Radeon Vega Architecture Thread Of Shame

[Qwertilot]

This is the collected nVidia discussion from the Radeon Vega Architecture Preview thread. The rule is simple but you just can't help yourselves. Shame on you all.

AT Moderator ElFenix

I don't see how a 500+mm^2 Vega doesn't easily beat a 1080 and at least compete with a Titan XP.

Ok, there are plenty of reasons why it might not do that For starters look at recent history: Fury vs the 980 - definitely not a blow out win. Polaris was notably further behind NV's small chips in terms of perf/watt and mm2 than they were in the Maxwell era, so they'll need some genuinely non trivial improvements over Polaris to even repeat that.

They've also got a lot of compute stuff in vega, and doing that has non trivial associated die size/power draw disadvantages. (See P100 vs P102.).

Of course that compute stuff may well end up being the reason that Vega makes sense - in gaming terms it is going to have serious problems unless it is way ahead of the year(+) old 1070/80 etc. In DP compute terms it will be! Behind P100 perhaps but that's a monster/very expensive card so scope to compete.

 

[CatMerc]

Had a thought...

Titan XP is 72% faster than a Fury X.
Fury X is clocked at 1050MHz.
Professional Vega is supposed to clock at around 1550MHz, so about 50% higher clockspeed.

That would mean that a Vega GPU with Fury X performance per clock would already account for having 87% the performance of a Titan XP, sitting between 1080 and Titan XP.
It literally wouldn't make sense for Vega to NOT be stronger than a 1080 unless performance per clock dropped since the Fury X, and that's a card with terrible perf/clock due to the amount of bottlenecks everywhere.
Then consider that professional clockspeeds tend to not be as high as consumer clockspeeds, and it's quite possible Vega would reach Titan XP performance by just having a Fury X clocking to 1700MHz.

And that's completely ignoring the massive reworks they mentioned in the preview, which should improve IPC by their own words. I doubt they would mention IPC if it dropped.
Does it make any sense what so ever for big Vega to NOT beat Titan XP?


Here are the possibilities IMO:
* Drivers are very VERY early, and are pulling less performance per clock than out of Fiji
* This being the first revision (first working silicon is just a few weeks old), it might not be hitting clockspeed targets right now
* Both of the above

 

[IEC]

This seems to be a lot of similar changes to when nVidia gained huge amounts of efficiency and throughput with the 9xx series. Hopefully this means that Vega will be competitive with at least the 1080, if not with future unreleased nVidia products.

 

[PhonakV30]

from Source

Device ID = 687f:C1
Driver 1.4.1

techbuyersguru is saying it's slower than our bench, Now this one from GameGpu :

Which one ?

 

[raghu78]

With Vega 10 now mentioned as supporting 1/16 DPFP I am not able to understand how a 4096 sp chip is almost 550 sq mm even without 1/2 FP rate support. Fiji with 4096 sp was 600 sq mm on 28nm. So in effect AMD has used up an entire node shrink of 50% area reduction and still are on the same sp count. To top it all they are using only 2 HBM2 stacks on Vega 10 instead of 4 HBM1 stacks on Fiji. So they are saving die space on the memory controller too. If AMD have not been able to increase perf/sp at the same clock in a big way then its a huge waste of die space and in fact a regression in perf/sq mm. Vega 10 is still restricted to 4 geometry engines though its capable of a throughput of 11 polygons/clock . Its still not known if thats just driver side updates which can exploit the increased geometry throughput or game side updates. If its game side updates then its a huge problem in current games. Future Scorpio games would definitely use the primitive shaders for increased geometry processing but current games would still be limited .AMD has not given enough information yet about the NCU and how IPC has been increased. This chip could end up either way. It could turn out a disaster which regresses on even Fiji in terms of perf / sq mm or turnout to be a gem like 9700 Pro or HD 4870. I hope its not a disaster.

Nvidia has been able to roughly double 980 Ti performance with GP102. The current Titan X is 70-75% faster and still not a full GP102. A full GP102 with slightly higher clocks can pretty much provide 2x the performance of 980 Ti. Nvidia has also improved perf/sq mm with GP102 which is just 471 sq mm compared to 980 Ti's 600 sq mm.So a roughly doubling of performance at well less than 2x the die size (since 16FF+ provides a 50% die shrink compared to 28nm) If AMD cannot match a full GP102 then they have regressed in terms of perf / sq mm over Fiji . I think Vega will either disappoint badly or surprise wildly with its perf, perf/sq mm and perf/watt. Here is hoping its the latter for the benefit of the consumers.

 

[Samwell]

First it's not 550 mm², it's a bit under 500 mm² as it seems. (Straight from Raja Koduri to the editor of golem.de in another forum). 2nd point is, we don't get a real 2x shrink with 14nm. Just look at Polaris. Transistor density goes up to 24,6 Mio/mm². Vega 10 might be at 25, but Fiji is at 14,9 already. So a straight shrink would get you a ~350 mm² die. Then you have 2xFP16 and 4xFP8 in Vega. Both takes space. Then this new memory controller which has access to hbm, but also to other types of ram won't be so small. That's mostly features for HPC space. The L2 probably also got bigger for helping deferred shading. Add the new front end features and it's quite possible to spend 150mm². A gpu is not just shaders.

Why does it even need more shaders? You see what's possible with less shaders looking at Nvidia, so wait for results before complaining about it. They are adressing major bottlenecks which they had since GCN1. I'm sure we get at least 50% more speed than fury which puts it between GP104 and GP102. Maybe it'll be even nearer to GP102, which would be great for a mixed gaming/computing chip against a gaming chip GP102.

 

[RussianSensation]

Raghu,

Many hypothesized that AMD's very high transistor/mm2 chip density impeded the GPUs from achieving high clock speeds.

There are 5 key areas you need to consider.

The first is that not all of the building blocks of the GPU scale linearly with a node shrink since not all types of transistors scale the same. From GTX480->GTX580, we have learned that not the same types of transistors make up the entire GPU die. Moving from Hawaii to Fiji resulted in a shader and TMU increase of 45%, but the number of ROPs stayed the same despite a 36% increase in die size.

The second is that transistor speed/frequency may negatively impacts density (read the first paragraph under the SRAM graph in the ExtremeTech article below). Packing more transistors into tinier areas causes hot spot formation, which increases GPU temperature and power usage, while hampering max sustained clock speeds. It's easily possible that the move from 1.05Ghz clocks on Fiji to 1.5-1.55Ghz clocks on Vega 10 would require a hit to AMD's transistor density and thus die size. Fiji also had what barely 5-8% overclocking headroom? We don't yet know if Vega will suffer the same fate.

The third is that higher die size (I.e., transistor count) does not always correlate linearly with more performance. That is because some transistors are spent on non-gaming performance logic. Transistors are often used on power gating, and other building block functions of the GPU (such as media engine, 8-bit and 16-bit floating point operations) that don't impact PC gaming performance.

https://www.extremetech.com/computi...unts-theyre-a-terrible-way-of-comparing-chips

If Vega comes in using 225-240W of power, that's a significant reduction to 280W of Fury X:
https://www.techpowerup.com/reviews/Zotac/GeForce_GTX_1080_Amp_Extreme/27.html

I personally don't care if the GPUs use even 700-800W (have owned R9 295X2+R9 390 TriFire), a lot more people are obsessed about power usage which may have forced AMD to move away from 280-300W flagship chips (and for me this is a huge step for the desktop since it means possibly 15-25% of extra performance would be left on the table to save $2-3 a month in electricity costs).

The fourth is that if AMD is designing a NGCN architecture on a relatively new for them 14nm/16nm node. Since Polaris 10 is not a NCU design, the team which designed Vega is unlikely to have benefited from many lessons learned on the Polaris design. The architectural changes between Polaris 10 and Vega 10 are far greater than between Maxwell and Pascal. We should not expect AMD's best NCU design out of the gate. We should rather expect 2nd or 3rd iteration of that design to maximize transistors/mm2 efficiency (for example when AMD shifts Vega to 7nm).

The fifth point is controversial and my own opinion. Looking at Steam users, and previous successes (or lack thereof) of ATI's/AMD's 9800XT, X800XT/X850XT, X1800XT, X1950XTX, HD4890, HD5870, HD6970, HD7970/Ghz, R9 290X, Fury X, how much more evidence do we need that the competitor's flagship will always outsell ATI's/AMD's flagship, no matter what! If you were AMD, how much resources and effort would you put into Vega 10 for the sake of flagship sales? Not much. The reason AMD MUST design chips like Vega is to iterate them as next-generation mid-range or use as a foundation for future designs. They need a fast card for workstation, neural network, deep learning, professional use this generation. They can't wait every 2-3 years.

AMD cannot be naive to think that 70-80% of high-end users won't buy their cards as long as they lose performance by even 5%. AMD isn't stupid to see that a lot of high-end buyers are already in the G-Sync eco-system. AMD already knows that they will have missed out on a whole year of high-end sales with consumers buying 1080s and Titan XPs. AMD isn't naive to think 1080Ti won't launch by March-April. When all of that is considered, the cost of achieving flagship crown performance at all costs (300W flagship) is probably not even worth it. The market simply isn't there. It's amazing AMD even keeps bothering since their cards were far more popular during HD4870/5870/6970 days. Almost no one bought the Fury X, a chip that beats a stock 980Ti at 1440p/4K:
https://www.techpowerup.com/reviews/Zotac/GeForce_GTX_1080_Amp_Extreme/29.html

Fury X is one of the most hated cards in the market but recall FX5800U/5950U, 6800U, 7900GTX, GTX680, GTX780/780Ti. Those cards lost to ATI/AMD's flagships and still sold as well or outsold AMD. Notice the disproportionate hate/sales penalty for AMD's flagship vs. the competitor's?

The point of this last rant is context. AMD has lost context during the last generation. They cannot price a flagship at the same price as the competitor and expect it to sell well since most high-end consumers buy the competitor, while objective gamers found 980Ti's bonus VRAM and 20-25% overclocking headroom superior. These factors weren't in place during HD7970/R9 290X generations, which is why these same objective customers bought those cards.

Vega 10 can be successful even if it's slower than the competition as long as it's priced correctly.

*** As noted above, Vega 10 is not 550mm2.

 

[CatMerc]

While shrinking Fiji would not necesssarily result in a smaller die, just doubling up Polaris would. Hence they better be utilizing that extra area for great effect, otherwise they would be better off just creating a 4096 shader Polaris.

As for transistor densities:
RX 480 - 24.56m/mm^2
GTX 1060 - 22m/mm^2
GTX 1080 - 22.92/mm^2
Titan XP - 25.477/mm^2
GTX 1050 Ti - 25m/mm^2 (14nm is slightly denser)

Density rises with chip size, with the exception of 1050 Ti being based on the denser 14nm node. AMD's density isn't THAT much higher, and even accounting for it, a straight doubling of Polaris would still probably do better. And in fact, with using the same process NVIDIA got higher clockspeeds with HIGHER density. Using 16nm might have made the RX 480 less dense than the 1060.

I don't know what they're doing there, but at the clock speeds it's supposed to run even a Fury X would do better than what was demoed. This better either be silicon that can't hit the targets yet, bugged silicon, low performance early drivers, or all of the above.

 

[CatMerc]

I have watched the PCWorld interview with Raja. One thing apparent about Vega is this: it is designed for the future. Tile-Based Rasterization will be immediately apparent as improvement. However, Primitive Shaders, and Geometry Pipeline are designed for the future. Doom for example has not been updated for this feature, and Vega does not benefit from it whatsoever.

Designing for the future is fine as long as the NOW is competitive. Less than TXP performance a year later isn't exactly competitive.

 

[PhonakV30]

Designing for the future is fine as long as the NOW is competitive. Less than TXP performance a year later isn't exactly competitive.

I don't think It's good strategy.this can hurt AMD more , because It's opposite Day one Ready.

 

[flopper]

Designing for the future is fine as long as the NOW is competitive. Less than TXP performance a year later isn't exactly competitive.

have you had anything that indicate it isnt competitive?
Beats the 1080 in battlefront so, whats that about, doomsday prophecy?
when they asked Raja he confirmed it beats the 1080 easily in Vulkan and doom.
so not sure how people reason.

I am all in for a Ryzen update and Vega 2017, going 4K screen later when HDR is out in mass.

 

[Glo.]

have you had anything that indicate it isnt competitive?
Beats the 1080 in battlefront so, whats that about, doomsday prophecy?
when they asked Raja he confirmed it beats the 1080 easily in Vulkan and doom.
so not sure how people reason.

I am all in for a Ryzen update and Vega 2017, going 4K screen later when HDR is out in mass.

Hold your horses, before you will go to far .

Vega is possibly most advanced GPU architecture to this day. BUT we have to wait and see how it performs. I am leaving open mind about it, because I know they do not have drivers ready, and most of hardware features have to be applied to the game, to be used. And we have to wait for DX12 to be updated to use some of the hardware features of the GPU arch.

Where Vega will excell is APU. Im in for 95W 4C/8T CPU with 16CU iGPU, and 4/8 GB of HBM2 with 512 GB/s.

Also the performance of the GPU in Doom is true. Make no mistake here. It will not be faster than Titan X. Until... the drivers will be ready, and the game will be updated to use the hardware features. Then it most likely will be much faster and much more efficient.

 

[RussianSensation]

Designing for the future is fine as long as the NOW is competitive. Less than TXP performance a year later isn't exactly competitive.

But TXP costs $1200. What's wrong with a $650 card that offers 87-90% of TXP performance? I get that TXP isn't realistically "worth" $1200 using historical metrics (i.e, $349 GTX570, $499 GTX580 or even full blown $699 GTX780Ti), but those days are done for. We should just accept that it's not worthwhile or cost effective for AMD to go for the performance crown at all costs. NV lives and breathes high-end GPUs, but this level of GPUs isn't as important for AMD's turn-around strategy. AMD would make 2-4X the profits selling $199-$499 Zen than selling GPUs. Even during AMD's best GPU years, they barely made $. We are lucky they are still even trying.

They have to design their GPUs to last like FineWine since AMD doesn't have resources to do a new GPU architecture every 2 years.

Look at $699 780Ti losing to AMD's $399 R9 290 from the same generation, and R9 290X = R9 390X ties all-new Maxwell 980!
https://www.computerbase.de/thema/grafikkarte/rangliste/#diagramm-performancerating-1920-1080

At 1440p, 390X beats 980 and smashes 780Ti by 25% now (wow!):

https://www.computerbase.de/thema/grafikkarte/rangliste/#diagramm-performancerating-2560-1440

That means just 1 AMD architecture and chip designed before 2013 knocked out both the 2013 780Ti flagship and 2014 Maxwell 980!

Think about that long and hard.

Fury X is 5% faster at 4K than 980Ti:
https://www.computerbase.de/thema/grafikkarte/rangliste/#diagramm-performancerating-3840-2160

From 2012, every single AMD GPU series has outlasted NV, no exceptions (the only thing that saved 980Ti was overclocking). The problem is this hasn't translated into sales success since the types of gamers who buy $600-700 cards seem to change them every 10-18 months and are also extremely loyal to AMD's competitor.

There is no question that if buying used or long-term, AMD makes better products than NV when comparing the same generations since 2012. If Fury X was 8GB and had 20-25% overclocking headroom, 980Ti would have been irrelevant.

When AMD designed Vega 10, no doubt they designed it as a foundation for Navi to compete with Volta. When NV designed 780Ti, 980Ti and 1080Ti, they could care less if those cards last more than 2 years. It's why 780Ti and 980Ti have no real DX12 or Asynchronous Compute future-proofing capability built into them.

Think about how hard it must be for AMD's engineers to design an architecture for 3-4 years when NV probably had one team working on Pascal for 2016-2017, and then another on Volta for 2018-2019.

The issue with AMD's approach is that you absolutely cannot maximize perf/watt or perf/mm2 if you start adding more and more "complex future-proofing fat" that requires specific developer optimizations or next gen games to take advantage of. You end up with a GPU that stays underutilized for 1-1.5 years like 7970Ghz or 290X that later ends up absolutely demolishing its 680 and 780Ti competitors in years 2-4.

I don't think It's good strategy.this can hurt AMD more , because It's opposite Day one Ready.

But then you would risk alienating loyal AMD consumers who stay away from the competitor because its GPUs age like rotten food. The minute AMD has something worth buying I know I am dumping my 1070s since keeping them is like a ticking time-bomb for next gen games. AMD can't have it both ways since they don't have $$ to design an all-new architecture every 2 years (and the design probably takes 3-4 years, which means you need at least 2 separate large design teams working on parallel GPU architectures).

Think about this, if AMD's GPUs didn't have Fine Wine technology, many AMD buyers who have gamed on HD7850->R9 390X wouldn't buy future AMD cards. Seeing RX480 outperforming the 1060 keeps reinforcing this theme.

If Vega flops, AMD won't even take a dent (just look at Fury X) as Zen would more than make-up for it. If Volta flops, NV's revenue and profits could drop 30-40%. GPU design success matters way more to NV than for AMD. NV's entire business is built around GPUs, AMD's isn't and people keep forgetting this. It would actually be flat out embrassing and shocking if AMD matched and beat Intel in CPUs and NV in GPUs. It should not actually happen under any circumstances given AMD's resources and node manufacturing partners. So if Vega 10 doesn't beat $1200 TXP, it's still not a failure.

 

[CatMerc]

If Vega and TXP released within similar times then sure. But Vega will come near a full year later, that's a lot of time for extra work on the GPU. If they don't catch up to NVIDIA Maxwell on steroids after a deep change in uArch that they CANNOT do every generation, and a year of extra time later, how will they compete with NVIDIA's deep uArch change with Vega's Pascal like iteration - Navi?

Trust me, I am well aware that in the grand scheme of things Zen is what matters and this is why I'm not worried about AMD's existence. I am however worried about NVIDIA becoming the only real option with GPU's.

 

[CatMerc]

It's clear AMD is targeting higher clocks with Vega. Which will likely increase the area needed per CU compared to Polaris, which is very densely packed.

NVIDIA's density is on par or higher than AMD's. The density isn't the problem, it's the pipeline.

 

[sirmo]

NVIDIA's density is on par or higher than AMD's.

I don't know that this is true. I think Nvidia cards just have less features. Like less fp64, no ACEs, no command processor, no TrueAudio.. Also it's clear AMD GPUs have more stream processors than competing Nvidia cards.

 

[CatMerc]

I don't know that this is true. I think Nvidia cards just have less features. Like less fp64, no ACEs, no command processor, no TrueAudio.. Also it's clear AMD GPUs have more stream processors than competing Nvidia cards.

RX 480 - 24.56m/mm^2
GTX 1060 - 22m/mm^2
GTX 1080 - 22.92/mm^2
Titan XP - 25.477/mm^2
GTX 1050 Ti - 25m/mm^2 (14nm is slightly denser)

Remember that 14nm is denser than 16nm, so accounting for that 1060 and 1080 would probably be denser than 480 on 14nm.

 

[sirmo]

RX 480 - 24.56m/mm^2
GTX 1060 - 22m/mm^2
GTX 1080 - 22.92/mm^2
Titan XP - 25.477/mm^2
GTX 1050 Ti - 25m/mm^2 (14nm is slightly denser)

Remember that 14nm is denser than 16nm, so accounting for that 1060 and 1080 would probably be denser than 480 on 14nm.

As I said Polaris has more transistors that aren't part of the compute units. We are talking about compute unit density here.

1050 Ti and Polaris cards are built on the same 14nm process. Compare the number of CUDA cores to the number of Stream Procesors on the AMD side, say rx460 which is actually a cut down chip.

1050Ti has 768 cuda cores - 132mm2
unlocked rx460 has 1024 SPs. - 123mm2

Tflop per area is also interesting. rx460 (unlocked) is 2.5Tflop, while 1050ti has 2 tflops, this is at 1300Mhz which the boost maintains.

 

[flopper]

If Vega and TXP released within similar times then sure. But Vega will come near a full year later, that's a lot of time for extra work on the GPU. If they don't catch up to NVIDIA Maxwell on steroids after a deep change in uArch that they CANNOT do every generation, and a year of extra time later, how will they compete with NVIDIA's deep uArch change with Vega's Pascal like iteration - Navi?

Trust me, I am well aware that in the grand scheme of things Zen is what matters and this is why I'm not worried about AMD's existence. I am however worried about NVIDIA becoming the only real option with GPU's.

Elon Musk said it best, it was a real close call when he choose Nvidia for his Tesla cars.
Not worried about GPU Radeon, they are happy and confident about it already.
That tells me they have done things right and no bigger stuff is in the way for its Vega starpower.

Its the card anyone wants to own soon.

 

[RussianSensation]

If Vega and TXP released within similar times then sure. But Vega will come near a full year later, that's a lot of time for extra work on the GPU. If they don't catch up to NVIDIA Maxwell on steroids after a deep change in uArch that they CANNOT do every generation, and a year of extra time later, how will they compete with NVIDIA's deep uArch change with Vega's Pascal like iteration - Navi?

Trust me, I am well aware that in the grand scheme of things Zen is what matters and this is why I'm not worried about AMD's existence. I am however worried about NVIDIA becoming the only real option with GPU's.

You mean the only real option at the ultra high-end? Like I said already, 95% of PC gamers were better off during HD2900-HD6900 eras. Back then we had fierce price competition which resulted in $250 8800GT, $300 HD4870, $260-370 HD5850-5870, $250-300 HD6950 that unlocked into a 6970. Even if the top AMD card wasn't the fastest, we could buy 80-90% of flagship NV/AMD performance in the $250-300 range. Heck, even 980Ti/Titan XP tier GTX470/570 cost $349.

So really, the ONLY consumers who enjoy today's GPU market are those who love buying expensive PC parts for bragging rights, those whose only/primary hobby is PC gaming or the top 10% of income earners. Everyone else lost. From a consumer's point of view, we should want as much performance as possible for as little $ as possible.

Sorry, but AMD cannot be held responsible on this one when consumers made the situation what it is. Along the way, I bought HD4890, HD6950, HD7970, R9 390, R9 295X2. To this day we have the same people complaining that AMD makes nothing worthwhile but they didn't buy a single AMD/ATI product in 5-20 years! It wasn't the sales of flagship cards the buried AMD's dGPU division's chances. It was all the bottom feeding $100-300 inferior products from the competitor that vastly outsold AMD every generation since Radeon 8500 no less. The only time AMD had a fighting chance was when NV flopped with Fermi and took 6-9 months to roll those out. The fact that GeForcr 5 and 7 turds were so popular is telling enough about consumer consensus. As I keep repeating, during those eras, ATI had superior IQ and performance.

Really, consumers got 100% what they deserved in the end. And now AMD stopped the price/performance strategy. I don't blame them. If you want the best, no doubt NV will have it in the $700-1200 range. If you are in that target market, then I would not even wait for Vega. I would just buy TXP SLI and enjoy it.

Think about the risk factor for AMD. At any point the competitor can release 10-20% faster $349 1170 and $499 1180 and drop $649-799 1080Ti, and release a fully unlocked 3840 CC TXP Black for $1200.

AMD needs Vega for future mid-range cards, new console designs wins, foundation for next gen APUs/Navi. AMD doesn't care if Vega 10 beats 1080Ti/TXP since the sales they would get are irrelevant to their financials. They would be better off refreshing RX460/470/480 instead. But AMD can't keep reusing the same designs forever so eventually they are forced to do something new like Vega.

If AMD released an $800-1200 Vega in 2016, outside of few hardcore gamers, almost no one would have purchased it. When R9 295X2 cost $650 and wiped the floor with a $550 980, no one bought it. If 295X2 was an NV card and 980 was an AMD card, 980 would have bombed in such a scenario.

That's why AMD has to wait almost a year for costs of HBM2 and 450-500mm2 die to fall so they can sell Vega 10 derivatives at $449-699 at most. Even at $550, the HD7970/R9 290X weren't great sellers.

What we are seeing from AMD is a natural course of action any firm would take at this point after having a track record of past events -- focus on what's most important and disregard anything that's a waste of time and resources for minimal potential gains. The cost-benefit analysis for Vega likely dictated it wasn't worth releasing in 2016. You can imagine that over the last 2-3 years as many resources as possible were also allocated towards Zen since that's where the $$ for AMD is. That would also explain how RTG had no contengicy plan for a GDDR5X 350-400mm2 Vega card.

I am disappointed that AMD didn't focus on smaller Vega 11 for late 2016 launch. Seems they must have really been strapped for time and resources. Considering how close Fury X is to the 1070, they should have been able to release a card as fast as the 1070 for $399-479 with GDDR5X.

 

[sirmo]

AMD needs Vega for future mid-range cards, new console designs wins, foundation for next gen APUs/Navi. AMD doesn't care if Vega 10 beats 1080Ti/TXP since the sales they would get are irrelevant to their financials. They would be better off refreshing RX460/470/480 instead. But AMD can't keep reusing the same designs forever so eventually they are forced to do something new like Vega.

I am disappointed that AMD didn't focus on smaller Vega 11 for late 2016 launch. Seems they must have really been strapped for time and resources. Considering how close Fury X is to the 1070, they should have been able to release a card as fast as the 1070 for $399-479 with GDDR5X.

Spot on. AMD is not rushing high end because they are making it for OEM design wins, and ip blocks for semi-custom. People don't really buy AMD high end cards anyways. The entire Fury line is less than 0.10% on Steam HW survey. And you can't tell me they were bad cards. Heck Nano had no competition, literally in terms of the form factor.

Consumers totally asked for this, they voted for a monopoly with their wallets. I mean even in mainstream, 1060 is outselling rx480 2:1. And rx480 is a better card imo, when you factor in FreeSync it's not even close. Also can't really complain about AMD's product release cycle cadence. A smaller company than Intel and Nvidia and yet.. look at their product stack this year. Potentially one of the fastest CPU archs in the world (Zen), high end Vega (two brand new GPU archs in less than 12 months). Custom SoCs, for Scorpio, potentially (unannounced) PS5.. SSG GPUs.. Instinct stack.. K12 perhaps?

 

[RussianSensation]

https://www.youtube.com/watch?v=Y8tDaPLHxiE

If Vega 10 engineering sample was demoed with that crappy blower, and knowing AMD learned from HD7970Ghz, it reaffirms the slides showing 225W TDP.

https://www.pcper.com/news/Graphics-Cards/Leak-AMD-Vega-10-and-Vega-20-Information-Leaked

Looks like AMD is backing off 280-300W TDP designs of R9 290X/390X/Fury X. As I hypothesized earlier, if they target 225-240W power usage, they are potentially losing 10-20% of the performance compared to a 280-300W AIO CLC Vega 10 "Fury X" successor. As I said, they are going more for balance not the performance crown. Maybe AMD may have a separate $100 premium AIO CLC SKU marketed against the $100 crappy FE blower.

We should have more realistic expectations of performance if the TDP is only 225W. 45-50% faster at 4K over Fury X in a 225W TDP would be good for $599-649.

For reference, Zotac AMP! Extreme 1080 peaks at 245W, and is 43% faster than Fury X at 4K:

https://www.techpowerup.com/reviews/Zotac/GeForce_GTX_1080_Amp_Extreme/27.html

Inevitably when people start comparing AIB 1080s to Vega 10, there will be a tendency to site 1080's stock 184W power usage and AIB 1080 performance numbers.....

If AMD matches AIB GP104 in perf/watt, that would be a huge accomplishment given the gap during the R9 290X/390X and Fury X generations.

What's everyone's opinion? I am going to start with 45-50% faster than Fury X at 4K for $599-649 as I noted above. Would this be disappointing to you guys?

 

[Glo.]

You guys are looking in the wrong direction for clues. RX 480 is the mark where you have to look, for performance, because it is the newest architecture.

For example, 3072 GCN core chip just on core count will be 30% faster, making it to GTX 1070 level of performance. Add core clocks, and GPU advancements, higher bandwidth, and you are already within GTX 1080 range of performance.

I will be extremely happy if the Vega 11 will be priced at 499$ level.

 

[Glo.]

If this is true, big Vega GPU can have lower power consumption than some aftermarket GTX 1080's.

 

[.vodka]

........

We should have more realistic expectations of performance if the TDP is only 225W. 45-50% faster at 4K over Fury X in a 225W TDP would be good for $599-649.

......

What's everyone's opinion? I am going to start with 45-50% faster than Fury X at 4K for $599-649 as I noted above. Would this be disappointing to you guys?

Considering the amount of bottlenecks that GCN has since its inception that have been resolved on Vega, 45-50% uplift vs Fury X at high resolutions from these tweaks and higher clocks down to a 225w TDP range should be the minimum expected, I think. It's too big of an architecture jump in the right direction for it to be faster by less than that, even when power constrained vs Fiji.

On the lower TDP, that's great. For Vega to shine and make a good impression, it also has to overclock well. Leave the extra 75w TDP to the user's desire. The Fury X although equaling a 980Ti or exceeding it stock vs stock, can't do anything when the 980Ti can OC >25%.

GCN apart from Tahiti and Pitcairn back in 2012 has not been a good overclocker... Vega has been made for higher clock speeds and higher performance per clock, so it's a big win.

I don't know if this was posted here before.



Only one PCI-E 6+2-pin

150w + 75w from the slot = 225w TDP... which is the same as shown in the leaked vega slides and roadmap from videocardz.

Those do seem like valid leaks.