AMD Radeon Vega GPU Architecture

[Kanan]

This is could be true, but it wouldn't explain why AMD's Die-size / SP is so much larger than before. Spacing out SP's almost always leads to higher sustainable clocks (Less density = less concentrated heat).

That, or it's because of Vega's new future tech. Remember, Vega isn't simply beefed up Polaris, it's a new architecture with new features.

 

[Captain_Tom]

That, or it's because of Vega's new future tech. Remember, Vega isn't simply beefed up Polaris, it's a new architecture with new features.

That's exactly what I am saying buddy

 

[Kanan]

That's exactly what I am saying buddy

A lot of people talk down Vega as a RX 470 x 2 - it's much more than that. And if not ... well, it's rare that I'm wrong but even a 2x RX 470 wouldn't be exactly "bad".

 

[Captain_Tom]

A lot of people talk down Vega as a RX 470 x 2 - it's much more than that. And if not ... well, it's rare that I'm wrong but even a 2x RX 470 wouldn't be exactly "bad".

It really is this simple: 2.25x the TFLOPS of the 480, 2,2x the bandwidth of the 480 (Due to memory compression), and an entirely new architecture.


The 1080 Ti is a tad over twice as strong as the 480, so overall Vega should compete with the 1080 Ti at a minimum. The massive architectural enhancements should give it a decent chance of stomping the 1080 Ti, but we will see.

At the very least it would have NEGATIVE IPC compared to Polaris if it doesn't match the 1080 Ti.

 

[Kanan]

It really is this simple: 2.25x the TFLOPS of the 480, 2,2x the bandwidth of the 480 (Due to memory compression), and an entirely new architecture.


The 1080 Ti is a tad over twice as strong as the 480, so overall Vega should compete with the 1080 Ti at a minimum. The massive architectural enhancements should give it a decent chance of stomping the 1080 Ti, but we will see.

At the very least it would have NEGATIVE IPC compared to Polaris if it doesn't match the 1080 Ti.

That's just tflops math, and quite useless in real world. In reality 2x 480 isn't even as fast as 1x 1080 (non-TI), don't forget Crossfire isn't perfectly scaling either, so even with tflops math it wouldn't be that simple.

That said the 1080 Ti easily stomps 2x 480 without a hitch. Even overclocked RX 480s have no chance. So to say the RX Vega will have 1080 Ti performance minimum is just a dream - that's the maximum it will have at DX11 apps. And I expect it to be faster in Vulcan and DX12, but not always either. Worst case: it's equal in Vulcan/DX12, or slower.

No negative IPC either, it will be faster than 2x 480, I expect performance between 1080 and 1080 Ti levels, and I'm very sure on that too. Nvidia knows the game, and if Vega would've been as strong as you think it will, they would've blown out the full GP102 chip and not again a cut down one. This already happened in the past when R9 290X was too fast and Nvidia was forced to push out full GK110 at 2880 shaders, even more than the overly expensive Titan has. This time Nvidia isn't doing this, and it's another hint at the performance Vega will probably have.

 

[Captain_Tom]

That's just tflops math, and quite useless in real world. In reality 2x 480 isn't even as fast as 1x 1080 (non-TI), don't forget Crossfire isn't perfectly scaling either, so even with tflops math it wouldn't be that simple.

Wrong:

https://www.techpowerup.com/reviews/ASUS/GTX_1080_Ti_Strix_OC/29.html


^Multiply the 480's performance summary by 2, and HELLO it beats the 1080 Ti lol.


And yeah TFLOP's between similar families of GPU's is an entirely valid comparison to make.

 

[FordGT90Concept]

Very few games get 100% scaling from multi-GPU. Most games don't support multi-GPU at all.

 

[Captain_Tom]

Very few games get 100% scaling from multi-GPU. Most games don't support multi-GPU at all.

Why are we talking about crossfire?! LOL Vega has over double the specs of the 480 without being a double GPU.

 

[Kanan]

Why are we talking about crossfire?! LOL Vega has over double the specs of the 480 without being a double GPU.

That's wrong too, Vega will have 4096 shaders, that's double RX 470 and not double RX 480. So, double RX 470 + architectural improvements = between 1080 and 1080 Ti speed, probably. Maybe the core clocks play a part in that game too, I suspect Vega to be clocked higher than RX 470 at release.

And btw. you can stop the silly 1+1 math, I know these reviews and more myself. It's much much more complicated and you will never understand it properly if you continue like this.

 

[Captain_Tom]

That's wrong too, Vega will have 4096 shaders, that's double RX 470 and not double RX 480. So, double RX 470 + architectural improvements = between 1080 and 1080 Ti speed, probably. Maybe the core clocks play a part in that game too, I suspect Vega to be clocked higher than RX 470 at release.

Alright buddy now you are just showing your ignorance lol. Read a blood article or press release!!!

Vega is 12.5 TFLOPS, 480 is 5.8. This is because Vega is clocked substantially higher than the 480. Now go to bed, I am sure you have school tomorrow.

 

[Kanan]

Alright buddy now you are just showing your ignorance lol. Read a blood article or press release!!!

Vega is 12.5 TFLOPS, 480 is 5.8. This is because Vega is clocked substantially higher than the 480. Now go to bed, I am sure you have school tomorrow.

The ignorant one calls me ignorant, I'm just laughing right now. To quote myself and counter your bullshit:

That's just tflops math, and quite useless in real world. In reality 2x 480 isn't even as fast as 1x 1080 (non-TI), don't forget Crossfire isn't perfectly scaling either, so even with tflops math it wouldn't be that simple.

I'm pretty sure the kid here is you, you're already behaving like a silly fanboy all the time.

Now, you can repeat your nonsense and silly 1+1 math again and again, it won't help your point, you're still wrong.

 

[FordGT90Concept]

Why are we talking about crossfire?! LOL Vega has over double the specs of the 480 without being a double GPU.

That didn't work out the best for Fury X, sadly. Unless AMD fixed the problems of underutilization in Vega, it's not going to compete with GTX 1080 Ti.

This is because Vega is clocked substantially higher than the 480.

Wishful thinking. RX 480 is a small chip and small chips tend to hold higher clocks than large chips. 1000 MHz is likely, 1200 MHz is best-case-scenario.

 

[Captain_Tom]

That didn't work out the best for Fury X, sadly. Unless AMD fixed the problems of underutilization in Vega, it's not going to compete with GTX 1080 Ti.


Wishful thinking. RX 480 is a small chip and small chips tend to hold higher clocks than large chips. 1000 MHz is likely, 1200 MHz is best-case-scenario.

This is a different architecture with more spaced out SP's and a more mature 14nm process. How about we cut to the chase and you can just tell me "You were right" later - sound good?


Either way AMD has released materials showing 12.5 TFLOPs. They have to get to that number somehow, so if you think it's by having a 6000 SP card @ 1000MHz - more power to you buddy!

 

[Kanan]

Tflops are calculated with a formula. http://optimisationcpugpu-hpc.blogspot.de/2012/10/how-to-calculate-flops-of-gpu.html

The reason why Nvidia GPUs even with less Tflops are faster than AMD GPUs is simply because they are way better utilized, a problem AMD has since at least the HD 5000 line of GPUs. That's also why I'm saying to not stress the Tflops numbers too much, they aren't conclusive for real world performance.

That said, RX line of GPUs made some improvements over the latest R9 products when it comes to utilization, but it's still not good enough. Vega will probably have better utilization than Fury but I don't expect any wonders, so Nvidia will most probably still be faster by a good amount, unless the game can utilize Vega entirely, for instance Doom.

 

[medi01]

Repeating myself:


So, shader count ("cuda cores") / chip size in mm2
1060 - 1280 / 200
1070 - 1920 / 314 (harvested 1080)
1080 - 2560 / 314
1080Ti - 3584 /471 (harvested Titan, although, number of shaders is the same) => 2.8 times 1060 shaders,, 2.35 times 1060 chip size

RX 480 - 2304 / 232
RX Vega - 4096 / 500 (?) => 1.8 times 480 shaders, 2.15 times 480 chip size

TFLOP/FPS gap with Nvidia

tflops/fps gap is hardly meaningful, it's like tflop/mhz gap.

The real gaps are perf/watt and perf/die size, the things mentioned above are merely architectural differences: you can decide to go with slower clock but denser chip (AMD normally does) or aim for higher clock (nvidia normally does).

 

[Kanan]

The real gaps are perf/watt and perf/die size, the things mentioned above are merely architectural differences: you can decide to go with slower clock but denser chip (AMD normally does) or aim for higher clock (nvidia normally does).

The funny thing is, Nvidia is simply doing both (density / high shader count + high clocks) at the same time - the last time I looked Titan X + 1080 Ti were still able to hit very high clocks. AMD really needs to fix their architectural problems. I want them to be at the very top again.

 

[CookieMonsta]

Wrong:

https://www.techpowerup.com/reviews/ASUS/GTX_1080_Ti_Strix_OC/29.html


^Multiply the 480's performance summary by 2, and HELLO it beats the 1080 Ti lol.


And yeah TFLOP's between similar families of GPU's is an entirely valid comparison to make.

Not really, even on a monolithic chip, simply doubling the resources rarely results in a perfect doubling of performance.
Just look at the 1060 vs the 1080. Once you normalize the clockspeed gap, you will find the 1080 lags behind dual 1060s (i'm talking theoretical doubling not SLI). There are internal bottlenecks that restrict perfect scaling.

So the only unknown/s that would allow Vega to completely surpass the 1080ti is if theres some secret-sauce type IPC improvements (e.g. NVIDIA's tile-based rendering when Maxwell was released) or if it scales clockspeed harder than GP102.

 

[medi01]

The funny thing is, Nvidia is simply doing both (density / high shader count + high clocks) at the same time - the last time I looked Titan X + 1080 Ti were still able to hit very high clocks.

1060 is able to hit "very high clock" doesn't help it that much vs 480.
1070 (a bigger chip, although hard to say how much bigger) is 1.43 times faster than 480, while running at 1.34-1.32 higher clock.

480 is denser than 1080. (5.7/230 vs 7.2/314)

We also have Glofo 14nm vs TSMC 16nm, I'm still confused which one is better.

Titan is denser than 480 (12/470) although not by much.
We see how that compares to AMDs high end in May.

So the only unknown/s that would allow Vega to completely surpass the 1080ti is if theres some secret-sauce type IPC improvements

There is die size parity, always been. It is can swing back n forth for 10-20%, but not much more than that.
500mm2 Vega (ok, Raja said "it is less") can not not compete with 472mm Titan.
10% slower? Possibly. (I rather expect perf to vary wildly from game to game) But closer to Titan, than 1080.

That didn't work out the best for Fury X, sadly. Unless AMD fixed the problems of underutilization

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)

 

[FordGT90Concept]

So the only unknown/s that would allow Vega to completely surpass the 1080ti is if theres some secret-sauce type IPC improvements (e.g. NVIDIA's tile-based rendering when Maxwell was released) or if it scales clockspeed harder than GP102.

And that might be the key. Xbox One and Xbox 360 utilized a tile-based rasterizer but Vega will be the first card by AMD to bring the technology to the consumer space. It could give Vega a boost similar to Maxwell:
https://www.pcper.com/reviews/Graph...w-Redesigned-Memory-Architecture/Primitive-Sh

Vega appears to focus on areas GCN has struggled in the past (tile-based rasterization and cache misses) so maybe it will be improved in a big way. Then again, maybe not. A lot of it depends on developers actually using it (e.g. packed math and culling invisible) and history tells us NVIDIA has developers by the balls.

Fury's problem was => bad OC.

The problem was that Fiji was literally as big as the interposer would allow and even then, it only tended to beat NVIDIA at the highest resolutions where all those shaders would be put to work.

 

[RejZoR]

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

 

[medi01]

it only tended to beat NVIDIA at the highest resolutions

1440p is not "highest resolutions".

as big as the interposer would allow

Give me a break.

 

[FordGT90Concept]

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

I don't think that actually matters. Developers largely develop games on NVIDIA hardware. The difference is that they optimize console versions because if they don't, it won't pass validation. Those optimizations rarely effect the PC side because, again, they're fundamentally programmed for NVIDIA. Case in point, Unreal Engine 4 still uses PhysX even though the consoles can't accelerate it (except Switch). Developers use PhysX acceleration on PC though because that's the platform they're working on.

And, there's a reason for this too. GTX 660 is still the bare minimum for game developers to target because it is the most common card when surveying gamers.

1440p is not "highest resolutions".

Higher than most gamers play at.

 

[RejZoR]

They use PhysX because someone cocked up Havok along the way...

 

[FordGT90Concept]

Havok is likely going to find its way into the DirectX package soon. DirectPhysics anyone?

Anyway, that was just one example. AMD has to pay developers to use OpenCL and TressFX, sadly.

 

[Captain_Tom]

Not really, even on a monolithic chip, simply doubling the resources rarely results in a perfect doubling of performance.
Just look at the 1060 vs the 1080. Once you normalize the clockspeed gap, you will find the 1080 lags behind dual 1060s (i'm talking theoretical doubling not SLI). There are internal bottlenecks that restrict perfect scaling.

So the only unknown/s that would allow Vega to completely surpass the 1080ti is if theres some secret-sauce type IPC improvements (e.g. NVIDIA's tile-based rendering when Maxwell was released) or if it scales clockspeed harder than GP102.

Not true at all. The 1080 may have roughly double the shader power, but it only has 67% more bandwidth. ( 1.67 + 2 ) / 2 = 83% more performance should be expected. Looking at benchmarks it seems to perform around 73% better in 4K, not too far from linear scaling.


No I don't expect perfectly linear scaling from the 480, but getting double performance when all the stats are over double isn't an insane prospect. Again, this is before we throw in the massive architectural enhancements.