vs the 275 watt LC Fury X . The fastest Vega card is not much more power efficient than a Fury X.

"
"The Radeon RX Vega 64 graphics cards are powered by dual 8 pin power connectors and will require beefy power supply units to keep them fed under heavy gaming loads. The TDP for the Liquid cooled models will be configured at 350W."
http://wccftech.com/amd-radeon-rx-vega-64-56-official-slide-performance-specs-price-leak/

 

[raghu78]

But efficiency has regressed in it's relative competitive position.

I think a big crux of the Vega letdown was endless hype creating expectations of a comeback for Vega, a shrinking of the gap between AMD and NVidia, and finding out that gap has instead grown larger.

If you look at it from any metric, and compare the relative position of:

Fury X vs 980/980Ti
To
Vega RX vs 1080/1080Ti

The gap has worsened in every metric. 2 years later, after endless hype from AMD, everything is in a worse relative position than it was for Fury X. Performance is relatively worse, die size is relatively worse, Power usage is relatively worse, time to market is worse, every metric is worse across the board.

Everything is relatively worse than it was for the Fury X launch, which in itself was no great leap.

In some ways I think it would have been better if AMD never mentioned/delivered Vega as a gaming card. I think if it hadn't been for mining propping up prices, it might not have feasible to deliver it against a $400 1080/$300 1070.

Nvidia has been able to get close to 2x performance (>80% improvement) from Titan Maxwell to Titan Pascal. AMD Vega 10 is struggling to improve performance by 40% on Fury X. The perf/watt, perf/sq mm of Vega is worse than Fiji and the gap between Nvidia and AMD has widened massively. AMD's marketing tried to build hype knowing very well Vega is a turd for gaming and effciiency. But all of that is of no use when actual reviews land in 2 weeks.

 

[leoneazzurro]

Nvidia has been able to get close to 2x performance (>80% improvement) from Titan Maxwell to Titan Pascal. AMD Vega 10 is struggling to improve performance by 40% on Fury X. The perf/watt, perf/sq mm of Vega is worse than Fiji and the gap between Nvidia and AMD has widened massively. AMD's marketing tried to build hype knowing very well Vega is a turd for gaming and effciiency. But all of that is of no use when actual reviews land in 2 weeks.

perf/W we will see at RX launch, but perf/mm^2? Fiji is 600 mm^2, Vega is 484 mm^2 and the latter is faster. Did you mean perf/transistor (which will be also debatable until the launch)?

 

[Stuka87]

Need to remember one of the reasons Vega has such a large die, and so many transistors is is because it has a lot of compute added on. Something GP104 does not have. So while gaming performance may match up with GP104, overall compute does not. Its closer to GP102 in that regard. Not that it match GP102 in efficiency either, just saying its efficiency cannot be directly compared to GP104, which was designed as a gaming card, and ONLY a gaming card. Vega is trying to be two different cards.

 

[swilli89]

perf/W we will see at RX launch, but perf/mm^2? Fiji is 600 mm^2, Vega is 484 mm^2 and the latter is faster. Did you mean perf/transistor (which will be also debatable until the launch)?

He's probably referring to its relative performance against Nvidia's offerings. And.. considering 14LPP offers

14LPE is the first foundry process technology manufactured in the foundry industry with the successful volume ramp-up. 14LPE offers 40% faster performance; 60% less power consumption; and, 50% smaller chip area scaling as compared to its 28LPP process.

14LPE is the first foundry process technology manufactured in the foundry industry with the successful volume ramp-up. 14LPE offers 40% faster performance; 60% less power consumption; and, 50% smaller chip area scaling as compared to its 28LPP process.

Keep in mind the above is 14nm LPE. LPP should be 10% faster. So for a process that claims 40% faster performance.. Vega roughly delivers. Looking like 40% will be the very best case against Fiji. Alright so AMD designed at least, to spec for 14nm and this is in-line with that.

How about the 60% less power consumption? Not even close, Vega basically equals Fiji and might even eclipse it. This is a complete fail, based on what we know.

How about 50% smaller chip? 600mm^2 should have scaled to 300mm^2 given the same basic chip layout. AMD barely managed 20% smaller chip, not to mention this new Vega chip only has 2 memory controller interfaces instead of the 4 that Fiji used die area on.

I just can't, for the life of me, understand what AMD has been doing. A 14nm-ized Fiji would have been a bit faster than the Vega we have today, would have consumed less than 200 Watts, and would have been a lot smaller and thus cheaper for AMD to produce.

 

[Mopetar]

Need to remember one of the reasons Vega has such a large die, and so many transistors is is because it has a lot of compute added on. Something GP104 does not have. So while gaming performance may match up with GP104, overall compute does not. Its closer to GP102 in that regard. Not that it match GP102 in efficiency either, just saying its efficiency cannot be directly compared to GP104, which was designed as a gaming card, and ONLY a gaming card. Vega is trying to be two different cards.

That would be okay if you could buy the compute card at gaming card prices, but you need to buy the more expensive professional or prosumer versions of AMD (and NVidia) cards in order to unlock some of those features.

All that aside though, you can still compare Vega to something like that Titan Xp which does fill the same niche and Vega still comes up short for the same reasons. Over the long run I think Vega will age better than Pascal, but that has as much to do with NVidia being quick to abandon older platforms as it does with AMD being able to wring out my performance through drivers and developers utilizing some of the various functionality in Vega that can't be brought out through drivers.

 

[crisium]

Wasn't Hawaii a "compute card"? It is only 10% larger than the pure gaming GM204 (GTX 980), and performance pretty similar now-a-days, although a bit behind at launch. But even at launch, Hawaii : Maxwell never had as bad size : performance ratio as Vega vs Pascal. Vega is more than 50% larger than GP104 (GTX 1080).

I think I heard that Hawaii was the last design with remnants of the old ATi guard, who were sacked and replaced afterwards. Budget matters. But I don't buy the "compute" excuse. It's poor execution from no money.

 

[Stuka87]

Wasn't Hawaii a "compute card"? It is only 10% larger than the pure gaming GM204 (GTX 980), and performance pretty similar now-a-days, although a bit behind at launch. But even at launch, Hawaii : Maxwell never had as bad size : performance ratio as Vega vs Pascal. Vega is more than 50% larger than GP104 (GTX 1080).

I think I heard that Hawaii was the last design with remnants of the old ATi guard, who were sacked and replaced afterwards. Budget matters. But I don't buy the "compute" excuse. It's poor execution from no money.

Oh, not trying to make up excuses. I agree budget was the #1 reason Vega is not the card we all hoped it would be. I was making a comment because many people above are comparing efficiency with GP104, when they should compare with GP102. Even if the performance does not match, they are closer in design with each other.

 

[french toast]

Not sure how it's less efficient. Vega RX 64 is a 295 watt TDP card. Fury X was 275 watt. So Vega RX 64 uses 7% more power (if we go by TDP) but offers 20-35% more performance depending on resolution.

This means Vega RX is more efficient not less.

But the uarch probably is worse, remember Vega benefits from a mature and optimised 14nm FF, nearly 2 node jumps from Fiji 28nm hp.
Tht would easily account for 35℅ efficiency.

 

[Veradun]

But the uarch probably is worse, remember Vega benefits from a mature and optimised 14nm FF, nearly 2 node jumps from Fiji 28nm hp.
Tht would easily account for 35℅ efficiency.

It's nearly 1 node jump, actually. Well, nope, it's exactly 1 node jump.

 

[PeterScott]

It's nearly 1 node jump, actually. Well, nope, it's exactly 1 node jump.

It's two nodes. 28->20 would be one node. 20->14 would be another.

Though the official ITRS Node designation are:
32->22->16

The above is the same scale just displaced by half node. 32->28 is half node. 22->20 is half node, 16-14 is half node.

 

[Qwertilot]

Well that's not fair. The LC Vega is a 350 watt card vs the 275 watt LC Fury X . The fastest Vega card is not much more power efficient than a Fury X.

Although I'm not sure if it is quite fair to use the FuryX/liquid cooled LC Vega as points to measure efficiency. They were both clocked way past the point of sensible returns. Might look better with stuff running at saner clocks - its not impossible that AMD were so desperate for every ounce of performance that they configured their stock clocks to utterly insane this time round.

The real issue with Fiji was in some ways that it didn't replace the 390(X). It logically should/would have done but they couldn't produce enough of them etc.

 

[Elixer]

Really. I think this is an extremely weak excuse, when he has been leading them for 4.5 years. Even if there were some preliminary design notes, there was plenty of time to review them and get Vega back on track.

You vastly underestimate how long it takes to make a GPU, let alone make a new GPU on a new node. Vega is really late because of all the technical challenges & the delayed HBM2 that still isn't performing within specs.

So, while he might be able to tweak a few things late in the game, there just wasn't any time to do a major design changes.

On average, it takes 3-6 years to do a whole new design depending on the complexity.
Look at Ryzen.

DON WOLIGROSKI: Off the top of my head, I believe it was 4 or 5 years ago now, around 2012. Before my time at AMD, I started my tenure here at the beginning of 2015. The promise of the Zen architecture is one of the reasons I came to AMD in the first place.

That took 4-5 years.
Navi will take 4-5 years as well. Say they started in preliminary design in 2014, that will make it 2018/19.

and updated to

They are once again doing a new node on a new design, so, I am betting they will run into issues yet again.
That is why Navi is looking to be late 2018 or most likely 2019, unless the 7nm goes super smooth.

 

[PeterScott]

You vastly underestimate how long it takes to make a GPU, let alone make a new GPU on a new node. Vega is really late because of all the technical challenges & the delayed HBM2 that still isn't performing within specs.

I don't think so. It is naive to think the design is set in stone 5 years before shipping. It could be 5 years or more if you never designed GPUs before, but for companies turning out annual updates no way. I didn't work in GPUs, but Telecom, and we developed new chips, and in many ways it was worse for us because we didn't turn out annual updates to the same chip families, but did it intermittently, still It only took our chip development about 2 years start to finish.

There may have been some goals/ideas 5 years ago, but not much else. This isn't a new architecture either, it still GCN based with some new tweaks. If it took AMD 5 years to do this, they are really screwed.

Pretending the guy in charge for 4.5 years isn't to blame for this product is absurd.

 

[Peicy]

You vastly underestimate how long it takes to make a GPU, let alone make a new GPU on a new node. Vega is really late because of all the technical challenges & the delayed HBM2 that still isn't performing within specs.

So, while he might be able to tweak a few things late in the game, there just wasn't any time to do a major design changes.

Just because they started to design Vega...maybe 4 years ago, does not mean that there are no major changes made along the way. Its not like after 6 months everything is set in stone and they are just waiting for silicon to come back. Raja was is full control of this thing, he is at AMD since mid 2013.
Another thing is that designing a large high-peformance GPU is a huge team efford. You can have a great captain at the helm as a plus.

On average, it takes 3-6 years to do a whole new design depending on the complexity.
Look at Ryzen.

That took 4-5 years.
Navi will take 4-5 years as well. Say they started in preliminary design in 2014, that will make it 2018/19.

Something else you are not considering here: Zen is a completey new design. Sure lessons learned from the past are incoporated (see energy saving features), but the architecture is completely new.
Vega is a revamped GCN architecture. Sure there is a lot of new stuff, but its by no means as much of a clean slate as Zen is.

What i do find much more likely than the whole "Raja had not so much control over this" is:

Zen and Vega were designed during phases when AMD was piss poor and bleeding talent. They may have just focused most money and efford on Zen, and it was worth it. GPU design may have taken a back seat in terms of ressources. Simple as that.

 

[Snarf Snarf]

I don't think so. It is naive to think the design is set in stone 5 years before shipping. It could be 5 years or more if you never designed GPUs before, but for companies turning out annual updates no way. I didn't work in GPUs, but Telecom, and we developed new chips, and in many ways it was worse for us because we didn't turn out annual updates to the same chip families, but did it intermittently, still It only took our chip development about 2 years start to finish.

There may have been some goals/ideas 5 years ago, but not much else. This isn't a new architecture either, it still GCN based with some new tweaks. If it took AMD 5 years to do this, they are really screwed.

Pretending the guy in charge for 4.5 years isn't to blame for this product is absurd.

5 years ago AMD assumed discrete GPU was a dying market, and cut their previous design team loose. Raja even said as much in his AMA on reddit a couple months back. His words were "we lost vision on the high end GPU market" which is why Polaris was designed, it was a mainstream product designed for higher volume. Vega essentially is the worst case scenario in every aspect for any leader, new team that hasn't made a large monolithic GPU before, small budget, and very short release window. They made obvious trade offs in the design of this card and at a high level it looks like GCN with different things tacked on to it.

I said this like 20 pages back but AMD's goals do not always have to agree with your expectations for a product. This could still be seen as an internal success due to only the upper management knowing how bad of a situation Raja was given. This is one of those situations where speculation is all we can do, if Koduri keeps his job, it's because he met his goals, and the company thinks he did a good job. If that's the case it's reasonable to assume upper management gave him very specific goals to meet (probably data center, machine learning, and production) and that's where a bulk of the design was focused on. Gamer's will be upset at the performance, but without knowing what volume this product was expected to sell, it may be a complete afterthought if AMD sells enough Vega to higher margin markets. Failure is an extremely subjective term, and depending on what angle you take can be viewed as success (mostly higher ASP/margins, or increased market share in new markets) to RTG management.

 

[Cookie Monster]

My train of thought is that Navi will again be an updated GCN architecture (and could be the last one), down shrinked to 7nm and an emphasis on perhaps MCM based solutions via infinity fabric (potentially useful for GPGPU/HPC market, NOT the gaming market - think along the lines of NVlink).

I think the change (that most users are looking for and that puts them back on the competitive landscape with nVIDIA) in architecture will happen in their "next gen" architecture, whatever that is.

 

[Stuka87]

It's two nodes. 28->20 would be one node. 20->14 would be another.

Though the official ITRS Node designation are:
32->22->16

The above is the same scale just displaced by half node. 32->28 is half node. 22->20 is half node, 16-14 is half node.

No, it isn't. 28 to 20 is a "half node". Just to like 20 to 14 is half a node.

 

[PeterScott]

My train of thought is that Navi will again be an updated GCN architecture (and could be the last one), down shrinked to 7nm and an emphasis on perhaps MCM based solutions via infinity fabric (potentially useful for GPGPU/HPC market, NOT the gaming market - think along the lines of NVlink).

I think the change (that most users are looking for and that puts them back on the competitive landscape with nVIDIA) in architecture will happen in their "next gen" architecture, whatever that is.

I also feel quite strongly that Navi is another GCN tweak.

That is kind of the vibe I get from the slide deck that shows Vega, then Navi, and then the thing after is called "Next Gen".

So wait for Next Gen?

 

[PeterScott]

No, it isn't. 28 to 20 is a "half node". Just to like 20 to 14 is half a node.

A full node is .7 linear shrink
A half node is a .9 linear shrink.

I think people are getting mixed up because they remember a full node is half size/double density, and it is. But that is based on area and node names are based on linear measurement.

.7 * . 7 = .49, so ~half the area, double the transistor density.

28 *.7 = 19.6 (rounded to 20) Full Node
20 *.7 = 14 Full Node

https://www.semiwiki.com/forum/f2/20nm-half-node-1402-2.html
https://en.wikichip.org/wiki/technology_node

28->14 was massive jump because they skipped a full node in between, so this is two full node jump. 1/2 the linear measure, 1/4 the area, 4 times the transistor density.

 

[swilli89]

A full node is .7 linear shrink
A half node is a .9 linear shrink.

I think people are getting mixed up because they remember a full node is half size/double density, and it is. But that is based on area and node names are based on linear measurement.

.7 * . 7 = .49, so ~half the area, double the transistor density.

28 *.7 = 19.6 (rounded to 20) Full Node
20 *.7 = 14 Full Node

https://www.semiwiki.com/forum/f2/20nm-half-node-1402-2.html
https://en.wikichip.org/wiki/technology_node

28->14 was massive jump because they skipped a full node in between, so this is two full node jump. 1/2 the linear measure, 1/4 the area, 4 times the transistor density.

But Samsung only claims 50% chip size going from 28nm to 14nm. Not 25%.

 

[PeterScott]

But Samsung only claims 50% chip size going from 28nm to 14nm. Not 25%.

I should have added. "In Theory".

Remember Intel complaining that it's competitors node names were bullshit? This is what it is all about.

The intention of Node Names were that full Nodes were to be double transistor density so ~.7 linear scaling.

But marketing departments have got too involved in the naming process, and thus we get a lot of fudging of supposed full node advances.

Do you have link where Samsung claims that. I would love to see in writing where they admit that 2 full Marketing nodes is only equal to one real node.

 

[Konan]

http://www.tweaktown.com/news/58635...-leaked-benchmarks-gtx-1070-killer/index.html

An industry source of mine has provided me with some raw benchmark numbers on Radeon RX Vega 56, which looks to be the new $400 mainstream king, beating the GTX 1070 in some of the biggest games on the market. My source said that the RX Vega 56 card was running on an Intel Core i7-7700K @ 4.2GHz, had 16GB of DDR4-3000MHz RAM, and was running Windows 10.


The benchmarks were run at 2560x1440 with the AMD Radeon RX Vega 56 easily beating NVIDIA's GeForce GTX 1070 in Battlefield 1, DOOM, Civilization 6, and even Call of Duty: Infinite Warfare. My source said that Battlefield 1 was run on Ultra settings, Civ 6 was on Ultra with 4x MSAA, DOOM was at Ultra with 8x TSAA enabled, and COD:IW was running on its High preset.

Radeon RX Vega 56 benchmark results:
Battlefield 1: 95.4FPS (GTX 1070: 72.2FPS)
Civilization 6: 85.1FPS (GTX 1070: 72.2FPS)
DOOM: 101.2FPS (GTX 1070: 84.6FPS)
COD:IW: 99.9FPS (GTX 1070: 92.1FPS)

 

[PeterScott]

http://www.tweaktown.com/news/58635...-leaked-benchmarks-gtx-1070-killer/index.html

I think the RX 580 beats the GTX 1060 in those same benchmarks:
http://www.guru3d.com/articles_pages/asus_radeon_rx_580_strix_review,17.html

Maybe the Vega 56 has a bit more of leg up on 1070, than 580 does on 1060 (if you think it does at all in wider field of benchmarks).