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

[Karnak]

Looks like someone got a Vega FE.

Just got a Vega FE. Going to test some fp16 optimizations tomorrow. I have a few juicy targets

At some point I also need to test the paging system. It is awesome that we finally have automated paging on graphics workloads. Let's see how it manages a 32 GB volume texture

Update: Maybe I should buy extra 32 GB RAM to test with 64 GB volume texture instead

https://forum.beyond3d.com/threads/...rs-and-discussion.59649/page-107#post-1989071

 

[Malogeek]

Don't go bugging poor sebbbi with benchmark requests He'll have great insight into the architecture though which he'll share there.

 

[Phynaz]

Why would AMD bench the Radeon Vega FE in VRMark Orange Room and show pathetic scores while stating "Performance may vary based on use of latest drivers" . What are AMD trying to say ? That their Radeon Vega FE sucks on launch day. This card is marketed at VR enthusiasts too along with machine learning and 3D professionals. So they should atleast have solid VR performance on launch day. AMD marketing is horrible.

They promised to ship H1 2017. This is what happens when a product is rushed to market. Give it a couple of months.

 

[Malogeek]

They promised to ship H1 2017. This is what happens when a product is rushed to market. Give it a couple of months.

Or the fact that it was benched on a Xeon E5, not exactly enthusiast CPU performance which would impact the score significantly.

Still terrible idea adding VR scores but it is highly targeted to game developers.

 

[swilli89]

Or the fact that it was benched on a Xeon E5, not exactly enthusiast CPU performance which would impact the score significantly.

Still terrible idea adding VR scores but it is highly targeted to game developers.

No one cares about VR anyways. Whole thing is way over hyped.

 

[Bouowmx]

Intel Xeon E5-1650 v3 is the Core i7-5930K (3.6-3.7 GHz all-core-maximum frequency), plus additional 0.1 GHz with 1 or 2 active cores, not like for example, Xeon E5-2620 v3.

Though, VRMark scores still have a big range: Intel Core i7-5930K and NVIDIA GeForce GTX 1080 Ti

 

[Malogeek]

Intel Xeon E5-1650 v3 is the Core i7-5930K (3.6-3.7 GHz all-core-maximum frequency), plus additional 0.1 GHz with 1 or 2 active cores, not like for example, Xeon E5-2620 v3.

Though, VRMark scores still have a big range: Intel Core i7-5930K and NVIDIA GeForce GTX 1080 Ti

Thanks. That's a huge disparity in scores lol.

 

[Phynaz]

Or the fact that it was benched on a Xeon E5, not exactly enthusiast CPU performance which would impact the score significantly.

Still terrible idea adding VR scores but it is highly targeted to game developers.

Good point, why wouldn't they use Ryzen, Threadripper, or Epyc?

 

[Malogeek]

Good point, why wouldn't they use Ryzen, Threadripper, or Epyc?

Indeed. The latter two aren't even released and as of a month ago (when these benchmarks were done) the platform probably wasn't considered mature enough for performance metrics? Ryzen though.... just weird release.

 

[Phynaz]

Don't know if this has been posted.

Did they actually benchmark a Vega FE card against a Gforce GTX card in pro apps? That's not a Titan, right?

 

[Phynaz]

Indeed. The latter two aren't even released and as of a month ago (when these benchmarks were done) the platform probably wasn't considered mature enough for performance metrics? Ryzen though.... just weird release.

Someday AMD's Engineers are going to come for the Marketing team.

 

[guskline]

In the PC World release video, They showed an AMD Vega FE vs a Nvidia Titan xp with both using Ryzen 1800x with same ram etc.

 

[Wall Street]

One of my worries is coming to fruition. From Techreport:

"In an interesting disclosure, AMD says the Vega FE will have peak polygon throughput of 6.4 GTris/s, a figure that may belie the fundamental organization of the Vega GPU as four main shader clusters."

It appears that Vega still only does four triangles per clock, which is hugely disappointing given that this isn't nearly enough for a high end card in 2017. This is what held the Fury cards back. The competing GTX 1070, 1080 and 1080 Ti can all do over 10 GTris/s, which isn't possible with the existing Radeon front end unless clock speeds go above 2500 Mhz. It looks like this is a hint that the geometry shaders aren't working on general gaming workloads.

 

[Panino Manino]

One of my worries is coming to fruition. From Techreport:

"In an interesting disclosure, AMD says the Vega FE will have peak polygon throughput of 6.4 GTris/s, a figure that may belie the fundamental organization of the Vega GPU as four main shader clusters."

It appears that Vega still only does four triangles per clock, which is hugely disappointing given that this isn't nearly enough for a high end card in 2017. This is what held the Fury cards back. The competing GTX 1070, 1080 and 1080 Ti can all do over 10 GTris/s, which isn't possible with the existing Radeon front end unless clock speeds go above 2500 Mhz. It looks like this is a hint that the geometry shaders aren't working on general gaming workloads.

Hey, may I ask something?
Why is the triangle rate so low in these moderns GPUs?
I ask this because more than a decade ago gaming consoles used to compete with specs and I remember they promising tens of millions of polygons per second. It's like the numbers have decreased, strange.

 

[Bouowmx]

Confusing Giga (billion) and Mega (million)?

 

[Wall Street]

Hey, may I ask something?
Why is the triangle rate so low in these moderns GPUs?
I ask this because more than a decade ago gaming consoles used to compete with specs and I remember they promising tens of millions of polygons per second. It's like the numbers have decreased, strange.

Triangle rate has been somewhat less emphasized because modern games can use normal mapping to simulate geometry in a way that wasn't possible in the 1990s.

Here is an example from wikipedia:

https://en.wikipedia.org/wiki/Normal_mapping#/media/File:Normal_map_example.png

This allows less geometry to be used to simulate textures, however this can only go so far as the geometry added this way is fake and will not enhance the silouette or the actual depth of the object when rotated. The lighting calculation done by the normal mapping process is done in the pixel shaders.

FYI, modern GPUs are tens of Gigatriangles which is tens of billions - several orders of magnitude faster than what you are thinking of. Modern in-game scenes often have multiple millions of triangles per frame.

 

[ThatBuzzkiller]

Hey, may I ask something?
Why is the triangle rate so low in these moderns GPUs?
I ask this because more than a decade agora gaming consoles used to compete with specs and I remember they promising tens os millions of polygons per second. It's like the numbers have decreased, strange.

Well that's because of changes in hardware design philosophy and software bottlenecks ... (graphics programmers also figured out small triangles are a bad idea in terms of performance since GPUs shade 2x2 quads and not individual pixels so 2x2 quads covering less than 4 pixels saw hit in efficiency)

Last gen gaming consoles also don't compete with current GPUs in terms of geometry performance. The most consoles could do was 500 million triangles per second with the ATI Xenos and the GTX 1080 Ti could do 9.5 billion triangles per second, 19x as much! (I suspect the only reason why Pascal GPUs get over 10 billion in these B3D test suites from tech report is that the benchmark has some overdraw that they can bypass with an immediate mode tiled rasterizer but if they rasterized a 16k*16k full screen pass where each 2x2 quad covered a pixel they wouldn't even be able to hit over 150FPS. Vega 10 will probably see similar numbers to the Titan X Maxwell in that same test aswell since it's getting a tiled rasterizer as well.)

Geometry performance has grown much more slowly just like bandwidth has ...

 

[ThatBuzzkiller]

Triangle rate has been somewhat less emphasized because modern games can use normal mapping to simulate geometry in a way that wasn't possible in the 1990s.

Here is an example from wikipedia:

https://en.wikipedia.org/wiki/Normal_mapping#/media/File:Normal_map_example.png

This allows less geometry to be used to simulate textures, however this can only go so far as the geometry added this way is fake and will not enhance the silouette or the actual depth of the object when rotated. The lighting calculation done by the normal mapping process is done in the pixel shaders.

FYI, modern GPUs are tens of Gigatriangles which is tens of billions - several orders of magnitude faster than what you are thinking of. Modern in-game scenes often have multiple millions of triangles per frame.

Games can go beyond the use of triangles for geometric representation such as voxels and point clouds (Dreams PS4) ...

Shadow map rendering also sucks up quite a bit of that geometry throughput ...

 

[Freestyle7]

Why would AMD bench the Radeon Vega FE in VRMark Orange Room and show pathetic scores while stating "Performance may vary based on use of latest drivers" . What are AMD trying to say ? That their Radeon Vega FE sucks on launch day. This card is marketed at VR enthusiasts too along with machine learning and 3D professionals. So they should atleast have solid VR performance on launch day. AMD marketing is horrible.

One of the reasons I won't support AMD's future GPU products is because they did the same thing with the 7970 release. I dealt with horrible drivers for 6 months with this card. This is what AMD does.

 

[Glo.]

Why would AMD bench the Radeon Vega FE in VRMark Orange Room and show pathetic scores while stating "Performance may vary based on use of latest drivers" . What are AMD trying to say ? That their Radeon Vega FE sucks on launch day. This card is marketed at VR enthusiasts too along with machine learning and 3D professionals. So they should atleast have solid VR performance on launch day. AMD marketing is horrible.

https://videocardz.com/70511/amd-radeon-vega-frontier-features-gaming-mode

Although no one will officially confirm this, AMD Vega was postponed few times (according to AIBs). The reason for the delay is not related to any yield issues, but performance. AMD is clearly trying to boost it by optimizing software to the very last moment. This might also explain missing reviews of Frontier (even from the most trusted tech sites). The reality is that the very first person to provide gaming benchmarks of Vega Frontier is probably going to have the most popular content in months.

I think you have the big picture right now.

It is new architecture, that requires A LOT of work, to get drivers work properly.

 

[Borealis7]

hmm...i wonder what percentage of sales of Vega FE will be from Hardware sites wanting to bench Vega
also...can the FE mine? there will be a price hike because of extreme pro miners??

 

[xpea]

It is new architecture, that requires A LOT of work, to get drivers work properly.

I think it's time to be honest and realistic. This weird launch, without any gaming numbers, tells us that Vega cant beat the 1080Ti so they hide Vega from mainstream consumers as long as they can.
These days, you know well in advance of any real silicon (by simulation) the performance level you can reach (Cadence is your friend in this area). And if, to reach maximum performance, you need an unreasonable number of months to work on the drivers, then the product was badly designed from the start...

 

[tamz_msc]

I think it's time to be honest and realistic. This weird launch, without any gaming numbers, tells us that Vega cant beat the 1080Ti so they hide Vega from mainstream consumers as long as they can.
These days, you know well in advance of any real silicon (by simulation) the performance level you can reach (Cadence is your friend in this area). And if, to reach maximum performance, you need an unreasonable number of months to work on the drivers, then the product was badly designed from the start...

Yeah it must be so easy for anybody to walk right out of AMD with the vhdl files and see how it performs in advance.

Maybe if people tempered their expectations somewhat and looked at their past GPU launches to figure out that they haven't been competing directly with NVIDIA in the high-end space for a while, people would be clearer with their purchases. That would save them from disappointment if one camp or the other fails to deliver.

Want to spend 700$+ for the latest and greatest right now? Go for NVIDIA
Want something at under 300$ that does the job? Maybe AMD has something for you.

Much simpler.

 

[3DVagabond]

I think it's time to be honest and realistic. This weird launch, without any gaming numbers, tells us that Vega cant beat the 1080Ti so they hide Vega from mainstream consumers as long as they can.
These days, you know well in advance of any real silicon (by simulation) the performance level you can reach (Cadence is your friend in this area). And if, to reach maximum performance, you need an unreasonable number of months to work on the drivers, then the product was badly designed from the start...

Or you had almost all your resources going towards your last product launch (Because it was going to make or break the company) and your small team for Vega just couldn't meet the performance by the planned deadline.

 

[Glo.]

I think it's time to be honest and realistic. This weird launch, without any gaming numbers, tells us that Vega cant beat the 1080Ti so they hide Vega from mainstream consumers as long as they can.
These days, you know well in advance of any real silicon (by simulation) the performance level you can reach (Cadence is your friend in this area). And if, to reach maximum performance, you need an unreasonable number of months to work on the drivers, then the product was badly designed from the start...

No, my friend. You can tell the performance of the silicon based on Clock speed, and throughput of the cores, not how the hardware features will actually impact the performance.

Everything what we know about this release is that they have to push the performance so far. Vega has 3 meaningful departures from previous versions of GCN, which require reworking the drivers: Memory System paging, Improved Load Balancing on the GPU, and Tile Based Rasterization.

This weird launch is exact emanation of drivers being not ready to be released. AMD wanted to release those GPUs, for some reason(most likely - earning moneyzzz!, on early adopters, which are willing to pay the premium. They even do not sample the GPUs to reviewers, but they are selling them, very happily).

I wonder why you did not brought what Ryan Smith from this very site wrote on Beyond3D forum. You are there, as well.
This is what he wrote: https://forum.beyond3d.com/posts/1989164/

To be fair, if I had one I probably wouldn't bother with gaming numbers either. Any performance data collected now will be invalid by the time Vega RX launches. Which would make the work going into producing that data more or less wasted.

https://forum.beyond3d.com/posts/1989183/

Oh it's totally valid to benchmark it. I just have to tread a little more carefully; as AnandTech people often take what we say as the final word, and while I'm proud of the trust readers have in us, it means I need to be careful not to write a preview article and have 10 news aggregators posting "AnandTech proves that Vega sucks!" in the morning. Most readers get that a preview is a preview, but not everyone does.

(Also, I'm on vacation next week anyhow )