AMD Radeon Vega GPU Architecture

[FordGT90Concept]

As said previously, Vega has twice the number of shaders as RX 470 (2048), not RX 480 (2304).

 

[Captain_Tom]

As said previously, Vega has twice the number of shaders as RX 470 (2048), not RX 480 (2304).

As said previously, Vega is clocked significantly higher than the 480.

As said previously, Vega is 12.5 TF vs 5.8TF 480.

 

[FordGT90Concept]

Vega will be clocked lower. RX 480 reference clock is 1266 MHz. Fury X is clocked at 1050 MHz. 1200 MHz seems the most probable max clock.

AMD hasn't announced TFLOPs yet. That said, TFLOPS is for highly parallelized compute tasks, not gaming. Gaming is all about rushing completed frames out the door as fast as possible.

AMD has put this out though:

Which until we know the clocks, doesn't answer anything.

 

[Kanan]

Fury's problem was => bad OC.
At stock it was on par with 980Ti day one (and even more so later on, when drivers matured)

Fury X problem was that it used the frontend of R9 380X/380/285 which is Tonga a much smaller GPU with only half the shaders, the first GCN 3 GPU (Fiji is the 2nd) - that meant that the utilization of it's shaders was only high on 4K and on some games at 1440p too, but almost never at 1080p. RX 480 / Polaris nearly solved that problem, using a new frontend - but basically all AMD GPUs still suffer by a a certain degree from it (that's why they scaled so nice on higher resolutions), even HD 7970 - 7870 was much better utilized, also the reason why it was so fast, it was a damn good performance grade GPU for around 200-300 bucks. Hawaii (R9 290/390) is generally the best GPU AMD did up to this day, it was well balanced and fought against 2 gen (Pascal would be the third) of Nvidia GPUs without losing other than on efficiency.

The one who controls console market controls the developer's balls. It's what's keeping AMD very relevant. Otherwise, yes, it would be NVIDIA only who was fiddling with ones balls... XD

Yeah too bad that never really happened on the PC market. Not really using irony here, it's my honest opinion. But then again it's the natural thing to do, if Nvidia controls the market, and not AMD. Consoles != PCs. I think you PCMR guys would know that.

 

[medi01]

Higher than most gamers play at.

Most gamers play on 2 core CPUs, what the fuck does it have to do with anything?

 

[Super XP]

I would assume VEGA's core clock will be 1,500 MHz or more. Probably between 1500 to 1800 MHz IMO. Though I heard clocking the HBM2 speed greatly benefits overall performance as much as clocking the core clock. Can't wait to see this baby in action.

Most gamers play on 2 core CPUs, what the fuck does it have to do with anything?

Older games perhaps, New games min 4 cores. And Dev's are stepping that up to 6-8 cores with New upcoming games.

 

[FordGT90Concept]

I think 1000-1200 MHz is more probable. Remember, it is a big chip on GloFo's not-so-fantastic 14nm process.

 

[Kanan]

I think it will be at 1200-1400 MHz tops, 1500 MHz is a fantasy that's not even done for Polaris, a half sized chip. This, or they are producing Vega at TSMC and it's vastly better, or Vega has a vastly improved architecture that allows for high clocks with any production node. Three possible outcomes.

 

[Captain_Tom]

I think 1000-1200 MHz is more probable. Remember, it is a big chip on GloFo's not-so-fantastic 14nm process.

Imo the only reason AMD is waiting this long to release Vega is an attempt to get clockspeeds up to a point where they can win.

They already announced the server chip is ~1530 MHz. I would highly doubt the gaming chip is less than that. Overclocked 580's already hit 1500 MHz, there's no reason the top-binned Vega couldn't by June.


Furthermore, Vega seems a bit less dense in SP/mm than Polaris is. The die size of Vega is over double that of Polaris, and it has less than twice as many SP's. Logically this extra space would be provided to facilitate higher clocks.



P.S. 1000 - 1200 MHz is a joke! What are you smoking?! This isn't 40nm lol.

 

[Super XP]

I still stand by the high clock of about 1,500 MHz none boost speed.

I have a good feeling about VEGA. High clock speeds may benefit it a lot more over anything AMD released to date. But you never know.

This is one Architecture AMD will not mess up. They most likely want to Ride on Ryzen's success.

 

[Captain_Tom]

I still stand by the high clock of about 1,500 MHz none boost speed.

I have a good feeling about VEGA. High clock speeds may benefit it a lot more over anything AMD released to date. But you never know.

"You never know". Absolutely!

None of us know for sure. But common sense dictates it will at least be 1500 MHz, but I fully expect the cut-down version to be closer to 1400 MHz.

 

[Super XP]

"You never know". Absolutely!

None of us know for sure. But common sense dictates it will at least be 1500 MHz, but I fully expect the cut-down version to be closer to 1400 MHz.

Yes makes sense.. in around 1300-1400 MHz for the cut down version.

But that 1,500MHz speed will still have room for overclocking.

 

[Captain_Tom]

Yes makes sense.. in around 1300-1400 MHz for the cut down version.

But that 1,500MHz speed will still have room for overclocking.

I sure hope so. After playing around with Fiji, Polaris, and Pascal I feel like Overclocking is almost dead. The newest complex chips simply don't have any headroom.


I remember clocking my launch 7970 to 1215/1835 MHz lol! That sucker performed 35% stronger than stock, and traded blows with a Titan (Beats it now due to Kepler's horrific aging)! If I could get close to that kind of boost again, I would be so happy...

 

[Super XP]

I sure hope so. After playing around with Fiji, Polaris, and Pascal I feel like Overclocking is almost dead. The newest complex chips simply don't have any headroom.


I remember clocking my launch 7970 to 1215/1835 MHz lol! That sucker performed 35% stronger than stock, and traded blows with a Titan (Beats it now due to Kepler's horrific aging)! If I could get close to that kind of boost again, I would be so happy...

Yes the good old days. Lol

As much fun as OC'ing is, I just hope VEGA is the design that really shows off AMD's GPU architecture. Because FURYX failed to. Mingling new and old tech together was a bad AMD move.


L

 

[Captain_Tom]

Yes the good old days. Lol

As much fun as OC'ing is, I just hope VEGA is the design that really shows off AMD's GPU architecture. Because FURYX failed to. Mingling new and old tech together was a bad AMD move.


L

Yeah well I suspect AMD was caught off-guard on 2 fronts in that situation:

1) They didn't expect Maxwell in 2014. In fact Nvidia decided to port Maxwell to 28nm at the last minute.

2) AMD thought they would be able to get 20nm by late 2014, once they realized it wasn't happening it was too late.


As for Vega? LOL just look at the slides!!!!

-2 x Geometry IPC
-100% more effective usage of framebuffers (Likely translates to an even higher effective bandwidth).
-Even higher color compression (At this point it's at least 30% better color compression compared to Fiji).

^If those numbers are even mostly accurate, then a 12.5 TFLOP Vega card would be 2.5 - 3x more powerful than the RX 480 lol, and that's before we even consider that Vega supposedly is switching to tiled rendering (Maxwell was the first Nvidia arch to do that, and you remember how that turned out?).

So either Vega will be a monster, or AMD just straight up lied lol.

 

[FordGT90Concept]

"You never know". Absolutely!

None of us know for sure. But common sense dictates it will at least be 1500 MHz, but I fully expect the cut-down version to be closer to 1400 MHz.

Common sense dictates Polaris should have been 1500 MHz too but it launched at 1266 MHz. The most similar chip to Vega is Fiji which launched at 1050 MHz.

They already announced the server chip is ~1530 MHz.

Source?

Edit: working backwards from the numbers here you'd need 1525 MHz to get 12.5 TFLOP.

If those numbers are even mostly accurate, then a 12.5 TFLOP Vega card would be 2.5 - 3x more powerful than the RX 480...

Imma let you finish but...
-12.5 TFLOP is derived from the Vega Cube (quad Vega, ~50 TFLOP).
-Titan Xp is also 12 TFLOP.
-Before you get all excited about the above, RX 480 is 5.8 TFLOP (215%) while RX 580 is 6.2 TFLOP (202%).
-AMD takes about a 30% TFLOP hit when doing games compared to NVIDIA (RX 580 should match GTX 1070 according to TFLOPs but it doesn't).

 

[Xzibit]

Yeah well I suspect AMD was caught off-guard on 2 fronts in that situation:

1) They didn't expect Maxwell in 2014. In fact Nvidia decided to port Maxwell to 28nm at the last minute.

2) AMD thought they would be able to get 20nm by late 2014, once they realized it wasn't happening it was too late.

They should have Nvidia told everyone 2013 then it slipped to 2014

09.21.10 - The Tech Report - Nvidia prepping Kepler in 2011, Maxwell in 2013

The delay in process node hit both of them.

 

[Captain_Tom]

Common sense dictates Polaris should have been 1500 MHz too but it launched at 1266 MHz. The most similar chip to Vega is Fiji which launched at 1050 MHz.


Source?

Edit: working backwards from the numbers here you'd need 1525 MHz to get 12.5 TFLOP.


Imma let you finish but...
-12.5 TFLOP is derived from the Vega Cube (quad Vega, ~50 TFLOP).
-Titan Xp is also 12 TFLOP.
-Before you get all excited about the above, RX 480 is 5.8 TFLOP (215%) while RX 580 is 6.2 TFLOP (202%).
-AMD takes about a 30% TFLOP hit when doing games compared to NVIDIA (RX 580 should match GTX 1070 according to TFLOPs but it doesn't).

1) The 380X was clocked at 1030 MHz, and thus Fiji was clocked higher than it's lower-tier brethren. I can do this all day:

-7850 was clocked lower than the 7970.
-The 6800 series clocked almost exactly the same as the 6900 series
-7790 was clocked the same as the 290X

The fact is there is absolutely no consistent pattern of AMD clocking their higher-end cards lower than their smaller cards. None.

It is Nvidia that has a tendency to clock the top cards slower, not AMD.



2) "you'd need 1525 MHz to get 12.5 TFLOP." Yes.



3) "AMD takes about a 30% TFLOP hit when doing games compared to NVIDIA"

Nope, Polaris does. We don't know what Vega's FPS/flops ratio is yet. But we do know this is the biggest architectural shift AMD has made in 6 years, and we also know AMD is switching to tiled rendering like Nvidia did with Maxwell. Thus it is not crazy to think AMD will come close to Nvidia's TFLOPS efficiency, but again we don't know yet!

AMD's confirmed professional version of Vega cards have over double the stats of the 480, and let's see how that would stack up:

^Hmmm right next to a 1080 Ti, and that is before ANY architectural enhancements are considered.

 

[Fluffmeister]

Crazy how things progress, that Palit GameRock is 72% faster than Fury X!

 

[FordGT90Concept]

1) The 380X was clocked at 1030 MHz, and thus Fiji was clocked higher than it's lower-tier brethren.

I wasn't talking about "lower-tier." Fury X was to R9 390X as RX Vega is to RX 580. RX Vega's architecture is mostly a die shrink of Fury X.

2) "you'd need 1525 MHz to get 12.5 TFLOP." Yes.

I wouldn't use that as a metric for consumer cards. They basically just wanted to hold 100 TFLOPs in their hand and declare it so. He never actually proved it. On top of that, it wasn't packaged in a way that can be used in the computing industry at all. It was intended for automotive industry.

But we do know this is the biggest architectural shift AMD has made in 6 years...

That's what was said about Polaris and it turns out it behaves like GCN 1.2 with more stream processors.

...we also know AMD is switching to tiled rendering like Nvidia did with Maxwell.

"Tile-based rasterization," not rendering. Maxwell was a completely new architecture pretty much from the ground up. There's a lot more to its efficiency than just tile-based rasterization. GCN was an architecture designed for DirectX 11.1 and hasn't been significantly changed since it's introduction in 2012. Maxwell, on the other hand, was designed for DirectX 12, feature level 12_1.

AMD's confirmed professional version of Vega cards have over double the stats of the 480, and let's see how that would stack up:

Polaris 10 has 36 compute units while Vega 10 has 64 compute units. Unless you failed elementary school math, that is not "double."

 

[Captain_Tom]

I wasn't talking about "lower-tier." Fury X was to R9 390X as RX Vega is to RX 580. RX Vega's architecture is mostly a die shrink of Fury X.


I wouldn't use that as a metric for consumer cards. They basically just wanted to hold 100 TFLOPs in their hand and declare it so. He never actually proved it. On top of that, it wasn't packaged in a way that can be used in the computing industry at all. It was intended for automotive industry.


That's what was said about Polaris and it turns out it behaves like GCN 1.2 with more stream processors.


"Tile-based rasterization," not rendering. Maxwell was a completely new architecture pretty much from the ground up. There's a lot more to its efficiency than just tile-based rasterization. GCN was an architecture designed for DirectX 11.1 and hasn't been significantly changed since it's introduction in 2012. Maxwell, on the other hand, was designed for DirectX 12, feature level 12_1.


Polaris 10 has 36 compute units while Vega 10 has 64 compute units. Unless you failed elementary school math, that is not "double."

I wasn't talking about "lower-tier." Fury X was to R9 390X as RX Vega is to RX 580. RX Vega's architecture is mostly a die shrink of Fury X.


I wouldn't use that as a metric for consumer cards. They basically just wanted to hold 100 TFLOPs in their hand and declare it so. He never actually proved it. On top of that, it wasn't packaged in a way that can be used in the computing industry at all. It was intended for automotive industry.


That's what was said about Polaris and it turns out it behaves like GCN 1.2 with more stream processors.


"Tile-based rasterization," not rendering. Maxwell was a completely new architecture pretty much from the ground up. There's a lot more to its efficiency than just tile-based rasterization. GCN was an architecture designed for DirectX 11.1 and hasn't been significantly changed since it's introduction in 2012. Maxwell, on the other hand, was designed for DirectX 12, feature level 12_1.


Polaris 10 has 36 compute units while Vega 10 has 64 compute units. Unless you failed elementary school math, that is not "double."

5.8 TFLOPS vs 12.5 TFLOPS. You may want to go back to school kiddie lol. [Hint: It's over double because Vega is clocked higher]

 

[Captain_Tom]

Crazy how things progress, that Palit GameRock is 72% faster than Fury X!

Ehh. 2 years for not even double the performance?

I wish it was more...

 

[FordGT90Concept]

5.8 TFLOPS vs 12.5 TFLOPS.

*cough*

...then a 12.5 TFLOP Vega card would be 2.5 - 3x more powerful than the RX 480...

RX 480 is 5.8 TFLOP (215%) while RX 580 is 6.2 TFLOP (202%).

[Hint: It's over double because Vega is clocked higher]

It's over double because Raja Koduri claimed it was for an application that is 100% FP16 compute-based. For all we know, those were chips designed specifically for deep learning (geometric, SIP, and graphics command blocks removed) so they could push the clocks higher at lower voltage. AMD needs to do this to compete with NVIDIA DGX-1 (170 TFLOP FP16, 21.25 TFLOP per Pascal chip).

On the other hand, all of NVIDIA's recent GeForce cards have their FP16 performance seriously gimped so customers looking for deep learning hardware are forced to dish out big bucks for DGX-1. Vega's goal numero uno is to capitalize on the market that NVIDIA is raping right now.

 

[Fluffmeister]

Ehh. 2 years for not even double the performance?

I wish it was more...

Well yeah, as it stands AMD have brought fuck all to the table.

Now think of those dodgy CPUs.

 

[Super XP]

Yeah well I suspect AMD was caught off-guard on 2 fronts in that situation:

1) They didn't expect Maxwell in 2014. In fact Nvidia decided to port Maxwell to 28nm at the last minute.

2) AMD thought they would be able to get 20nm by late 2014, once they realized it wasn't happening it was too late.


As for Vega? LOL just look at the slides!!!!

-2 x Geometry IPC
-100% more effective usage of framebuffers (Likely translates to an even higher effective bandwidth).
-Even higher color compression (At this point it's at least 30% better color compression compared to Fiji).

^If those numbers are even mostly accurate, then a 12.5 TFLOP Vega card would be 2.5 - 3x more powerful than the RX 480 lol, and that's before we even consider that Vega supposedly is switching to tiled rendering (Maxwell was the first Nvidia arch to do that, and you remember how that turned out?).

So either Vega will be a monster, or AMD just straight up lied lol.

Here is what is quite interesting with VEGA IMO, though there's many other cool features. Well what we know now based on the AMD slides.
1) "Vega is designed to handle up to 11 polygons per clock with 4 geometry engines."
2) "AMD is promising a better than 2x or more improvement in peak geometry throughput per clock"
3) "AMD is also confirming, at a high level, that the Vega NCU is optimized for both higher clockspeeds and a higher IPC."
4) "AMD - A new shader stage (Called Primative Shader) that runs in place of the usual vertex and geometry shader path, the primitive shader allows for the high speed discarding of hidden/unnecessary primitives."

AMD has completely Re-Designed its GPU Architecture. Can't wait to see this thing in action.

Well yeah, as it stands AMD have brought fuck all to the table.

Now think of those dodgy CPUs.

The only dodgy CPU I can think of is the 7700K that has a habit of causing micro stuttering when PC gaming. Based on several reviews and the like.

ZEN - Ryzen on the other hand are superior products. AMD has done an amazing job. Thank you Jim Keller, one of the worlds greatest CPU Architects to date.