Vega/Navi Rumors (Updated)

[CatMerc]

Its funny. Scott Wasson has in-depth explanation on reasons why Nvidia hardware clocks so high.

Make your minds yourselves guys. It is also a clue to what AMD meant for Vega: "optimized for higher IPC and higher core clocks".

It's simple, for early 28nm GCN as a design was the best balance AMD could think of in terms of width and speed. Problem is that 28nm matured, and then 16nm/14nm happened, but GCN didn't evolve to accomodate that. Right now the best approach is higher clocks than what GCN is built to do.

 

[lobz]

Its funny. Scott Wasson has in-depth explanation on reasons why Nvidia hardware clocks so high.

Make your minds yourselves guys. It is also a clue to what AMD meant for Vega: "optimized for higher IPC and higher core clocks".

I sweared to myself some time ago, that I'd never watch or read anything again that has anything to do with the gamersnexus guy, so I don't care even if he's praising vega.

 

[ElFenix]

So to be clear.. Not that I have ANY intention of, but we could waltz into the Nvidia Volta thread and just start threadcrapping with FUD, claiming it as a rumour.. And that's okay?


But calling someone out on their FUD, is bad.

What was wrong with the second of Valantar's quoted posts was the mocking nature of it, not the challenge to Mockingbird for his speculation.

As Valantar said:

However, I don't believe in ignoring trolls, misinformation spreaders or people showing awful judgement - allowing their statements to stand unchallenged is too easily read as acquiescence (or at the very least lack of objection). Whether it's a raving tech fanboy/hater, some bumbling MAGA-Trumpist, anti-vaxxers, climate deniers, deregulation-obsessed libertarians, some alt-right neo-nazi a**hole, or just some logic-challenged overly impressionable teenager, they all need to be challenged and met with thoughtful, logical and thorough argumentation at every step. If not, they soon start drowning out the sensible people. Which makes for awful, awful discussion.

The bolded is what makes good forums good. We need more of that.

AT Moderator ElFenix

 

[Crumpet]

What was wrong with the second of Valantar's quoted posts was the mocking nature of it, not the challenge to Mockingbird for his speculation.

As Valantar said:

The bolded is what makes good forums good. We need more of that.

AT Moderator ElFenix

Thank you for the clarification.

 

[tential]

I hope Vega is very fast. If it is close to a 1080ti, I may be in for 2, depending on price. I suspect it to be somewhere between 1080 and 1080ti, or perhaps very close to 1080ti with less OC headroom. Of course, only time and playing the silicon lottery will tell.

I want 2, but I'm trying to realize that. 1 may be slower sure, but amd shines with multiple gpus at highend which is why I waited. Sli sucks.

 

[Mockingbird]

Not even close.. Paying $3.2Billion too much for dodgy ATI drivers in 2006 is probably their biggest mistake.

That's why I said "one of".

AMD's real problem was that it had a revolving door of CEOs who think that they need to do something big to change the dynamic of the company, when what AMD really needed was a CEO that ran the company well from day to day.

 

[n0x1ous]

I want 2, but I'm trying to realize that. 1 may be slower sure, but amd shines with multiple gpus at highend which is why I waited. Sli sucks.

multi-gpu sucks on both sides with DX12 and Vulkan taking over. Multi-gpu is just dying.....

 

[guskline]

I think Dr. Lisa Su will be the CEO of AMD for awhile considering her performance so far.

 

[krumme]

That's why I said "one of".

AMD's real problem was that it had a revolving door of CEOs who think that they need to do something big to change the dynamic of the company, when what AMD really needed was a CEO that ran the company well from day to day.

Without ATI purchase where would amd stand today?

Who would have poured money into a company with no gpu no consoles only a bd arch? - no obligations. Mubadala have effectively acted as cash reserve when the situation was most bad.

Besides from the way to high price for ATI in hindsight it still seems fine strategic way.

Yeaa Hector was a crook but hey he was battling Otellini that was an outright business gangster.

 

[Paratus]

If we are trying to predict Vegas performance compared to an NV competitor based on die size it's good to point out that last generation Fury X and 980Ti were both right at 600mm^2. In general they performed on par. Fury taking a slight lead at 4K and Vulcan, 980Ti taking the lead back when OC'd.

If Vega has a significantly larger die size than the 1080Ti/ Titan Xpp it's not unreasonable to assume it will have greater performance.

If that seems weird consider the reportedly similar specs to Fury, (4096 SPs) that previous GPUs from different generations with the same basic layout have had significantly different performances.

Going all the way back to the X1800XL vs X850XT, both cards had similar clocks speed and were 16 pipe cards yet the XL was between 10% and 50% faster

http://www.anandtech.com/show/1810/10

More recently the 7970/280x vs the 470. Both are 2048 shader parts but the 470 is about 35% faster. I'd expect Vega will remove any internal bottlenecks Fury had.

If Vega does have the larger die size I expect it to outperform the 1080Ti. It will still be less power efficient than the 1080Ti mostly because I expect it to include professional capabilities that NV dropped from Pascal because AMD can't afford separate mainstream and professional GPUs.

 

[itsmydamnation]

It is very possible for Vega to only be 10% faster than the 1080. Even considering improvements.

The 2304SP RX 580 has 80% more shaders than the GTX 1060. It has more bandwidth. It has fewer ROPs. It has lower frequency. It performs the same as the GTX 1060, which tends to have more OC headroom.

Using the same shader ratio, Vega at 4608 shaders would equal a GTX 1080, and have less OC headroom.

So only having 4096 shaders and being 10% faster is an improvement. And we can see why. Vega has a relatively larger bandwidth advantage. Vega has equal ROPs, not fewer. Vega can (hopefully) lessen the frequency discrepancies. Vega has all the NCU features (tile-based rasterization, front-end, bla-bla-bla).

So what am I leaving out? Seems like a good improvement over Polaris 10 vs Pascal, right?

The only thing really off here is the die size. RX 580 is about 15% larger than the GTX 1060. Even at 500mm^2, which tends to be among the lower range of the estimates, Vega is more than 50% larger than the 1080. It would be very bad for this massively larger chip to be barely faster, and equal or possibly slower after both max OC. I hope it isn't true.

What's almost certainly not true is that Vega is cancelled.

This is a terrible comparison, it make no attempt to understand the strengths or deficits of the uarch and then forward extrapolates as a result the confidence factor is extremely low. NV have a massive bandwidth and power advantage as well as raster performance they gained when they updated their ROP's and ROP caches with maxwell. So AMD is also doing simlar things with Vega that NV did from kepler to maxwell, There are the updates to the ROP's and ROP caches, There is also a change of we dont know what size to the compute engine, There is also a change to the front end and geometry.


What you have just done is use Kepler to extrapolate Maxwells performance. So by your logic a 780ti beats a 980ti right..............

 

[Ajay]

Its funny. Scott Wasson has in-depth explanation on reasons why Nvidia hardware clocks so high.
Make your minds yourselves guys. It is also a clue to what AMD meant for Vega: "optimized for higher IPC and higher core clocks".

Wow, that was painful to watch. Scott is new to being on the other side of the discussion, so some hesitation is to be expected. I suppose he was purposefully dumbng down the material. Anyway, optimizing for both power and clock speed is more challenging that doing one or the other. That being said, both have always mattered to some degree - so it is a plainly understood problem for ASIC designers. In the age of sub 45nm tech, optimizing for power is becoming increasingly necessary. Let's all just hope AMD was successful and brings competition back to the enthusiast gamer market.

I am curious to know if there was something that may have pushed out AMD's timeline on Ryzen. It still seems to me that a Ryzen/Vega one-two punch would have delivered higher sales than the larger stagger we are seeing.

Oh, one last point - it seems that TSMC 16FF+ was already geared a higher clocks than GF 14LPP. NVidia and AMD have both designed accordingly.

 

[crisium]

This is a terrible comparison, it make no attempt to understand the strengths or deficits of the uarch and then forward extrapolates as a result the confidence factor is extremely low. NV have a massive bandwidth and power advantage as well as raster performance they gained when they updated their ROP's and ROP caches with maxwell. So AMD is also doing simlar things with Vega that NV did from kepler to maxwell, There are the updates to the ROP's and ROP caches, There is also a change of we dont know what size to the compute engine, There is also a change to the front end and geometry.

I'm just playing devils advocate. AMD can claim to have new features, and even parity with Nvidia (tile based rasterizer, L2 cache improvements, memory compression, etc). But that doesn't mean they will execute it as efficiently as Nvidia. A reminder:
http://cdn.wccftech.com/wp-content/uploads/2016/01/AMD-Polaris-15.jpg

And it often loses to the two-generation older Hawaii in performance-per-tflop, despite "New" improvements (mainly due to the unbalanced design, imo). I won't rule out the possibility that AMD's implementations of new features are done poorly, and/or won't be made manifest by software for many months or several years. Nvidia has far more resources to make sure it is done right.

But again, I am using the possibility that they will offer improvements already in this possible scenario. Remember, 480:1060 is 4608SP AMD card to 1080. Not 4096SP and 10% faster. I give credit to possible Vega uarch improvements, just highlight that they may not be spectacular. It comes down to your optimism in AMD's execution.

It's really the die size that is best is Vega's favour. When I remove the devil from my shoulder, I find it hard to believe it won't be closer to 1080 Ti than 1080. It'd be shocking not to. I'm hopeful for even more.

What you have just done is use Kepler to extrapolate Maxwells performance. So by your logic a 780ti beats a 980ti right..............

This doesn't make any sense. 980 Ti would have to regress in performance-per-TFLOP, significantly, for that to happen. I expect some performance-per-TFLOP improvement in Vega compared to Polaris 10, even in a "worst case scenario" where it only 10% faster than a 1080 that I already highlighted (and hope it doesn't happen).

Just playing a part. Bottomline: Vega can improve over Polaris 10 and have a situation where it is only 10% faster than a 1080, simply because Polaris 10 x2 would trade with the 1080. I hope it to be better than that though.

 

[itsmydamnation]

This doesn't make any sense.

Exactly!
I'm just playing devils advocate on your post.

 

[crisium]

The better comparison to what I am saying is if I expected Nvidia to have poor execution and the 980 non-Ti would only trade even with a 780 Ti, or slightly lose - both of which it could do and still have architectural improvements. No where did I say AMD would regress, which is what the 980 Ti would have had to do.

Read again. It's hard to have a discussion with someone who fundamentally does not understand the other viewpoint. Devil's advocacy is different from trolling.

 

[piesquared]

Doesn't sound like a miracle to me, just a good but expected job. I doubt AMD designed it for any less than it launched at, and the GloFo process had time to mature.


Actually, Intel clocks a full GHz higher than Ryzen, so I don't see what you mean.

As far 14nm vs 16nm, we can look at 1050 Ti. The thing maxes out at around 1.9GHz, as compared to the 2.1GHz of its 16nm brethren.

Like I said, Glo Fo's 14nm is more efficient or similar to intel's 14nm at the 3.5 GHz and lower range. Some people are saying Glo Fo's 14nm is a liability for Vega, clearly it could reach quite high if the architecture is designed to do so.

Also, can you show me the 16nm chips running at 4 GHz?
No we can't look at the 1050Ti because we don't know the clocks it would operate at on 16nm.

 

[antihelten]

But again, I am using the possibility that they will offer improvements already in this possible scenario. Remember, 480:1060 is 4608SP AMD card to 1080. Not 4096SP and 10% faster.

You can't really scale things this way though. A 1080 has 100% more shaders than the 1060 as you correctly note, but it only has 33% more ROPs and 43% more bandwidth, which results in it only being 68% faster overall instead of 100%. In other words the 1080 is clearly ROP and/or bandwidth bottlenecked compared to the 1060.

By comparison Vega is currently rumoured to have 78% more shaders (4096 vs. 2304) along with 18% higher clocks (1500 MHz vs 1266 MHz), for a total of 110% more potential shading performance, this is then combined with 100% more bandwidth (512 GB/s vs 256 GB/s), and 100% more ROPs (64 vs 32). As such Vega shouldn't really be ROP or bandwidth bottlenecked compared to RX 480, since both of those areas increase proportionally with shading performance.

As such Vega looks to be a much more balanced upgrade over the 480 than the 1080 is over the 1060, and as such should get much better scaling. If we assume a simple 100% scaling (following the smaller increase found for ROPs and bandwidth), then Vega would be ~15% faster than 1080, and that's before we even take any architectural improvements into account, if we use the 10% improvement that you mentioned before, then it will end up 25-30% faster than a 1080 and more or less neck and neck with a 1080 TI.

Of course this is all speculation and AMD may somehow figure out a way to actually regress in performance for Vega compared to Polaris, who knows.

 

[crisium]

I like your logic. I already mentioned the ROPs and buswidth will be better this time around, but I really like your overall explanation. Agree with your points.

Btw, your quote right now is someone else's name with my text. Forum bug?

 

[antihelten]

I like your logic. I already mentioned the ROPs and buswidth will be better this time around, but I really like your overall explanation. Agree with your points.

Regarding the 1080 being ROP/bandwidth limited, it is also interesting to look at the 1080 Ti. The 1080 Ti only has 28% more shading performance on paper, but it has 37.5% more ROP and bandwidth throughput, and in real life performance it is upwards of 35% faster, further indicating that the 1080 is ROP/bandwidth bottlenecked.

Btw, your quote right now is someone else's name with my text. Forum bug?

Not a forum bug, just me being sloppy. Fixed now.

 

[zinfamous]

Regarding overclocking headroom this is the one thing that has disappointed me over successive generations. I know you can make the argument that
it doesn't matter but I find it exciting seeing how far I can push my new hardware. OC headroom was the key difference between 980ti and Fury X (apart from the obvious 4gb RAM) and while at stock they perform similarly as an enthusiast I want the fastest possible performance and enjoy spending time fiddling to get there. If Vega is better value per $ than the 1080ti or in the unlikely event it outperforms it then I might be tempted to trade in the 1080ti. Probably not since AMD have been under delivering in GPU terms for a couple of years and starving their dual GPU cards of RAM was another annoying practice that meant they would not outlast their single GPU equivalents. I think the silence from AMD is a good thing but gaining the performance crown does seem a bit beyond them at this stage. 5% Better than 1080ti OC for OC and I'm in just to go red again.

I would pee myself if that happened. Not that I would get one, because that kind of performance is out of my price range, lol. But it would simply be shocking, and hopefully be a sign of what kind of great performance I can get in a ~$400 Vega card.

Now, how about this: Big Vega = +5% 1080Ti stock, but at -10% OC/OC (on release--you know, expecting some AMD maturity to end up some year(s) later at near equal OC/OC on those same cards...well, maybe), and ~$100 cheaper? (lol--not sure if the math is possible on that +5 stock /-10 overall, but I'm just tossing junk out there)

Would that be tempting?

 

[Mockingbird]

If we are trying to predict Vegas performance compared to an NV competitor based on die size it's good to point out that last generation Fury X and 980Ti were both right at 600mm^2. In general they performed on par. Fury taking a slight lead at 4K and Vulcan, 980Ti taking the lead back when OC'd.

If Vega has a significantly larger die size than the 1080Ti/ Titan Xpp it's not unreasonable to assume it will have greater performance.

If that seems weird consider the reportedly similar specs to Fury, (4096 SPs) that previous GPUs from different generations with the same basic layout have had significantly different performances.

Going all the way back to the X1800XL vs X850XT, both cards had similar clocks speed and were 16 pipe cards yet the XL was between 10% and 50% faster

http://www.anandtech.com/show/1810/10

More recently the 7970/280x vs the 470. Both are 2048 shader parts but the 470 is about 35% faster. I'd expect Vega will remove any internal bottlenecks Fury had.

If Vega does have the larger die size I expect it to outperform the 1080Ti. It will still be less power efficient than the 1080Ti mostly because I expect it to include professional capabilities that NV dropped from Pascal because AMD can't afford separate mainstream and professional GPUs.

You have to remember that the Radeon R9 Fury/Fury X was made using TSMC's fabrication process.

AMD has obviously moved away from TSMC after the Radeon Rx 300 Series.

I don't think die sizes are particularly predictive of performances when the products are made using difference fabrication processes.

 

[Saylick]

I would pee myself if that happened. Not that I would get one, because that kind of performance is out of my price range, lol. But it would simply be shocking, and hopefully be a sign of what kind of great performance I can get in a ~$400 Vega card.

Now, how about this: Big Vega = +5% 1080Ti stock, but at -10% OC/OC (on release--you know, expecting some AMD maturity to end up some year(s) later at near equal OC/OC on those same cards...well, maybe), and ~$100 cheaper? (lol--not sure if the math is possible on that +5 stock /-10 overall, but I'm just tossing junk out there)

Would that be tempting?

That would be, but what would be more tempting was if Vega was 10% slower than 1080Ti in average FPS but was 5% or 10% better in minimum FPS. Would you go with this option?

 

[guachi]

I would go for that. It would be a better choice for a freesync setup and freesync is much cheaper than GSync.

And I already have a freesync monitor.

 

[keymaster151]

I've been assuming that Vega will actually be made on Samsung's LPC/LPU process instead of the LPP at Global Foundries. Part of this assumption is because I don't think the 14 nm LPP was as good as the 16 nm TSMC process, which has further support if you look back at the Apple SoCs that were made on both processes. Also, with Ryzen, AMD is going to be using a lot of wafers and Vega is a big die that is going to need a lot of wafers itself. I don't think GF has enough capacity to support all of AMDs products, especially since they'll be releasing their server CPUs and desktop APUs in the near future as well.

I've been thinking the same thing based on some news stories from last year. First there was this story which mentioned that AMD had already used Samsung's fabs to produce hardware.

https://www.extremetech.com/computi...samsung-could-tap-foundry-for-future-products

"AMD has strong foundry partnerships and our primary manufacturing partners are GLOBALFOUNDRIES and TSMC. We have run some product at Samsung and we have the option of enabling production with Samsung if needed as part of the strategic collaboration agreement they have with GLOBALFOUNDRIES to deliver 14nm FinFET process technology capacity."

The other interesting thing was AMD's and GlobalFoundries amendment to the WSA last year.

http://ir.amd.com/phoenix.zhtml?c=74093&p=irol-newsArticle&ID=2198716

A few key points from the press release are.

• Provides AMD with the flexibility to manufacture certain products with another wafer foundry;
• Make a $100 million cash payment to GF, paid in installments beginning in Q4 2016 through Q3 2017.
• Make quarterly payments to GF beginning in 2017 based on the volume of certain wafers purchased from another wafer foundry.

If I'm reading between the lines correctly, this means that AMD is paying GlobalFoundries $100 million for the ability to use other foundries for certain products. The fact that AMD also has to pay a certain amount to GF for wafers fabbed at other foundries leads me to believe that this option will only be used for products with high profit margins and where performance is paramount, such as high-end GPUs.

Samsung's 14nm LPU is supposed to offer increased performance over the LPP process, which could be important for efficiently reaching the higher clock speeds that Vega is supposedly designed for.

This could also open the door for the rumored Vega 20 die-shrink to be done with Samsung's 10nm process which obviously isn't as good as the upcoming GF 7nm process, but it does have the advantage of actually existing/working. I for one have doubts that GF's 7nm will be ready in 2018 H2 as they are promising.

All of this is of course pure speculation, since I'm neither and insider nor and expert on the subject, but it is interesting to think about.

 

[Paratus]

You have to remember that the Radeon R9 Fury/Fury X was made using TSMC's fabrication process.

AMD has obviously moved away from TSMC after the Radeon Rx 300 Series.

I don't think die sizes are particularly predictive of performances when the products are made using difference fabrication processes.

It's not die size so much, but transistors clock speed and xtor density.

I think the best way to estimate what Vega would need to do to equal Pascal would be:

Pascal Max Boost X #xtors = Polaris Max Boost X #xtors X %Vega Efficiency

(We'll use Polaris Max Boost clock as a stand in for Vegas max clock as the RX 580 has already been refined on the current process.)

Let's take a look at full Polaris and full Pascal and try and make some predictions for Vega. We'll look at xtors/mm^2 and clock speed to see how Polaris compares and how much Vegas architecture efficiency is going to have to catch up.

TPU puts the 1080TI /Titan Xpp at around 200% faster than the RX 580
https://www.techpowerup.com/gpudb/2938/radeon-rx-580

Let's look at their stats:

• GPU | mm^2 | # xtors | xtors/mm^2 | Max mhz
• Polaris | 232 | 5.7B | 24.6M/mm^2 |1340
• Pascal | 471 | 12.0B | 25.5M/mm^2 |1580

Based on TPU we could just say that all Vega needs to do is double the # of xtors from the 580 (11.4B, 464 mm^2) and bam we're tied with big Pascal. This would suggest Polaris is more architecturally efficient than Pascal.

That doesn't seem quite right especially since Pascal has an 18% clock speed advantage and even a 3.5% xtors density advantage.

It probably makes more sense to compare the RX 580 vs the GTX1060 6Gb to estimate architectural efficiency.

The 1060

• 1060 | 200mm^2 | 4.4B | 22M/mm^2 | 1709mhz

TPU puts it at 106% of the RX 580

The 1060 has 28% more clocks vs the 580s 30% more xtors. With it being 6% faster overall Pascal ends up around 4.4% more performance per combination of xtor(MHz)

(1060)1709mhz x 4.4B x 1.06 = 7970
(580) 1340mhz x 5.7Bx 1.00 = 7638
7970/7638 = 1.044

(This is actually an improvement. If you do the same calc for FuryX vs 980Ti the TI ends up 18% more performance per combination of xtor and MHz)

Now that we know the relative performance efficiency of Pascal over Polaris we can calculate the worst case die size for Titan Xpp equivalent Vega.

1580mhz X 12B xtors x 1.044 = 1340mhz X # of Vega xtors
# of Vega xtors = 14.8B

14.8B / 24.6M/ mm^2 = Vega Die Size: 602 mm^2

So if Vega is no better in clock speed, xtor density, and efficiency than Polaris then AMD will have to go right to the reticule limit to compete.

This also doesn't seem correct for a few reasons. First Vega is supposed be an improvement on Polaris, likely closing the gap or better between the architectures.

Second, Vega is AMDs second architecture on this process. If you look at previous GPUs AMD is very good at increasing xtor density on the same node:

• Tahiti - 11.8 M/mm^2
• Hawaii - 14.2
• Fury - 14.9

AMD increased xtor density by 20% their first rev on 28nm.

So if we assume Vega is equivalent to Pascal, has a mild 5% clock boost over Polaris, and a 20% density increase suddenly AMD only needs a die size equal to:

458mm^2

If assume the rumors of a 500mm^2 die are true suddenly Vega is 9-10% faster.