AMD RX Vega Leaked Benchmark Shows It Ahead Of GTX 1080 – Specs Confirmed, 1630MHz Clock Speed, 8GB HBM2 & 484GB/s Of Bandwidth

[cj09beira]

Just now, Atmos said:

And here I was hoping we'd not see another boom like bitcoin caused last time... Damn, doesn't look like im upgrading to new gpus for the next year or so. Time to check out the used market in my area then I suppose.

At the start Nvidia Gpus were spared as they are slower but because there is no stock of rx500 series Gpus anywhere they started buying Nvidia ones too. 

In the last month 1 million Gpus were bought for mining which is crazy

[MageTank]

While it's nice to see Vega score decently in this synthetic benchmark, it seems the results were cherry picked in their favor. Most overclocked 1080's (around roughly 1850mhz) score around 34k in 3D Mark 11 Performance. 

 

http://www.3dmark.com/3dm11/11403703

http://www.3dmark.com/3dm11/11457489

http://www.3dmark.com/3dm11/12204470

 

 

In fact, some of the even higher clocked 1080's have scored as high as 36k:

 http://www.3dmark.com/3dm11/11986710

http://www.3dmark.com/3dm11/12185370

 

My point is, it's hard to use one single score on a synthetic benchmark that mostly reports DMI strings, as an accurate metric to gauge performance between two cards. One would have to go through the entire list and and average out the graphics score for the 1080, and then compare it against this one single Vega result, and hope that it's an accurate representation of Vega as a whole. If it's Vega at stock (or a very good bin) then it only muddies the water even further.

 

What bothers me the most, as I scroll through this list of 3dmark 11 results, is that I am about 6000 results in, and the weakest 1080 is scoring 30k+ still on it's graphics score. The result in OP's post, is scoring 27.6k. Did the article purposely look for the absolute worst 1080 result, to compare against Vega? For proof of my claims, here is a GTX 1080 result, ranked at 6021 on the list: http://www.3dmark.com/3dm11/12048737 

 

Now, here is an image to show the URL at the bottom left matching the result at the top, along with the current time as of posting this on the bottom right: 

 

6000+ results in, well beyond the "average" 1080, and it's still showing over 15% higher graphics scores on 3dmark 11 performance. Can we honestly say this article is making a fair comparison? Or are they intentionally deceiving their readers to build hype/start a flamewar for relevancy? 

 

I have a average GTX 1070. I'll be running 3dmark 11 performance shortly to confirm whether or not that 1070 score is inflated or deflated intentionally. Either way, this stinks to the high heavens.

[Atmos]

6 minutes ago, cj09beira said:

At the start Nvidia Gpus were spared as they are slower but because there is no stock of rx500 series Gpus anywhere they started buying Nvidia ones too. 

In the last month 1 million Gpus were bought for mining which is crazy

This actually straight up makes me depressed... thanks

[cj09beira]

Just now, Atmos said:

This actually straight up makes me depressed... thanks

dont worry you are not alone in this :(

[DXMember]

19 minutes ago, PCGuy_5960 said:

No ID, I just found a random result and edited the HTML 

I kno... but they have a result ID

[Master Disaster]

5 hours ago, BuckGup said:

It's made for gaming and the VEGA FE isn't

Like how Quadro is made for WS work and GeForce is made for gaming yet when Linus tested a Quadro in gaming workloads it performed exactly the same as its cheaper "gaming" equivalent?

 

Vega is Vega.

 

5 hours ago, mach said:

What makes you think software isn't important? It determines how gpu is utilized, like a racing driver in a sport car.

I didn't say software isn't important, I said I doubt driver optimisation could give that much of a boost but it what others are saying about FE being launched on Fiji drivers is true then that could explain the huge difference.

 

 

[cj09beira]

1 minute ago, Master Disaster said:

Like how Quadro is made for WS work and GeForce is made for gaming yet when Linus tested a Quadro in gaming workloads it performed exactly the same as its cheaper "gaming" equivalent?

 

Vega is Vega.

 

I didn't say software isn't important, I said I doubt driver optimisation could give that much of a boost but it what others are saying about FE being launched on Fiji drivers is true then that could explain the huge difference.

 

 

its been proven that the new hardware features of vega aren't working, so there is that 

[Master Disaster]

4 hours ago, Valentyn said:

Base speed of 1630Mhz compared to Frontier Edition base of 1380Mhz with Boost to 1600Mhz. The FE during all tests with PcPer never exceeded 1440Mhz.

So it makes sense than if it was at a stable 1630Mhz it would do that well.

Especially if like the Fury X, the top RX Vega comes with an AIO water cooler on it.

Well damn, it was 5:30am when I read the numbers and I seriously misread them, thanks for pointing that out.

[Hunter259]

3 minutes ago, cj09beira said:

its been proven that the new hardware features of vega aren't working, so there is that 

Of which the rasterization is something that is hardware level and can't be changed by a driver unless AMD did some fuckery to hide the tiled rendering. I found 0 evidence that you can change how a card rasterizes through software.

[cj09beira]

Just now, Hunter259 said:

Of which the rasterization is something that is hardware level and can't be changed by a driver unless AMD did some fuckery to hide the tiled rendering. I found 0 evidence that you can change how a card rasterizes through software.

unless vega Fe is using a older stepping of the silicon, it might be somehow switched of at the driver/bios level

[Hunter259]

Just now, cj09beira said:

unless vega Fe is using a older stepping of the silicon, it might be somehow switched of at the driver/bios level

That also doesn't make sense as Tile rasterization has always been apart of the VEGA archetecure for as long as we've known about. It would be stupid to make two different versions of this chip with one that's gimped. And like I said. Unless they have done some massive fuckery, I can't find any evidence that you can change how it does rasterization outside of making a whole new architecture.

[cj09beira]

1 minute ago, Hunter259 said:

That also doesn't make sense as Tile rasterization has always been apart of the VEGA archetecure for as long as we've known about. It would be stupid to make two different versions of this chip with one that's gimped. And like I said. Unless they have done some massive fuckery, I can't find any evidence that you can change how it does rasterization outside of making a whole new architecture.

i think there might be a register that tells the rasterizer how to rasterize maybe, or both methods are implemented i dont know

[Hunter259]

Just now, cj09beira said:

i think there might be a register that tells the rasterizer how to rasterize maybe, or both methods are implemented i dont know

That could explain why the damn silicon is fucking huge. Could also be that GCN is reaching it's end. Maybe Navi will finally be a change of pace.

[Taf the Ghost]

1 minute ago, Hunter259 said:

That could explain why the damn silicon is fucking huge. Could also be that GCN is reaching it's end. Maybe Navi will finally be a change of pace.

Some of the interior connections should have been moved to Infinity Fabric, making GCN more modular at the architecture level. Vega is a stepping stone to Navi and the true multi-die GPUs. Which is where AMD should head because their entire Graphics Card department has been focused on making the best iGPUs for a decade.

[Hunter259]

4 minutes ago, Taf the Ghost said:

Some of the interior connections should have been moved to Infinity Fabric, making GCN more modular at the architecture level. Vega is a stepping stone to Navi and the true multi-die GPUs. Which is where AMD should head because their entire Graphics Card department has been focused on making the best iGPUs for a decade.

Infinity Fabric isn't in VEGA and multi-die is still a fantasy at the moment.

[cj09beira]

2 minutes ago, Hunter259 said:

That could explain why the damn silicon is fucking huge. Could also be that GCN is reaching it's end. Maybe Navi will finally be a change of pace.

gcn isn't a bad design at all, the silicon is big and i cant figure out why.

1 it cant be the memory controller as the added space because of what it can do is balanced by the fact that hbm has most of the controller in the modules themselves.

maybe the new compute units but even then.. adding fp16 support wouldn't increase its area by much 

so the only thing that comes to mind is pipelining, maybe they added another a "level"

[cj09beira]

Just now, Hunter259 said:

Infinity Fabric isn't in VEGA and multi-die is still a fantasy at the moment.

yes it is, they used it to connect everything up inside the gpu itself

[Hunter259]

Just now, cj09beira said:

yes it is, they used it to connect everything up inside the gpu itself

ah nvm. I'm behind on that part. Not like it matters much. It's just another interconnect.

 

2 minutes ago, cj09beira said:

gcn isn't a bad design at all, the silicon is big and i cant figure out why.

1 it cant be the memory controller as the added space because of what it can do is balanced by the fact that hbm has most of the controller in the modules themselves.

maybe the new compute units but even then.. adding fp16 support wouldn't increase its area by much 

so the only thing that comes to mind is pipelining, maybe they added another a "level"

If I remember correctly; All GCN cards are completely backwards compatible with earlier revisions meaning if a new GCN instruction is added that is better than an old one the old still has to stay due to compatibility. That in it self would bloat the architecture pretty heavily.

[cj09beira]

10 minutes ago, Hunter259 said:

ah nvm. I'm behind on that part. Not like it matters much. It's just another interconnect.

 

If I remember correctly; All GCN cards are completely backwards compatible with earlier revisions meaning if a new GCN instruction is added that is better than an old one the old still has to stay due to compatibility. That in it self would bloat the architecture pretty heavily.

good point, it might be time to ditch some old stuff then, but because gcn was always so focused on future there probably arent too many old instructions 

[ltguy]

Why are they using 3dmark 11... just to make it impossible to compare against all other data using current 3dmark? So much for comparing the data to RX480/RX580 crossfire reviews.

[MageTank]

1 hour ago, MageTank said:

While it's nice to see Vega score decently in this synthetic benchmark, it seems the results were cherry picked in their favor. Most overclocked 1080's (around roughly 1850mhz) score around 34k in 3D Mark 11 Performance. 

 

http://www.3dmark.com/3dm11/11403703

http://www.3dmark.com/3dm11/11457489

http://www.3dmark.com/3dm11/12204470

 

 

In fact, some of the even higher clocked 1080's have scored as high as 36k:

 http://www.3dmark.com/3dm11/11986710

http://www.3dmark.com/3dm11/12185370

 

My point is, it's hard to use one single score on a synthetic benchmark that mostly reports DMI strings, as an accurate metric to gauge performance between two cards. One would have to go through the entire list and and average out the graphics score for the 1080, and then compare it against this one single Vega result, and hope that it's an accurate representation of Vega as a whole. If it's Vega at stock (or a very good bin) then it only muddies the water even further.

 

What bothers me the most, as I scroll through this list of 3dmark 11 results, is that I am about 6000 results in, and the weakest 1080 is scoring 30k+ still on it's graphics score. The result in OP's post, is scoring 27.6k. Did the article purposely look for the absolute worst 1080 result, to compare against Vega? For proof of my claims, here is a GTX 1080 result, ranked at 6021 on the list: http://www.3dmark.com/3dm11/12048737 

 

Now, here is an image to show the URL at the bottom left matching the result at the top, along with the current time as of posting this on the bottom right: 

 

6000+ results in, well beyond the "average" 1080, and it's still showing over 15% higher graphics scores on 3dmark 11 performance. Can we honestly say this article is making a fair comparison? Or are they intentionally deceiving their readers to build hype/start a flamewar for relevancy? 

 

I have a average GTX 1070. I'll be running 3dmark 11 performance shortly to confirm whether or not that 1070 score is inflated or deflated intentionally. Either way, this stinks to the high heavens.

Well. I ran my 1070 at 2050mhz, my CPU at 5ghz, and used my ram's 3200 XMP profile for an "average overclocked" run. My results: http://www.3dmark.com/3dm11/12258063

 

26,775. This is over 4k higher than the 1070 result in the OP. A massive 20% difference in graphics score, and for what? To make the Vega look better than it really is? This is absolutely sad, and yet another reason why WCCFTech should never be trusted. They create this massive hype by skewing the results intentionally, and ride the hypetrain they generate for more clicks.

 

If anyone would like me to underclock my 1070 to see exactly how crippled it needs to be in order to achieve their graphics score, I'll gladly do so. 

[cj09beira]

1 minute ago, ONOTech said:

No, not necessarily.

 

The test used an extreme amount of polygons. The Fury X only has 4 GB of HBM memory compared to the 16 GB of HBM2 memory in the Vega card. The Vega card will decisively win based on the Fury X's VRAM limitations alone.

 

Plus, the Fury X is a gaming card. Even with bad drivers, the Vega has a much better chance in CAD and Workstation applications compared to the Fury X because...well, you know...it was made for that type of stuff, lol.

if you think a fury is slow in this pro apps, you wrong, it it was a memory issue we would see performance tanking which we only saw once ( where vega was 800% ahead)

[BuckGup]

1 hour ago, Master Disaster said:

Like how Quadro is made for WS work and GeForce is made for gaming yet when Linus tested a Quadro in gaming workloads it performed exactly the same as its cheaper "gaming" equivalent?

 

Vega is Vega.

 

I didn't say software isn't important, I said I doubt driver optimisation could give that much of a boost but it what others are saying about FE being launched on Fiji drivers is true then that could explain the huge difference.

 

 

I never said it would perform worse. I was just saying it wasnt for gaming

[cj09beira]

Just now, ONOTech said:

I never once stated it's slow. Just said that it's inevitably slower than Vega FE - a workstation card. That does not directly correlate to bad drivers. It could just be that FE is a beefier card with more computational power and VRAM.

you must have not seen this yet then 

this is vega vs fury x at the same frequency, meaning same compute perf, this will show improvements in things like geometry processing etc.

gcn by it self is a workstation class architecture, as its balance between compute and graphics is more to the side of compute, thats why a rx480 die can be as fast or faster than a 1070 when both are using certified drivers (in compute work loads)

[cj09beira]

Just now, ONOTech said:

Image is broken ): Original source link please?

http://www.gamersnexus.net/guides/2977-vega-fe-vs-fury-x-at-same-clocks-ipc 

its the last image