For reference, that was completely unclear in your original 2-word post. If that was your intention, it might've been nice to explain in a little more detail what you were asking for. Personally, I'd read a wall of text about how inefficient Vega is compared with modern NVIDIA chips.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
The Vega RX Thread (Rumors and Info)
- Thread starter fendersrule
- Start date
- Sort by reaction score
For reference, that was completely unclear in your original 2-word post. If that was your intention, it might've been nice to explain in a little more detail what you were asking for. Personally, I'd read a wall of text about how inefficient Vega is compared with modern NVIDIA chips.
It is because AMD GCN is compute oriented GPU, whereas Nvidia CUDA architecture is Geometry throughput oriented architecture, at least it was pre GP100/GV100.
If we talk about code execution we have to get to bases. AMD GCN has shorter pipeline of execution, which not allows the architecture to clock very high, but is able to do more work per clock cycle, because it has 64 KB instructions(wavefront), whereas each Warp in CUDA architecture has 32 KB size.
Each CU in GCN is made from 4, 16-Wide SIMD's which accounts for total 64 cores, in each CU.
Pre-Vega GPUs of never were graphics throughput masters - this is why weak efficiency of the architecture in games, and much stronger compute throughput of the GPUs - great efficiency in compute.
Both architectures work with 256 KB Register File Sizes, but differ in core counts that has access to those 256 KB RFS.
Nvidia Kepler had 192 cores/ 256 KB RFS.
Nvidia Maxwell Had 128 Cores/256 KB RFS.
Nvidia Consumer Pascal had 128 Cores/256 KB RFS.
GP100 has 64 cores/256 KB RFS.
GV100 has 64 cores/256 RFS.
Why we have seen increase both in gaming and compute performance per clock and per core with Maxwell vs Kepler? Because the cores in Maxwell were less starved for resources.
Why we haven't seen increase in performance with Pascal consumer GPUs per clock and per core? Because it is essentially Maxwell architecture on 16 nm process, with higher core clocks.
Why we see increase in performance with Pascal GP100 per clock and per core? Because those core are less starved for resources compared to consumer Pascal.
On a side note, Nvidia will reuse this architecture(actually - they already did with GV100) in consumer GPUs, so also expect per clock and per core increase in performance, and efficiency.
How did it looked like with GCN?
GCN1 - 64 cores/256 KB Register File Size
GCN2 - 64 cores/256 KB RFS
GCN3 - 64 cores/256 KB RFS
GCN4 - 64 cores/256 KB RFS.
Compute performance per clock were always dead flat, compared to previous generations of GCN, per clock, and per core.
What AMD lacked always was software optimization, and underutilization of its GPUs, both on compute side(if you optimize your codes and boarders, for not conflicting when you are operating on 32 KB Warp, and 256 KB RFS per 128 cores, on an architecture where you have 64 KB Wavefront, 256 KB Register File Size per 64 cores - this is what you will get - underutilization of the latter architecture). It can be partially blamed on lazy developers, partially not, because actually it requires quite a lot of optimization of your code to get things done on GCN properly. But if you will get it, expect two times higher performance on for example Hawaii(R9 390X, 2816 GCN cores) GPU compared to GTX 980 Ti(2816 CUDA Cores). This is why I have called many, many times on this forum to you, professional: demand optimization of software for all IEMs.
This is compute side.
What AMD GCN always lacked is Geometry throughput, hence its bad efficiency in games. GCN was able to register 1 Triangle per clock per Shader Engine. If you have had 1 GHz GPU this meant your GPU was able to register 1 mln triangles per clock, with each cycle, per shader engine.
Nvidia on the other hand with Kepler was able to register the same amount of triangles, but per SM(compute cluster). So not only they were able to clock higher, but also Nvidia GPUs had more SM's in their layout on high level.
With Maxwell Nvidia added, however, Tile Based Rasterization, which resulted in massively increased efficiency. Tile Based Rasterization removes the need for moving the data strictly from memory to the GPU each time it has to be refreshed, because needed data are stored in ROP caches, and L2 cache. So all of graphics important parts in the execution pipeline(Polymorh Engine, Pixel Engines, etc) were connected directly to L2 caches, not Memory controller. L2 cache then was connected to memory controller.
GCN on the other hand always L2 cache, and other important parts of the graphics pipeline(Pixel Engine, Geometry Engine, etc) were connected to Memory controller, rather than L2 cache. This increased inefficiency, because GCN always in situation with unoptimized software caused stalls of the pipeline, and underutilization of the Compute Units. To be fair, Asynchronous Compute is actually one way to deal with it, but its like trying to fix sinking Titanic with a bandage.
GCN4 resolved quite a lot of problems however with GCN in geometry throughput, but it had few design flaws(unbalanced design in the first place). In the end, Polaris in Geometry Throughput is exactly on par with consumer Pascal GPUs and Maxwell GPUs, but it outpaces it in Compute throughput per clock and per core. My company did a little experiment.
We have tried to declock GTX 1050 Ti with 768 CUDA cores to 855 MHz(actually we have been able to declock it to 900 MHz only, but close enough), the same level of core clock Radeon Pro 555 has, with the same core count(768 GCN4 cores, with 855 MHz). End result? In Nvidia optimized games we have seen 3 FPS difference for Nvidia GPU. AMD GPu was still faster than Nvidia in compute.
As for Vega GPU. It has new execution pipeline, which is longer, thats why it allows to clock it to 1.6 GHz. It has Tile Based Rasterization, which should increase efficiency of the memory bandwidth, and less stalls in the piepline(increased utilization). Vega has better load balancing, but what AMD is doing is they are dynamically declock, or overclock different Shader Engines to manually balance the load on short length shaders, and long length shaders. Pretty neat feature, which helps balancing the load even more. To even more balance the load, and help fully utilize the GPU is Infinity Fabric, which underneath everything connects every part of the GPU, together. And on top of everything - GCN has finally geometry based features connected to L2 cache, not memory controller, which will help balance the load even more, and should fix its efficiency.
AMD GCN on the other side, always had "some form" of Tile Based Rasterization. Each Shader Engine was working on its own part of display/screen. If you have had for example 4 Shader Engines, display was partitioned into 4 pieces, and each Shader Engine was working on its own. However, differences in complexity of the parts of the screen ended up in smaller load on some Shader engines, and bigger on the others - creating unbalanced pipeline, and lowering the performance of the GPU, because it created stalls. Previous Generations of GCN, as I have said, were able to register only 1 triangle per clock.
One last thing, to put on top of Vega architecture.
One of ways AMD could icnrease throughput of the GPU was by enabling higher amount of L2 cache, available to the cores. It would affect both: Compute and graphics.
Here is compute comparison per clock Vega versus Fiji(Fury X):
And here is comparison of Vega with Fury X in most AMD friendly games:
Now, the question is: Is AMD incompetent?
Simplest answer - no.
Previously, there was a little argument on this forum between few people on this topic. Some people believed that Vega outpaces Titan Xp in compute just because it is Geometry based benchmark(SpecPerf). If that would be the case, then Vega should be smoking over Titan Xp in games. Wiping the floor with it. Flushing it in the toilet. And yet, in games, per clock the architecture is slower, than Fiji. At least that is how it appears on the first glance.
What is the problem, then? Unfortunately, I do not have an answer. Most probable is that software is not ready: both drivers, and applications(Applications have to be rewritten to use for Example Primitive Shaders, which increases throughput of the GPU on geometry front). I do not believe that AMD is incompetent. Vega has almost the same features as Pascal, and Volta, but it still will have 64 KB Wavefront, so it will do more work per clock(slightly...) cycle, even if finally Nvidia paired their architecture layout in compute throughput with Vega(64 cores/256 KB RFS).
As a side note. In perfect world, in GCN optimized games, Vega 10 should be around 40% faster than Fiji XT was, per clock, per core, in gaming scenarios. It has around 15% higher core throughput, and much, much more robust graphics pipeline, which should end in that level performance. Second note, just increasing the clock on Vega, without the new features, from 1050 MHz on Fiji, to 1.6 GHz on Vega should make it on par with GTX 1080 Ti/Titan Xp. Something is bottlenecking the architecture. But on hardware level - there is absolutely nothing that could cause this. Absolutely NOTHING. It has the same features, the same layout of graphics pipeline as Nvidia GPUs has. If Nvidia can have great gaming performance, Vega also should.
If you have got to the end of this post, congratulations.
If we talk about code execution we have to get to bases. AMD GCN has shorter pipeline of execution, which not allows the architecture to clock very high, but is able to do more work per clock cycle, because it has 64 KB instructions(wavefront), whereas each Warp in CUDA architecture has 32 KB size.
Each CU in GCN is made from 4, 16-Wide SIMD's which accounts for total 64 cores, in each CU.
Pre-Vega GPUs of never were graphics throughput masters - this is why weak efficiency of the architecture in games, and much stronger compute throughput of the GPUs - great efficiency in compute.
Both architectures work with 256 KB Register File Sizes, but differ in core counts that has access to those 256 KB RFS.
Nvidia Kepler had 192 cores/ 256 KB RFS.
Nvidia Maxwell Had 128 Cores/256 KB RFS.
Nvidia Consumer Pascal had 128 Cores/256 KB RFS.
GP100 has 64 cores/256 KB RFS.
GV100 has 64 cores/256 RFS.
Why we have seen increase both in gaming and compute performance per clock and per core with Maxwell vs Kepler? Because the cores in Maxwell were less starved for resources.
Why we haven't seen increase in performance with Pascal consumer GPUs per clock and per core? Because it is essentially Maxwell architecture on 16 nm process, with higher core clocks.
Why we see increase in performance with Pascal GP100 per clock and per core? Because those core are less starved for resources compared to consumer Pascal.
On a side note, Nvidia will reuse this architecture(actually - they already did with GV100) in consumer GPUs, so also expect per clock and per core increase in performance, and efficiency.
How did it looked like with GCN?
GCN1 - 64 cores/256 KB Register File Size
GCN2 - 64 cores/256 KB RFS
GCN3 - 64 cores/256 KB RFS
GCN4 - 64 cores/256 KB RFS.
Compute performance per clock were always dead flat, compared to previous generations of GCN, per clock, and per core.
What AMD lacked always was software optimization, and underutilization of its GPUs, both on compute side(if you optimize your codes and boarders, for not conflicting when you are operating on 32 KB Warp, and 256 KB RFS per 128 cores, on an architecture where you have 64 KB Wavefront, 256 KB Register File Size per 64 cores - this is what you will get - underutilization of the latter architecture). It can be partially blamed on lazy developers, partially not, because actually it requires quite a lot of optimization of your code to get things done on GCN properly. But if you will get it, expect two times higher performance on for example Hawaii(R9 390X, 2816 GCN cores) GPU compared to GTX 980 Ti(2816 CUDA Cores). This is why I have called many, many times on this forum to you, professional: demand optimization of software for all IEMs.
This is compute side.
What AMD GCN always lacked is Geometry throughput, hence its bad efficiency in games. GCN was able to register 1 Triangle per clock per Shader Engine. If you have had 1 GHz GPU this meant your GPU was able to register 1 mln triangles per clock, with each cycle, per shader engine.
Nvidia on the other hand with Kepler was able to register the same amount of triangles, but per SM(compute cluster). So not only they were able to clock higher, but also Nvidia GPUs had more SM's in their layout on high level.
With Maxwell Nvidia added, however, Tile Based Rasterization, which resulted in massively increased efficiency. Tile Based Rasterization removes the need for moving the data strictly from memory to the GPU each time it has to be refreshed, because needed data are stored in ROP caches, and L2 cache. So all of graphics important parts in the execution pipeline(Polymorh Engine, Pixel Engines, etc) were connected directly to L2 caches, not Memory controller. L2 cache then was connected to memory controller.
GCN on the other hand always L2 cache, and other important parts of the graphics pipeline(Pixel Engine, Geometry Engine, etc) were connected to Memory controller, rather than L2 cache. This increased inefficiency, because GCN always in situation with unoptimized software caused stalls of the pipeline, and underutilization of the Compute Units. To be fair, Asynchronous Compute is actually one way to deal with it, but its like trying to fix sinking Titanic with a bandage.
GCN4 resolved quite a lot of problems however with GCN in geometry throughput, but it had few design flaws(unbalanced design in the first place). In the end, Polaris in Geometry Throughput is exactly on par with consumer Pascal GPUs and Maxwell GPUs, but it outpaces it in Compute throughput per clock and per core. My company did a little experiment.
We have tried to declock GTX 1050 Ti with 768 CUDA cores to 855 MHz(actually we have been able to declock it to 900 MHz only, but close enough), the same level of core clock Radeon Pro 555 has, with the same core count(768 GCN4 cores, with 855 MHz). End result? In Nvidia optimized games we have seen 3 FPS difference for Nvidia GPU. AMD GPu was still faster than Nvidia in compute.
As for Vega GPU. It has new execution pipeline, which is longer, thats why it allows to clock it to 1.6 GHz. It has Tile Based Rasterization, which should increase efficiency of the memory bandwidth, and less stalls in the piepline(increased utilization). Vega has better load balancing, but what AMD is doing is they are dynamically declock, or overclock different Shader Engines to manually balance the load on short length shaders, and long length shaders. Pretty neat feature, which helps balancing the load even more. To even more balance the load, and help fully utilize the GPU is Infinity Fabric, which underneath everything connects every part of the GPU, together. And on top of everything - GCN has finally geometry based features connected to L2 cache, not memory controller, which will help balance the load even more, and should fix its efficiency.
AMD GCN on the other side, always had "some form" of Tile Based Rasterization. Each Shader Engine was working on its own part of display/screen. If you have had for example 4 Shader Engines, display was partitioned into 4 pieces, and each Shader Engine was working on its own. However, differences in complexity of the parts of the screen ended up in smaller load on some Shader engines, and bigger on the others - creating unbalanced pipeline, and lowering the performance of the GPU, because it created stalls. Previous Generations of GCN, as I have said, were able to register only 1 triangle per clock.
One last thing, to put on top of Vega architecture.
One of ways AMD could icnrease throughput of the GPU was by enabling higher amount of L2 cache, available to the cores. It would affect both: Compute and graphics.
Here is compute comparison per clock Vega versus Fiji(Fury X):
And here is comparison of Vega with Fury X in most AMD friendly games:
Now, the question is: Is AMD incompetent?
Simplest answer - no.
Previously, there was a little argument on this forum between few people on this topic. Some people believed that Vega outpaces Titan Xp in compute just because it is Geometry based benchmark(SpecPerf). If that would be the case, then Vega should be smoking over Titan Xp in games. Wiping the floor with it. Flushing it in the toilet. And yet, in games, per clock the architecture is slower, than Fiji. At least that is how it appears on the first glance.
What is the problem, then? Unfortunately, I do not have an answer. Most probable is that software is not ready: both drivers, and applications(Applications have to be rewritten to use for Example Primitive Shaders, which increases throughput of the GPU on geometry front). I do not believe that AMD is incompetent. Vega has almost the same features as Pascal, and Volta, but it still will have 64 KB Wavefront, so it will do more work per clock(slightly...) cycle, even if finally Nvidia paired their architecture layout in compute throughput with Vega(64 cores/256 KB RFS).
As a side note. In perfect world, in GCN optimized games, Vega 10 should be around 40% faster than Fiji XT was, per clock, per core, in gaming scenarios. It has around 15% higher core throughput, and much, much more robust graphics pipeline, which should end in that level performance. Second note, just increasing the clock on Vega, without the new features, from 1050 MHz on Fiji, to 1.6 GHz on Vega should make it on par with GTX 1080 Ti/Titan Xp. Something is bottlenecking the architecture. But on hardware level - there is absolutely nothing that could cause this. Absolutely NOTHING. It has the same features, the same layout of graphics pipeline as Nvidia GPUs has. If Nvidia can have great gaming performance, Vega also should.
If you have got to the end of this post, congratulations.
Last edited:
AMD in even in Budapest compared RX Vega to GTX 1080, claiming that one system is 300$ cheaper than the other, and you do not see the difference between them.
The thing is... Both systems were exactly the same apart from GPUs, and Monitors.
System with AMD GPU and Freesync monitor was priced 300$ less, than Nvidia system with GTX 1080, and Asus PG34Q.
AMD Freesync monitor cost - 799$
Nvidia G-Sync Monitor - 1299$.
Both systems were the same, so RX Vega costs 200$ more, than GTX 1080.
The thing is... Both systems were exactly the same apart from GPUs, and Monitors.
System with AMD GPU and Freesync monitor was priced 300$ less, than Nvidia system with GTX 1080, and Asus PG34Q.
AMD Freesync monitor cost - 799$
Nvidia G-Sync Monitor - 1299$.
Both systems were the same, so RX Vega costs 200$ more, than GTX 1080.
I figured out some things, as well... Thanks!
Seems to me AMD had to release Vega Pioneer in an unfinished state (for gaming.) However, all they actually need is optimized software. The platform has solid bones and updates can always be added later.
I understand now, Vega has great potential, and even could be a major game changer ...in the very near future,
Last edited:
There is a possibility that Vega FE does not have ready BIOS, and voltage control, over different parts of the architecture(Infinity Fabric shenanigans).
Remember, Vega has improved load balancing that relies on different clock states between differently loaded Shader Engines, and it requires tremendous amount of control both on software and hardware side, to utilize this feature.
I have not mentioned few features that Vega has like High Bandwidth Cache Controller, and its affect on performance, simply because I forgot about it.
However, where Vega will excel is large data sets, and high resolution workloads(video editing, gaming).
Remember, Vega has improved load balancing that relies on different clock states between differently loaded Shader Engines, and it requires tremendous amount of control both on software and hardware side, to utilize this feature.
I have not mentioned few features that Vega has like High Bandwidth Cache Controller, and its affect on performance, simply because I forgot about it.
However, where Vega will excel is large data sets, and high resolution workloads(video editing, gaming).
AMD Radeon RX Vega 64 rumoured clocks
AMD Radeon RX Vega 64 rumoured prices
- AMD Radeon RX Vega 64 Liquid: base 1406 MHz / boost 1677 MHz
- AMD Radeon RX Vega 64 Limited: base 1247 MHz / boost 1546 MHz
- AMD Radeon RX Vega 64: base 1247 MHz / boost 1546 MHz
AMD Radeon RX Vega 64 rumoured prices
- AMD Radeon RX Vega: $499
- AMD Radeon RX Vega Limited Edition Model: $549
- AMD Radeon RX Vega Liquid Cooler Model: $599
- AMD Radeon RX Vega Liquid Cooler Limited Edition Model: $649
The dollar prices look good but in Europe we always get screwed with 1:1 conversion of dollars to pound and euros. The loss of competition has increased costs for us.
You were right about the need for increased software/firmware control. Tweaking will most likely allow HUGE increases in efficiency. AMD seems committed to improved control of this significant advance in hardware. Company management is asking for additional help from even more beta testers.
Sign up here... https://gaming.radeon.com/en-us/radeonsoftware/crimson-relive/vanguard/
I signed up but stated that I am with MacOS. Just wonder will they let me in, and what driver / software they can give me to test, haha
I do not see a correlation here. BIOS control over load balancing is not something that can be added through drivers. If there is no control in BIOS, the applications will not see this, because there is nothing to report to the scheduler, because there is no feature.
There are three layers: Hardware, Hardware BIOS - 1 layer, Drivers - 2 layer, Application - 3 layer.
ALL of those layers have to be perfectly synced to properly use this feature. DX12 and Vulkan Games should automatically use this feature without any Developer interaction apart from adding code that allows for use of this/acknowledging, by the engine. But first it has to be reported by the BIOS of the GPU.
Newegg leaked the prices starting at $499
https://www.reddit.com/r/Amd/comments/6qgq1v/newegg_accidentally_leaked_prices/
Your rumour comes from videocardz and refered to a post at guru3d which was about the vega 64, not the 56.
In any case we'll see soon enough what the msrp and actual street price will be.
You know better than me, where my rumor comes from?
http://www.2020techblog.com/2017/07/here-is-leaked-price-of-amd-rx-vega-64.html
[UPDATED – 7:36 AM ET 30/7/2017] We just got word that RX Vega 56 will cost $399 and the Limited Edition liquid cooled RX Vega 64 will retail for $699 and be part of a “special bundle”. We don’t know yet what this bundle will look like. We’ve now heard similar price quotes for the entire lineup from multiple sources, which is a good sign.
Last edited:
And your blog isn't citing their sources for that $399 v56 price either.
They probably just took that $499 and extrapolate the v56 price from it.
Just because someone on a blog says something, doesn't mean it's true. It may well end up to be true in the end but it's better to take things with a big shovel of salt when it comes to un-sourced rumors.
<facepalm>
My rumor comes from retail line. And so far they were correct, that RX Vega will cost 699$.
No your rumor comes from the blog you've linked to by your own admission... They just put their spin on what other have previously leaked, mostly the newegg leak that was first posted on guru3d
Post #43 by your own admission... That is where you got it from.
Or are you telling me, in a roundabout way, that this is your own personal blog?
If so, don't you find it a bit weird to try use yourself as your source for the veracity of this rumor?