AMD Vega (FE and RX) Benchmarks [Updated Aug 10 - RX Vega 64 Unboxing]

[Muhammed]

I do not see a reason why architecture that has higher throughput, but requires software optimization, would be slower than weaker architecture in optimized software.

Vega doesn't have higher throughput than GP102 in any thing (not even in HBM2 memory bandwidth), learn how how to calculate numbers before you make such a grand false statements.

 

[Glo.]

Vega doesn't have higher throughput than GP102 in any thing (not even in HBM2 memory bandwidth), learn how how to calculate numbers before you make such a grand false statements.

On what do you base this on?

Vega, as other GCN has 64 cores per 256 KB register file size.
Pascal has 128 core/256 KB register file size.

The Register File size is what defines the throughput of the cores. The lower amount of cores, are able to work with larger register file size the better.

Example? 192 Cores/256 KB RFS in Kepler vs 128 cores/256 KB RFS in Maxwell. 128 Maxwell cores have had essentially the same amount of performance as 192 cores in Maxwell.

Vega has 64 cores/256, and those cores are also working with 64 KB wavefront compared to 32 KB warp in Nvidia. Hence the higher throughput.

Bah. Even previous generations of GCN are per clock faster than Nvidia GPUs in simplest thing as Compute. For example I can provide two proofs for RX 480 compared to GTX 1060 using CUDA, and AMD using OpenCL, and RX 480 compared to GTX 1070 in Adobe premiere applications, which only would solidify my statements.

 

[beginner99]

The only reason why its not faster: software developers did not implemented Primitive Shaders, and not used Programmable Geometry Pipeline, and drivers of said GPU are not ready.

Or some part of the hardware is broken and they could not fix it anymore. AMD could actually market upcoming huge boosts if they believed it in it. They have not. The whole thing was delayed and strange. Hardware error seems possible.

And if it actually is software, then shame on AMD. The haven fallen into that trap multiple times already most notably Bulldozer. If you are the underdog anything requiring each piece of software to be specially optimized is a non-feature because barley anyone will use it (only the ones you pay for it). Better to either use those transistors for more general hardware or omit them completely.

The most likely scenario however is HBM2 failure. They aren't reaching target speeds even at higher power use. The overvolted HBM2 probably also eats heavily into the power budget. It's very likely Vega is in gaming almost fully bandwidth limited. This also explains why the Vega 56 chip will come with slower memory. It has to because else it would maybe be same speed mostly as Vega 64. We will see soon enough if a memory OC of vega 56 will lead to Vega 64 performance. If that memory can even OC that high and isn't lower binned...

 

[Glo.]

Or some part of the hardware is broken and they could not fix it anymore. AMD could actually market upcoming huge boosts if they believed it in it. They have not. The whole thing was delayed and strange. Hardware error seems possible.

And if it actually is software, then shame on AMD. The haven fallen into that trap multiple times already most notably Bulldozer. If you are the underdog anything requiring each piece of software to be specially optimized is a non-feature because barley anyone will use it (only the ones you pay for it). Better to either use those transistors for more general hardware or omit them completely.

The most likely scenario however is HBM2 failure. They aren't reaching target speeds even at higher power use. The overvolted HBM2 probably also eats heavily into the power budget. It's very likely Vega is in gaming almost fully bandwidth limited. This also explains why the Vega 56 chip will come with slower memory. It has to because else it would maybe be same speed mostly as Vega 64. We will see soon enough if a memory OC of vega 56 will lead to Vega 64 performance. If that memory can even OC that high and isn't lower binned...

May I ask you a question? Why do you post your opinion as a fact?

Did you studied the architecture? Did you studied both: GCN4 and Nvidia Pascal architectures, from ground up?
There is no hardware reason why Vega FE would perform so poorly apart from software.

GCN5 is blend between Nvidia CUDA architecture, and GCN, with adding Primitive Shaders and Programmable Geometry Pipeline.

 

[leoneazzurro]

Nope, not gonna happen.
And theoretical numbers don't mean squat, if the architecture is inefficient or has multiple bottlenecks that prevent it from achieving it's maximum potential. And Vega appears to still have loads of these bottlenecks, memory compression is behind NV, polygon throughput is behind NV, (primitive shaders needs to be coded for by the driver and the developers, and it's efficacy is still unkown), ALU utilization is still behind NV (needs a lot of Async to increase utilization), ROPs and TMUs effective throughput is still behind NV, nearly every hardware aspect is behind NV. No surprise it's barely GP104 level. AMD's GPUs needs alot of hand holding and software intervention to be usable, which sucks, but is perfect for a console environment. NV's GPUs just work behind the scenes, transparently. This is the way to do it, that's why they are superior out of the box.

While I don't agree with Glo and its optimistic view, you and all other doomsayers are exactly at the opposite end of the spectrum, negating even the cases where Vega is good.
As I wrote before, games are only one of the scenarios where vega will be used. And one thare relies heavily on software optimizations. And many of the things you are saying about Vega disadvantages are SOFTWARE issues, not HARDWARE, meaning that an adequate programming can get overcome most of them . Memory compression is behind NV: yes it is, but Vega has also massive BW advantage compared to GP104 and we don't know yet if the Tiled rasterization in Vega will close part of the gap. Promitive shaders have to be coded? Yes, but they are there, and there is also something in the work on the driver side (transparent to the application), see the presentation slide deck. ROP and TMU REAL throughput (read: useful) depends also on how the game is coded. AMD needs more software intervention? Maybe, but hardware capabilities are there, async compute included, which is there to be used, not for the lazy programmer going in DX10 compatibility mode. This is why AMD GPUs usually mature better than their green counterpart, which sometimes had nasty surprises, like Kepler legacy status. Not only, but if you are interested in Vega as a GPU as a whole, you should have not missed that for professional applications it is already close to the GP102, and for compute tasks (except FP64) it gives GP100 more than a run for its money. And you already decided the NV cards are "superior out of the box" without seeing a single review, and for everyone, at whatever price. Good, it is quite telling about your bias.
For me (I have a laptop with a 1070 which works very good for me and I've had lots of "green" and "red" cards in the past, so for you information I am not so biased) the biggest flaw is the power draw. That makes high end Vega unsuitable for mobile space, which is not so good for the customers as there is a de facto monopoly in the mobile dicrete GPU market.
Die size is an issue for AMD not for the customer in the first place, but due to the high professional and especially compute performance it can be overcome by selling Vega in the high profit markets.
For the rest, you have a product that is quite probably on GTX1080 level, at GTX1080 price, which has the drawback of power draw and has the plus of the lower Freesync costs. Moreover, it was hinted that the Vega 56 variant may be a "best bang for the buck" case. Why it should be a complete failure, only God knows.

 

[Muhammed]

you should have not missed that for professional applications it is already close to the GP102

Nope, SpecView numbers put it around Quadro GP104 as well.

, and for compute tasks (except FP64) it gives GP100 more than a run for its money.

And how did you know that?

Vega, as other GCN has 64 cores per 256 KB register file size.
Pascal has 128 core/256 KB register file size.

Don't compare different architectures with some arbitrary factors, NV and AMD have different scheduling and command processor structure, direct comparisons don't apply here. AMD has always maintained higher ALU throughput than NV, but their utilization suffered. Same thing applies to Vega.

 

[leoneazzurro]

Nope, SpecView numbers put it around Quadro GP104 as well.

Spec view is only a part of the picture, here is a notoriously AMD fan site:
http://www.tomshardware.com/reviews/amd-radeon-vega-frontier-edition-16gb,5128-6.html

And this with EARLY drivers.

And how did you know that?
.

In the same manner as you say that It will suck for gaming.

 

[Glo.]

Don't compare different architectures with some arbitrary factors, NV and AMD have different scheduling and command processor structure, direct comparisons don't apply here. AMD has always maintained higher ALU throughput than NV, but their utilization suffered. Same thing applies to Vega.

On what do you base this on?

Utilization? Utilization suffered because of software, not hardware.

Blender for example added split kernel to their OpenCL compute pipeline, which made OpenCL implementation mimic CUDA. What happened with performance between both architectures?

Here is comparison between GCN(RX 480) and CUDA Pascal(GTX 1060) GPUs:

https://wiki.blender.org/index.php/Dev:Source/Render/Cycles/OpenCL
GTX 1060 is using CUDA, while RX 480 is using OpenCL.

Thanks to adding Split Kernel into the pipeline for OpenCL, the kernels are able to be processed much faster, reducing stalls in the pipeline. This is the effect comparing two GPUs: one which uses architecture that has lower throughput, and one that has higher throughput but cannot clock as high, as the other one.

The same effect we should see with games, when they will become optimized for Vega.

GCN5 solved all of utilization problems. I have no idea why people stubbornly are repeating the same argument, over and over again. Vega FE results are not showing anything, about true potential of Vega architecture.

And one last thing. Both architectures on high-level architecture picture are very similar. So yes, they direct comparisons apply, when trying to compare low-level architecture pictures.

 

[Glo.]

Nope, SpecView numbers put it around Quadro GP104 as well.

And how did you know that?

Do you know that Nvidia purposely is gimping features that are in consumer GPU drivers? If you will look at the benchmarks you will see that Titan Xp in Vega FE reviews in SpecPerf was slower than smaller GPU, based on GP104. Why? Because of market segmentation.

The catia-04 viewset was created from traces of the graphics workload generated by the CATIA V6 R2012 application from Dassault Systemes. Model sizes range from 5.1 to 21 million vertices.

The viewset includes numerous rendering modes supported by the application, including wireframe, anti-aliasing, shaded, shaded with edges, depth of field, and ambient occlusion

The creo-01 viewset was created from traces of the graphics workload generated by the Creo 2™ application from PTC. Model sizes range from 20 to 48 million vertices.

The viewset includes numerous rendering modes supported by the application, including wireframe, anti-aliasing, shaded, shaded with edges, and shaded reflection modes.

The energy-01 viewset is representative of a typical volume rendering application in the seismic and oil and gas fields. Similar to medical imaging such as MRI or CT, geophysical surveys generate image slices through the subsurface that are built into a 3D grid. Volume rendering provides a 2D projection of this 3D volumetric grid for further analysis and interpretation.

At every frame, depending on the viewer position, a series of coplanar slices aligned with the viewing angle are computed on the CPU and then sent to the graphics hardware for texturing and further calculations such as transfer function lookup, lighting and clipping to reveal internal structures. Finally, the slices are blended together before the image is displayed.

Energy-01 benchmark is the one which AMD claimed had big improvement with DrawStream Binning Rasterizer, which was not enabled in Vega FE drivers.

SpecPerf uses OpenGL, and hardware features that are available through drivers. For example - Antialiased Lines.

I am not interested in fanboyism. I am interested in hardware. If you, or anyone is using the SpecPerf numbers to tell which one hardware is better shows level of his/hers intellect, or cluelesness, about what he is talking about.

Vega FE has very basic drivers that do not even have optimizations from Radeon Pro lineup. Its driver lacked most of Vega features, and behaved just like Fiji.

Even then, Vega was around 30% faster per clock than Fiji.

It is because L2 cache is now 4 MB, instead of 2 that was in Fiji.

Its also interesting how different reviews get different scores in SpecPerf.

What to draw from this? Only one thing. Software is not ready for Vega. The same situation as is with Ryzen and Skylake-X CPUs.

 

[Muhammed]

Spec view is only a part of the picture, here is a notoriously AMD fan site:
http://www.tomshardware.com/reviews/amd-radeon-vega-frontier-edition-16gb,5128-6.html

By your logic, these tests are also part of the picture, P6000 is superior in most of them though. Vega FE is close in CPU limited tests.

In the same manner as you say that It will suck for gaming.

I never said that, all I said is that Vega is 1080 level (AMD's own statement), as opposed to 30% faster than a 1080Ti!!! Stop putting words into my mouth and stop defending false hype and wishful thinking.

Utilization? Utilization suffered because of software, not hardware.

Nope, AMD's design has bubbles that trashes it's utilization, it's front end is limited by polygon throughput, scheduling and by clock speed.

What to draw from this? Only one thing. Software is not ready for Vega. The same situation as is with Ryzen and Skylake-X CPUs.

And Bulldozer! don't forget that! Meanwhile NV gets to reign supreme unhindered by software+hardware drawbacks that plagues AMD hardware, great plan AMD!

GCN5 solved all of utilization problems. I have no idea why people stubbornly are repeating the same argument, over and over again. Vega FE results are not showing anything, about true potential of Vega architecture.

Yeah right, As evident by AMD's numbers, Vega FE numbers and RX Vega leaks. Can't believe after all of this someone has the guts to still hope for magic drivers.

 

[Glo.]

Nope, AMD's design has bubbles that trashes it's utilization, it's front end is limited by polygon throughput, scheduling and by clock speed.

This is correct. However GCN5 solves the Geometry Throughput limitation, with Primitive Shaders, and Programmable Geometry Pipeline.

And Bulldozer! don't forget that! Meanwhile NV gets to reign supreme unhindered by software+hardware drawbacks that plagues AMD hardware, great plan AMD!

It isn't supreme. Just because you repeat it constantly does not make it true. The software matures with time, and AMD did a lot of work lately on software side, and devs are now getting a lot of improvements from that work. Blender is one of examples.

Yeah right, As evident by AMD's numbers, Vega FE numbers and RX Vega leaks. Can't believe after all of this someone has the guts to still hope for magic drivers.

Drivers can make the GPU run at best 20% faster than it is running right now. Primitive Shaders and Programmable geometry pipeline - 40-50% faster.

If Geometry throughput was the only bottleneck for GCN, and primitive Shaders increases Geometry Throughput 2.5 times, why is it hard for anyone to believe that Vega is going to outpace GTX 1080 Ti, by 30%?

 

[Muhammed]

The software matures with time, and AMD did a lot of work lately on software side, and devs are now getting a lot of improvements from that work. Blender is one of examples.

Software matures with time compared to perfect software from the get go, yeah I know which one is supreme alright!

Drivers can make the GPU run at best 20% faster than it is running right now. Primitive Shaders and Programmable geometry pipeline - 40-50% faster.

Unsubstantiated numbers fueled by wishful thinking and hype. "Primitive Shaders" give 50% faster fps? even clockspeeds didn't give Vega this much, how did you arrive at that number anyway?

This tells me that you completely refuse to look at the hardware architecture, and you are prematurely judging the hardware. Its time for me to end conversation with you.

This tells me how far you've steered the thread into the realm of fantasy and dreams, you refuse to acknowledge official AMD statements and instead venture deep into your own imagination. Ending the conversation is the least thing to do with you.

 

[Glo.]

Software matures with time compared to perfect software from the get go, yeah I know which one is supreme alright!

Blender which I have already quoted, does not agree with you.

Unsubstantiated numbers fueled by wishful thinking and hype. "Primitive Shaders" give 50% faster fps? even clockspeeds didn't give Vega this much, how did you arrive at that number anyway?

Clock speed does not give Vega that much, because of software and drivers. If Vega has longer pipeline, if Draw Stream Binning rasterizer is not working with Vega FE, if load balancing is not working because of not ready BIOS of the GPU, and its voltage control over different parts of GPU, drivers alone can improve performance by 20%. But that remains how immature software was tied with Vega FE.

IF You praise Nvidia for such superior hardware, why you cannot see that AMD solving one bottleneck - Geometry Throughput, by 2.5 times will increase the Vega architecture performance dramatically?

Its not my job to make you believe. But failing to see obvious thing erodes your credibility as a valid conversation partner.

Currently Vega is registering 4 Triangles per clock. What will happen with performance when developers will implement Primitive Shaders and it will register 10 triangles per clock? What effect this will have on performance of the GPU? Simple question, simple answer.

This tells me how far you've steered the thread into the realm of fantasy and dreams, you refuse to acknowledge official AMD statements and instead venture deep into your own imagination. Ending the conversation is the least thing to do with you.

Ad personam attack is typical emanation of lack of arguments. Which is actually obvious, when all you can do is say: "Nvidia is better because it is better".

 

[IntelUser2000]

The most likely scenario however is HBM2 failure. They aren't reaching target speeds even at higher power use. The overvolted HBM2 probably also eats heavily into the power budget. It's very likely Vega is in gaming almost fully bandwidth limited.

The other chips with HBM2.

GP100 - Effective 1.4GT/s
GV100 - Effective 1.75GT/s

Vega 64 at the highest spec has 1.89GT/s.

So I'd say HBM2 will eventually reach 2GT/s, but right now, Vega has the highest clock of them all. So in that aspect nothing is out of the ordinary. Low yields due to difficulty in producing the chips are almost always the reason for chips that don't perform to spec. If you run them at lower clocks the percentage of chips that are functional rises.

Vega doesn't have higher throughput than GP102 in any thing (not even in HBM2 memory bandwidth)

This is true. PCGH tests show Vega is behind GP104 in bandwidth tests. Obviously it won't be the case everywhere, but we should take note. Actually their tests show it doesn't beat GP104 in everything except FLOPS. AIDA64 memory tests show Fury X is 20% ahead of Vega in memory bandwidth, way more than the 6% theoretical test indicate.

Paper specs can't always speak for reality, because implementation details aren't exposed. People believed the hype about Netburst because of double pumped ALU and so on. In the case of double pumped ALU for Netburst, we found out it could only do so with operations that weren't common. That didn't stop marketers claiming otherwise.

 

[IntelUser2000]

Mike Mantor of RTG says use of HBM isn't necessary to be classified as being compatible with HBCC. When asked regarding whether Notebook and APU Vega would use HBM2 because ALL Vega chips are going to use the HBCC feature, he says not necessarily and other memory types(like GDDR5) will qualify.

 

[Glo.]

Mike Mantor of RTG says use of HBM isn't necessary to be classified as being compatible with HBCC. When asked regarding whether Notebook and APU Vega would use HBM2 because ALL Vega chips are going to use the HBCC feature, he says not necessarily and other memory types(like GDDR5) will qualify.

Link, please.

 

[guskline]

Well. Vega will be on pre-order here soon. However I am not comfortable with pre-ordering it. Why do I keep thinking gtx1080ti? Specifically the Gigabyte AORUS waterforce Xtreme edition. Why?

I just got a Gigabyte AORUS Extreme with attached WB and it is a GREAT gpu. Salty but great in a custom loop. It is in my 5960x rig below. Moved my GTX1080 with a WB to my Ryzen 1800x rig since it will match the Vega64 right now.

 

[leoneazzurro]

By your logic, these tests are also part of the picture, P6000 is superior in most of them though. Vega FE is close in CPU limited tests..

Interesting. How do you know they are CPU limited? Also, in some test is ahead and P6000 costs 5x more. Price/performance is not your forte, it seems.

I never said that, all I said is that Vega is 1080 level (AMD's own statement), as opposed to 30% faster than a 1080Ti!!! Stop putting words into my mouth and stop defending false hype and wishful thinking.

And? 1080Ti costs a lot more. You defined Vega a bad GPU, for your own words standard 1080 should be a fail, too.

 

[beginner99]

The other chips with HBM2.

GP100 - Effective 1.4GT/s
GV100 - Effective 1.75GT/s

Vega 64 at the highest spec has 1.89GT/s.

So I'd say HBM2 will eventually reach 2GT/s, but right now, Vega has the highest clock of them all. So in that aspect nothing is out of the ordinary. Low yields due to difficulty in producing the chips are almost always the reason for chips that don't perform to spec. If you run them at lower clocks the percentage of chips that are functional rises.

GP100 has 720Gb/sec bandwidth (4 stacks, 4096 bit bus) compared to 480 GB/s (2 stacks, 2048 bit) in Vega FE. GP102 has roughly 432GB/s but for gaming due to NVs superior compression it has far more actual bandwidth than Vega.

 

[insertcarehere]

And? 1080Ti costs a lot more. You defined Vega a bad GPU, for your own words standard 1080 should be a fail, too.

Well, GTX1080 came out a year ago, and is also efficient enough to be used in SFF/notebooks without losing much (if any) performance, both of which Vega has not been shown as capable of doing.

 

[leoneazzurro]

Well, GTX1080 came out a year ago, and is also efficient enough to be used in SFF/notebooks without losing much (if any) performance, both of which Vega has not been shown as capable of doing.

The second one I already wrote about. I am not blind. And it's also the one thing that concerns me most, because I am a laptop gamer and having a monopoly in the discrete GPU area means that there is less money to spare compared to the situation where there is true competition. The "one year late" is irrilevant if someone wants to upgrade now.

 

[Qwertilot]

One year late is always relevant in this market. Its still moving forwards quite fast: ~30% year on year......

 

[maddie]

Mike Mantor of RTG says use of HBM isn't necessary to be classified as being compatible with HBCC. When asked regarding whether Notebook and APU Vega would use HBM2 because ALL Vega chips are going to use the HBCC feature, he says not necessarily and other memory types(like GDDR5) will qualify.

Link, please.

I too would very much like to see this.

 

[leoneazzurro]

One year late is always relevant in this market. Its still moving forwards quite fast: ~30% year on year......

It is relevant for the manufacturer. It is not relevant for the customer who buys NOW. In one year, scenario will change but if someone want to buy now, it is not important because otherwise you will wait forever for the "perfect" card.

 

[PhonakV30]

Well, GTX1080 came out a year ago, and is also efficient enough to be used in SFF/notebooks without losing much (if any) performance, both of which Vega has not been shown as capable of doing.

There is very big different GTX 1080 Mini and Vega Nano.Vega/Fiji Nano are True ITX cards While GTX 1080 Mini Isn't.Don't talk about Perf/W or Power draw or etc We both know story about Vega.about notebook , It's meh.Vega = a long path