[Updated with final Review] Vega FE Benchmarking by PCPER

[laminutederire]

Gamer Nexus review is out and is more thorough with their approach. They tested a bunch of things like thermal, noises, noticed some bugs pc per might not have (if have to read theirs another time to be sure). If that interest some people here

[Taf the Ghost]

1 hour ago, laminutederire said:

Gamer Nexus review is out and is more thorough with their approach. They tested a bunch of things like thermal, noises, noticed some bugs pc per might not have (if have to read theirs another time to be sure). If that interest some people here

http://www.gamersnexus.net/hwreviews/2973-amd-vega-frontier-edition-reviewed-too-soon-to-call/page-3

 

The professional stuff test is really interesting. In 3D design stuff, the Titan Xp is head by 15-20% for the rendering output, but in CAD-focused stuff, Vega FE is ahead by 25-35%. That's really quite interesting. It does explain that PCWorld piece where one of the project guys was explaining the use case. FE is less for Game Devs (though that would work) and more for small businesses that use CAD a lot.

 

http://www.gamersnexus.net/guides/2975-amd-radeon-vega-fe-vrm-pcb-analysis-by-buildzoid

 

For the breakdown of the parts. The really interesting part about Vega FE is massive amount of space on the end of the PCB that's not taken up by anything. I wonder what they're going to put there.  

[Taf the Ghost]

Vega FE card shot:

Spoiler

EVGA 1080 Classified card shot:

Spoiler

Obviously, the memory is right on the package for the Vega FE while the 1080 has the memory around the outside, but there's an empty spaces that are really interesting. Just to the right of the die and to the left of the VRM line. Even on the back of the card, there's pretty much a completely empty space on the upper portion.  If we assume that the Vega FE is a cut-down reference design, when Raja talked about RX Vega & Infinity Fabric, I'm now really curious if there is something there on the RX version. (Could be for the professional version as well, maybe some specialized hardware functions. I don't know that space well enough to take a stab. Maybe something Tensor Core-like?) 

[MoonSpot]

39 minutes ago, Taf the Ghost said:

-snip-

Don't forget to checkout buildzoids breakdown too.  Think it just went up, haven't seen it but am looking forward to it.  Might get a bit of a peak at what the man behind the curtain is up to.

[Taf the Ghost]

Just now, MoonSpot said:

Don't forget to checkout buildzoids breakdown too.  Think it just went up, haven't seen it but am looking forward to it.  Might get a bit of a peak at what the man behind the curtain is up to.

Actually listening to it now. 12-phase VRM and well designed. Apparently AMD's reference power design are well done.

[Jito463]

50 minutes ago, Taf the Ghost said:

For the breakdown of the parts. The really interesting part about Vega FE is massive amount of space on the end of the PCB that's not taken up by anything. I wonder what they're going to put there.  

It looks to me like the extra space is used for the cooler system.  That's right under where the fan sits, so I'm guessing it was just longer to give a place to put the fan.  Of course if that's the case, then they could make a shorter PCB for the AIO version, but they may not want to mess with multiple versions of the card right now.  This way they can take any card off the production line and either slap either a fan+shroud or an AIO on it and call it a day.

18 minutes ago, Taf the Ghost said:

Vega FE card shot:

  Reveal hidden contents

 

[Agosto]

13 minutes ago, MoonSpot said:

Don't forget to checkout buildzoids breakdown too.  Think it just went up, haven't seen it but am looking forward to it.  Might get a bit of a peak at what the man behind the curtain is up to.

Yep, it's up

[Valentyn]

1 hour ago, Taf the Ghost said:

Vega FE card shot:

  Hide contents

EVGA 1080 Classified card shot:

  Hide contents

Obviously, the memory is right on the package for the Vega FE while the 1080 has the memory around the outside, but there's an empty spaces that are really interesting. Just to the right of the die and to the left of the VRM line. Even on the back of the card, there's pretty much a completely empty space on the upper portion.  If we assume that the Vega FE is a cut-down reference design, when Raja talked about RX Vega & Infinity Fabric, I'm now really curious if there is something there on the RX version. (Could be for the professional version as well, maybe some specialized hardware functions. I don't know that space well enough to take a stab. Maybe something Tensor Core-like?) 

 

1 hour ago, Jito463 said:

It looks to me like the extra space is used for the cooler system.  That's right under where the fan sits, so I'm guessing it was just longer to give a place to put the fan.  Of course if that's the case, then they could make a shorter PCB for the AIO version, but they may not want to mess with multiple versions of the card right now.  This way they can take any card off the production line and either slap either a fan+shroud or an AIO on it and call it a day.

The Mi25 Instinct has a special NVME ssd spot on the pcb add as extra cache if needed. 

 

It could be they used the same PCB, but just removed the mounting and connection point which creates that big blank spot. 

 

Could explain why the water cooled Frontier Edition is also the same size. Probably also using the same PCB, instead of being Fury X size. 

[Jito463]

47 minutes ago, Valentyn said:

The Mi25 Instinct has a special NVME ssd spot on the pcb add as extra cache if needed. 

That's true, I'd forgotten about the SSG version.

[TOMPPIX]

http://www.3dmark.com/compare/sd/3829386/sd/3323533#

workstation vs gaming card.

 

if AMD tell us Vega FE is a workstation card, then the only driver's we will get at launch is workstation drivers.

[Taf the Ghost]

5 hours ago, Valentyn said:

 

The Mi25 Instinct has a special NVME ssd spot on the pcb add as extra cache if needed. 

 

It could be they used the same PCB, but just removed the mounting and connection point which creates that big blank spot. 

 

Could explain why the water cooled Frontier Edition is also the same size. Probably also using the same PCB, instead of being Fury X size. 

One thing about the PCB shots, which should be quite clear: we're definitely going to see a RX Vega Nano at some point.

[laminutederire]

8 hours ago, Taf the Ghost said:

http://www.gamersnexus.net/hwreviews/2973-amd-vega-frontier-edition-reviewed-too-soon-to-call/page-3

 

The professional stuff test is really interesting. In 3D design stuff, the Titan Xp is head by 15-20% for the rendering output, but in CAD-focused stuff, Vega FE is ahead by 25-35%. That's really quite interesting. It does explain that PCWorld piece where one of the project guys was explaining the use case. FE is less for Game Devs (though that would work) and more for small businesses that use CAD a lot.

 

http://www.gamersnexus.net/guides/2975-amd-radeon-vega-fe-vrm-pcb-analysis-by-buildzoid

 

For the breakdown of the parts. The really interesting part about Vega FE is massive amount of space on the end of the PCB that's not taken up by anything. I wonder what they're going to put there.  

To me it's driver related. It seems that in CAD everything is drawn on both cards; and the Titan get crushed because of that, while in 3D design, as in games, not everything is drawn on the Titan but significantly more is on Vega FE. That could explain the performance gap, which could be fixed with drivers at least partly.

But Gamer Nexus review was more informative to me than pcper's

[Captain_Tom]

I am sorry but I actually made an account to comment on these conversations.

 

 

I found this thread while I was in search for a decent discussion on what people thought of the latest Vega results.   A decent discussion for the most part.

 

 

However I have to say that the people doubting there is a driver issue...... Well you are idiotic to be frank.   

 

1. Tiled Rasterization is turned OFF.   This could bring a full 50% increase in performance.
2. If something as important as Tiled Rasterization is turned off, then it would not be surprising if things like Rapid Packed Math and other Architectural Enhancements are still dormant as well.  As people have said, it seems like this thing is effectively using Fiji drivers.
3. "Why would AMD launch a $1000 card that isn't ready?"  Simple:
   1. HBM stock is still terrible, but AMD promised an H1 launch to shareholders.  Thus a $1000 price is a great way to artificially limit demand.
   2. The professional drivers are mostly ready.  In some applications Vega FE is 30-50% faster than the Fury X at the same clocks.  That is the massive IPC increase AMD was talking about.

I somewhat doubt AMD will have good drivers ready for RX Vega's launch, but I fully expect them to bring at least a 20% performance increase over Vega FE.  Over time it should become anywhere from 50-100% stronger.

 

P.S.  I could certainly be wrong about the final increase, but come on: Vega matched the 1080 Ti is some games, and then lost to the 1070 in others.  A monkey could see that is clearly due to terrible drivers.

[xAcid9]

8 minutes ago, Captain_Tom said:

Tiled Rasterization is turned OFF.   This could bring a full 50% increase in performance.

Why 50%?

8 minutes ago, Captain_Tom said:

I somewhat doubt AMD will have good drivers ready for RX Vega's launch, but I fully expect them to bring at least a 20% performance increase over Vega FE.  Over time it should become anywhere from 50-100% stronger.

So they got all the time they need while waiting for HBM2 to be ready yet they can't prepare a good launch driver?

[mr moose]

3 hours ago, Captain_Tom said:

I am sorry but I actually made an account to comment on these conversations.

 

 

I found this thread while I was in search for a decent discussion on what people thought of the latest Vega results.   A decent discussion for the most part.

 

 

However I have to say that the people doubting there is a driver issue...... Well you are idiotic to be frank.   

 

1. Tiled Rasterization is turned OFF.   This could bring a full 50% increase in performance.
2. If something as important as Tiled Rasterization is turned off, then it would not be surprising if things like Rapid Packed Math and other Architectural Enhancements are still dormant as well.  As people have said, it seems like this thing is effectively using Fiji drivers.
3. "Why would AMD launch a $1000 card that isn't ready?"  Simple:
   1. HBM stock is still terrible, but AMD promised an H1 launch to shareholders.  Thus a $1000 price is a great way to artificially limit demand.
   2. The professional drivers are mostly ready.  In some applications Vega FE is 30-50% faster than the Fury X at the same clocks.  That is the massive IPC increase AMD was talking about.

I somewhat doubt AMD will have good drivers ready for RX Vega's launch, but I fully expect them to bring at least a 20% performance increase over Vega FE.  Over time it should become anywhere from 50-100% stronger.

 

P.S.  I could certainly be wrong about the final increase, but come on: Vega matched the 1080 Ti is some games, and then lost to the 1070 in others.  A monkey could see that is clearly due to terrible drivers.

 

Welcome to the forum.

 

1. don't call people idiots because they have a different opinion than you, it's against the standards and you will be reprimanded for it.

 

2. all companies release products that aren't ready at some point in their existence.  Usually it's becasue they are running behind and have no money.  I wouldn't make excuses outside of that because we just don't know.

 

and

3. don't bandy around numbers like "20%" and "50%" unless you have some citations or are willing to place make qualifiers that it is your assumption.

 

 

Enjoy the rest of the thread.

[LAwLz]

5 hours ago, Captain_Tom said:

Tiled Rasterization is turned OFF.   This could bring a full 50% increase in performance.

[Citation Needed]

Based on other implementations and just logical sense 50% seems way overboard. The numbers I've seen being talked about by far more knowledgeable people than myself is more along the lines of ~5%, and that's for very high resolution gaming (like 4K) where there is pressure on the memory.

Where did you get that 50% number from? Tile based rasterization has never been touted as a performance increasing feature. It's about efficiency, not performance.

 

5 hours ago, Captain_Tom said:

If something as important as Tiled Rasterization is turned off, then it would not be surprising if things like Rapid Packed Math and other Architectural Enhancements are still dormant as well.  As people have said, it seems like this thing is effectively using Fiji drivers.

That's just wild speculation. Maybe TBR is disabled because if not implemented properly it will break a ton of applications? Maybe the other features are simpler to implement and are active already?

 

5 hours ago, Captain_Tom said:

I somewhat doubt AMD will have good drivers ready for RX Vega's launch, but I fully expect them to bring at least a 20% performance increase over Vega FE.  Over time it should become anywhere from 50-100% stronger.

Is that 20% across the board? Because if it performs the same in 90% of applications/games but about 20% better in one game then I think it'd be a stretch to say that you're right. 50 to 100% faster is a very bold claim by the way.

I am looking forward to when RX Vega launches, because I will be saving this post and go back to it to see how right/wrong you were.

 

I mean, it would be funny if you went around calling other people idiots and then were completely off the mark yourself... Right?

[leadeater]

On 7/4/2017 at 3:20 AM, laminutederire said:

To me it's driver related. It seems that in CAD everything is drawn on both cards; and the Titan get crushed because of that, while in 3D design, as in games, not everything is drawn on the Titan but significantly more is on Vega FE. That could explain the performance gap, which could be fixed with drivers at least partly.

But Gamer Nexus review was more informative to me than pcper's

I believe the CAD type applications make a lot of use of half precision and double precision computation which is very limited on Titan/GeForce cards compared to Quadro cards. Vega FE is strong in all those areas so performs accordingly when asked to do those tasks.

 

Vega FE is much along the lines of no limits but no guarantees, unlike FirePro/Radeon Pro which is no limits with guarantees. Nvidia doesn't have this middle option which isn't really an issue, the cases where you are limited when using a GeForce card is not wide spread so AMD really does have a niche card for a subset of users in the professional industry.

 

A lot of CAD users can get away with using a GeForce card, unless you are doing rather complex work to hit performance limitations or need a feature of the software only available on a Quadro. A business should never buy a GeForce card though, unless they really know what they are doing and have a specific use case in mind where paying extra for Quadro makes little sense and you need to save the money (small business typically).

 

Edit:

Someone with a bit more understanding of the differences than me:

 

Quote

The difference is in view-port wire-frame rendering and double-sided polygon rendering, which is very common in professional CAD/3D software but not in games.

The difference is almost 10x-13x faster in single-fixed rendering pipeline (now very obsolete but some CAD software using it) rendering double sided polygons and wireframes:

 

Thats how entry level Quadro beats high-end GeForce. At least in the single-fixed pipeline using legacy calls like glLightModel(GL_LIGHT_MODEL_TWO_SIDE, GL_TRUE). The trick is done with driver optimization (does not matter if its single-fixed pipeline Direct3D or OpenGL). And its true that on some GeForce cards some firmware/hardware hacking can unlock the features.

If double sided is implemented using shader code, the GeForce has to render the polygon twice giving the Quadro only 2x the speed difference (it's less in real-world). The wireframe rendering remains much much slower on GeForce even if implemented in a modern way.

Todays GeForce cards can render millions of polygons per second, drawing lines with faded polygons can result in 100x speed difference eliminating the Quadro benefit.

Quadro equivalent GTX cards have usually better clock speeds giving 2%-10% better performance in games.

---

So to sum up:

The Quadro rules the single-fixed legacy now obsolete rendering pipeline (which CAD uses), but by implementing modern rendering methods this can be significantly reduced (virtually no speed gain in Maya's Viewport 2.0, it uses GLSL effects - very similar to game engine).

Other reasons to get Quadro are double precision float computations for science, better warranty and display's support for professionals.

That's about it, price-vise the Quadros or FirePros are artificially overpriced.

https://stackoverflow.com/questions/10532978/difference-between-nvidia-quadro-and-geforce-cards

[leadeater]

2 hours ago, LAwLz said:

Based on other implementations and just logical sense 50% seems way overboard.

If AMD had 50% more performance in the bag they would be WAY more smug about it and we would very much know already. No way AMD would keep quite about that just for the surprise, unless they really are completely stupid and want people to give up waiting and buy Nvidia.

[leadeater]

8 hours ago, Captain_Tom said:

I somewhat doubt AMD will have good drivers ready for RX Vega's launch, but I fully expect them to bring at least a 20% performance increase over Vega FE.  Over time it should become anywhere from 50-100% stronger.

50%-100%, err that has never happened in the history of GPU drivers for modern graphics APIs i.e. DirectX 7+ and OpenGL of same era. If it were possible to get that kind of performance increase from drivers alone why bother spending billions in GPU development when you could spend 10%-20% of that on driver optimization and rebadging cards pretending you have new hardware.

 

The R9 390X didn't gain 50% over the R9 290X as an example.

[Taf the Ghost]

4 hours ago, leadeater said:

If AMD had 50% more performance in the bag they would be WAY more smug about it and we would very much know already. No way AMD would keep quite about that just for the surprise, unless they really are completely stupid and want people to give up waiting and buy Nvidia.

RX Vega at 1600 Mhz with a 15% uplift from drivers is all AMD really needs. And if they were going to toast the 1080 Ti, they'd be really teasing that.

 

Given where the Compute performance is, they should land right between the 1080 and 1080 Ti, as they're just off the GP102 in the Pro side. (With trade offs depending on workflow.)