Vega/Navi Rumors (Updated)

[JDG1980]

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.

RX 480 has 2304 shaders at a maximum boost clock of 1266 MHz. That means 5.83 peak TFlops.
GTX 1060 has 1280 shaders at a maximum boost clock of 1708 MHz. That comes out to 4.37 peak TFlops.
Assuming both cards are reasonably well balanced in terms of ROPs, bandwidth, etc., this means that Polaris needs about 33% more TFlops to get the same performance as Pascal. Or, looked at the other way, a Pascal card with about 75% of the TFlops of Polaris can provide equal performance.

We have reason to believe that the top Vega SKU will do 12.5 TFlops or more. That was the figure given for the Radeon Instinct MI25 accelerator, and it would be unusual for a professional card to have higher clocks than the consumer equivalent - usually it's the other way around. So if we assume Vega is just a scaled-up Polaris, no architectural improvements at all, that means it would be roughly equivalent to a Pascal card with ~9.4 TFlops. Yes, that's only about 5% more than GTX 1080 (8.87 TFlops) - but keep in mind that assumes no architectural improvements. And that's not the case - there are some big improvements with Vega, most notably tiled rendering, which was what vaulted Maxwell past GCN in perf/TFlop in the first place.

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

The problem with that argument is that Pascal almost certainly won't have nearly as much of a clock speed edge over Vega as it has over Polaris. AMD specifically stated that Vega was optimized for higher clock speeds. As mentioned above, the Radeon Instinct MI25 is said to do 12.5 TFlops, which, assuming 4096 shaders, equates to a clock speed of 1525 MHz.

 

[beginner99]

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

Hard to say. Yes, they panicked and overpaid. Then due to lack of money R&D especially also process R&D got delayed and they slowly started to fall behind Intel. Still looking back AMD having to sell the fabs was inevitable. They where just to small of a company to support their own process R&D and fabs. And as Krumme said where would they be now? They would have needed to create their own GPU for iGPUs just like intel. That would not have been cheap as well.

One could also argue the biggest mistake was to sell of ATI mobile GPU division for 65 mio to qualcomm if we look at that market now.

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.

The fact that bandwidth and ROPs increase more than shaders just goes to show once more that RX 480 is bandwidth limited and instead of upping clocks AMD should have shipped the 580 with faster memory instead and lower power use.

 

[EnzoLT]

So TLDR, around which performance tier is Vega supposed to be at?

I am looking to buy a video card now for an X99 system and would either just wait it out for Vega or buy 1080TI right now.

 

[.vodka]

Truth is, no one knows. There's been too many positive changes in Vega as to even try to predict where performance will land. It's still GCN, but it's the first time in six years it's been tweaked and updated this much. GCN1-4 were mostly evolutionary changes... unlike Vega. Given the size of the known chip so far (4096sp, similar to Fiji/Fury X) I'd say GTX1080 at the very least, and that's a worst case scenario.

It'll launch soon, within Q2 (April, May, June). Two months to go, whenever AMD decides to launch. We know HEDT 16C32T X399 Ryzen will most likely launch in Computex, and that's May 30 - June 3. Maybe we get Vega too...

 

[AdamK47]

How many pages of this thread will we get to before Vega is actually released? I'm going to be optimistic and say 135.

 

[EnzoLT]

Wish we had some sort of "direction".

Sent from my SM-G930F using Tapatalk

 

[richaron]

How many pages of this thread will we get to before Vega is actually released? I'm going to be optimistic and say 135.

I'm betting it will be locked a few days before the actual release when (once again; looks at Ryzen rumor thread) it finally becomes obvious to mods that the OP has no intention of updating the OP with anything positive about an AMD product.

They can feel like they've done their job and OP has kept positive rumors off page 1 for ages. It's a win-win

 

[Krteq]

AMD shows off Vega's ability to handle 8K graphics at NAB

 

[Valantar]

Not a Vega rumor, but I take it this might interest a few people here:

AMD has quietly released a dual-polaris (2304SP x2) Radeon Pro Duo.

Now, I do not take this as a sign that a 480/580x2 will ever show up, as this makes far more sense for compute and VR development than gaming. But still interesting. Also very interesting that they're giving it the same name as the outgoing Fiji-based Pro Duo, which implies at least similar performance.

 

[Dribble]

So TLDR, around which performance tier is Vega supposed to be at?

I am looking to buy a video card now for an X99 system and would either just wait it out for Vega or buy 1080TI right now.

The consistent rumours say Vega performance is a 1080 + 10%. It's a 500mm2 die using 2 stacks of HBM2. So it'll be another fury basically (big, hot, expensive, quite fast). I think it'll start off competing with the o/c 1080's (not Ti) but after a year or two as DX12/Vulcan usage grows will be a bit faster (close to a 1080Ti).

 

[itsmydamnation]

The consistent rumours say Vega performance is a 1080 + 10%. It's a 500mm2 die using 2 stacks of HBM2. So it'll be another fury basically (big, hot, expensive, quite fast). I think it'll start off competing with the o/c 1080's (not Ti) but after a year or two as DX12/Vulcan usage grows will be a bit faster (close to a 1080Ti).

tTere have been no "consistent rumors" , just the same one or two people reaping something they cant backup over and over.

I would call that a negative rumor.

i would say a neutral rumor would be equal with GP102
i would say an optimistic position would be 15%+ ahead.

 

[beginner99]

This means 1 CU is a lot smaller than in hawaii or polaris and uses much less power. Where does the 4096 figure actually comes from? AFAIK this is just a rumor. Maybe it's actually more because they are smaller.

Quoting myself. I want to come back to this. Where does the 4096 gcn cores rumor comes from?

Polaris 10 has 2304 cores with a die size of 232mm2. That would make roughly 415mm2 for 4096 cores. Of course this is naive as it doesn't count for ROPs and memory controller. ROPs will use more in vega compared to P10, memory controller less. Still for what is the rest of the die space uses? Vega NCU is a lot smaller than a polaris CU due to variable SIMD width ALUs. Is that space saving all eaten up by more cache? Or what else?

 

[Valantar]

tTere have been no "consistent rumors" , just the same one or two people reaping something they cant backup over and over.

I would call that a negative rumor.

i would say a neutral rumor would be equal with GP102
i would say an optimistic position would be 15%+ ahead.

Actually, the "1080 + 10%" rumor is mostly based off AMDs SW:BF demo of a pre-production Vega card playing it at 4k. This has a lot of fundamental issues, though.

• SW:BF is a relatively Radeon-friendly title performance wise.
• They didn't tell/show us the specific graphics settings used.
• The demo was done with pre-production hardware and software, with no word on clocks or other specs.
• It's very reasonable to expect clocks to rise and drivers to improve significantly in the several months between that demo and launch, yet there's no way of knowing by how much.
• Not to mention that we don't know if drivers were mature enough to make use of any hardware innovations such as the HBCC or other more fundamental restructurings of the GCN core and the Vega NCU.

As such, the only thing we can really say is that AMD, roughly 6 months before launch, showed a pre-production Vega card slightly outperforming the 1080 in a single title at 4k with unspecified settings. Anything else is (still) unfounded.

Now, other rumors say some pre-production cards run at 1200MHz (corroborated by some online benchmark result logs), while the persistent idea is that the retail card will run at 1500MHz (no concrete corroboration of this yet). AMD has said that the Vega architecture is optimized for "higher clock speeds", which ought to be relative to Polaris 10/11, and as such should mean noticeably higher than the ~1350MHz that Polaris clocks to without catching fire. 1500 sounds reasonable here, but again: we don't know. We might be disappointed by a 1400MHz card, or they might pull another Ryzen-style overdelivery and ship a card clocking to 17-1800. There's no way of knowing.

Tl;dr: there are too many unknowns and variables in play here for us to know anything much. We need more info, more leaks, more of everything. Or a product launch, that would be even better

 

[antihelten]

The fact that bandwidth and ROPs increase more than shaders just goes to show once more that RX 480 is bandwidth limited and instead of upping clocks AMD should have shipped the 580 with faster memory instead and lower power use.

Whether RX 480 is bandwidth limited or shader performance limited isn't really relevant as far as Vega scaling goes though, since both areas will be improved by roughly 100% and as such we can probably assume roughly 100% scaling as well.

Now, other rumors say some pre-production cards run at 1200MHz (corroborated by some online benchmark result logs), while the persistent idea is that the retail card will run at 1500MHz (no concrete corroboration of this yet).

Actually there has been corroborating evidence of the 1500 MHz number, namely the Instinct MI25. For the MI25 to reach 25 GFLOPS with 4096 shaders it must be running at roughly 1500 MHz*.

Now technically it is of course possible that the consumer version of Vega is clocked lower than the professional version, but that would be highly unusual.

*Alternatively it could run at 1200 MHz, but then it would need to have 5120 shaders and I don't think there has ever been any indication of this.

 

[T1beriu]

Where does the 4096 gcn cores rumor comes from?

AMD engineer on LinkedIn and CompuBench.

 

[Harmaaviini]

I did a few rough estimtates using these charts
https://www.techpowerup.com/reviews/ASUS/GTX_1080_Ti_Strix_OC/29.html
https://www.techpowerup.com/reviews/ASUS/GTX_1080_Ti_Strix_OC/13.html

Doom&vulkan 4k:
Vega engineering sample 1.1x GTX 1080(50.2fps) =~ 55fps
55fps/fury(46.5fps) =~1.18

If we assume overall performance uplift of Vega vs Fury to be that same 1.18 then looking at the performance summary chart:

1.18x Fury =~ 65
With 1.2x clock scaling from engineering sample to final product-> ~78

Which is 10% faster than GTX 1080. But what about the driver then? 15% gain should be quite manageable over a pre-production driver I think. 1.15x 78 =~90. That would be ~27% above GTX 1080. Interesting consideration though: Is 1.18xFury really a lowball estimate for a general uplift because Fury is allready pretty efficient with Vulkan API? So I wonder what the gains could be against Fury's weakneses... All in all I find the 10% gain "rumor" to be quite unbelievable.

 

[Glo.]

I did a few rough estimtates using these charts
https://www.techpowerup.com/reviews/ASUS/GTX_1080_Ti_Strix_OC/29.html
https://www.techpowerup.com/reviews/ASUS/GTX_1080_Ti_Strix_OC/13.html

Doom&vulkan 4k:
Vega engineering sample 1.1x GTX 1080(50.2fps) =~ 55fps
55fps/fury(46.5fps) =~1.18

If we assume overall performance uplift of Vega vs Fury to be that same 1.18 then looking at the performance summary chart:

1.18x Fury =~ 65
With 1.2x clock scaling from engineering sample to final product-> ~78

Which is 10% faster than GTX 1080. But what about the driver then? 15% gain should be quite manageable over a pre-production driver I think. 1.15x 78 =~90. That would be ~27% above GTX 1080. Interesting consideration though: Is 1.18xFury really a lowball estimate for a general uplift because Fury is allready pretty efficient with Vulkan API? So I wonder what the gains could be against Fury's weakneses... All in all I find the 10% gain "rumor" to be quite unbelievable.

AMD demoed Vega flagship GPU in Doom 4K.

And they have used the same settings as Techpowerup uses, however on different platform, tho.
It can be seen in LinusTechTips video:

TSAA times 8. TechpowerUp uses standard TSAA setting.

We can see it is 687F:C1 so it has 1.2 GHz and air cooling. Water cooled GPU will have 1.5 GHz core clocks.

Draw your own conclusions on this. This is almost best case scenario of performance for AMD GPUs in case of software. However it still does not use most important feature: Primitive Shaders. There is hell of a lot of performance sleeping in the GPU, and will be extracted in future.

P.S. I will absolutely not be staggered if AMD will decide to sell RX Vega with air cooling and 1.2 GHz for 499$, and the limited edition with water cooling and 1.5 GHz for 699$.

Just my 2 cents on this whole situation.

 

[antihelten]

P.S. I will absolutely not be staggered if AMD will decide to sell RX Vega with air cooling and 1.2 GHz for 499$, and the limited edition with water cooling and 1.5 GHz for 699$.

Just my 2 cents on this whole situation.

That would be pretty nice. Instinct MI25 runs at 1.5 GHz passively cooled (obviously servers racks have fairly beefy case fans, so effectively it will be about equal to a an air cooled solution), so an RX Vega with air cooling at 1.2 GHz should have plenty of OC headroom (at least 25%) and for $500 it needs to be at least as fast as the 1080, and arguably a bit faster to be competitive (given Nvidia's better name recognition).

So if it's a bit faster than a 1080 (say 5% on average) and overclocks by 25%, then basically said $500 would match a stock 1080 Ti when overclocked.

 

[Glo.]

That would be pretty nice. Instinct MI25 runs at 1.5 GHz passively cooled (obviously servers racks have fairly beefy case fans, so effectively it will be about equal to a an air cooled solution), so an RX Vega with air cooling at 1.2 GHz should have plenty of OC headroom (at least 25%) and for $500 it needs to be at least as fast as the 1080, and arguably a bit faster to be competitive (given Nvidia's better name recognition).

So if it's a bit faster than a 1080 (say 5% on average) and overclocks by 25%, then basically said $500 would match a stock 1080 Ti when overclocked.

In the demo's by AMD the 687F:C1 GPU, which according to leak has 1.2 GHz averages around 71-72 FPS. In the same settings as TechpowerUp Uses. And this is exactly GTX 1080 Ti level. So make what you want out of this.

P.S. The GPU test was done on Ryzen+Vega platform, on completely not ready drivers, and software, as we know. So end results can be much better .

One more thing. Vega architecture is designed to deal with very high resolution content. 8K@60 Hz. That is why I have said before: nobody has correctly guessed Vega arch. performance so far.

 

[Valantar]

Actually there has been corroborating evidence of the 1500 MHz number, namely the Instinct MI25. For the MI25 to reach 25 GFLOPS with 4096 shaders it must be running at roughly 1500 MHz.

That is assuming zero IPC improvement from Fiji (8.6Tflops@1050MHz, so 8600/1050 = 8.19 Mflops/MHz, or in other words 12500/8.19= 1.526GHz required), which ... well, AMD has stated that there will be. Which is part of why this is all so up in the air right now.

Most enterprise/workstation/server parts are clocked lower than their consumer counterparts (for longevity), so what if the MI25 instead might be clocked at 1200MHz (which due to ES leaks we have an inkling is a possible stable clock for Vega)? That would mean a (12500/1200 = 10.42 Mflops/MHz) 27% increase in IPC. Given the information we have, I don't see how this is any less likely than the scenario you are drawing up.

 

[Valantar]

P.S. The GPU test was done on Ryzen+Vega platform, on completely not ready drivers, and software, as we know. So end results can be much better .

If you're talking about the Doom demo in the video above, that's clearly labeled at the end of the video as a custom-built i7-4790K system. No Ryzen in sight.

 

[Glo.]

If you're talking about the Doom demo in the video above, that's clearly labeled at the end of the video as a custom-built i7-4790K system. No Ryzen in sight.

LinusTechTips video. Thats where is Ryzen+Vega setup .

P.S. Framerate is consistent on both platforms. Intel, and Ryzen. Both are averaging around 70 FPS in 4K Doom, Ultra Settings, Vulkan.

 

[exquisitechar]

That is assuming zero IPC improvement from Fiji (8.6Tflops@1050MHz, so 8600/1050 = 8.19 Mflops/MHz, or in other words 12500/8.19= 1.526GHz required), which ... well, AMD has stated that there will be. Which is part of why this is all so up in the air right now.

FLOPS are independent from "IPC", aren't they?

 

[Paratus]

My predictions for a 500mm^2 Vega based on my mental gyrations above:

Assumptions:
Low end - no better than a big Polaris
High End

• No of Transistors - 20% increase in density from Polaris based on historical improvement at 28nm from Tahiti to Hawaii
• Clock Speed - 1500 max boost based on rumors
• Efficiency - +5% architectural efficiencies so Vega is equivalent to Pascal
• 1080P to 4K: Assume a Vega 4% relative improvement to GTX 1080 going from 1080P to 4K based on Furys performance (Titan Xpp assume 22% based on TPU)

My Predicted Range of Vega Stats:

• No of Transistors: 12.28B - 14.74B
• Max Boost Clock Speed: 1340mhz - 1500mhz
• Relative Efficiency to Pascal: -4.4% to +0.6%

My Predidicted Relative Ranking of Low and High Vega at 1080p (4K) to the GTX 1080 and Titan Xpp:

1. Vega High - 178 %(182%)
2. Titan Xpp - 151% (173%)
3. Vega Low - 126%(130%)
4. GTX 1080 - 100%(100%)

 

[Valantar]

FLOPS are independent from "IPC", aren't they?

Eh... no. IPC = instructions per clock. Any operation (such as a FLoating point OPeration, or FLOP) performed by a GPU means processing one or more instructions. As there are many different types of instructions this isn't something directly quantifiable, but is rather used as a general term meaning roughly "how quickly can a given number of cores/compute units finish a given task at a given clock speed." It quite directly lends itself to comparisons between architectures and components doing similar tasks.

As such, "IPC improvements" can mean both across-the-board improvements and improvements to certain types of instructions/instruction sets, but is generally assumed to mean across-the-board due to the other use being quite misleading.