AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

CarstenS said:
Alas, that's a rather ancient OpenGL Test. Will see if I can run it tomorrow in the office. But IIRC the results have been... strange for a couple of other cards a few years back, so I stopped using it on a regular basis. I still don't see, however, how I can correlate certain filtering modes to ALUs. Except the results between filtering modes differ wildly from the one in Fiji/Polaris - which the ones tested with the modern B3D suite do not indicate.
Click to expand...
Coming back to this:

Code: 
Nearest
Bilinear 
Trilinear

DXT1	 	DXT3	 	DXT5	 	INT8-1	 	INT8-2	 	INT8-3	 	INT8-4	 	FP10	 	RGB9E5	 	D16	 	D24	 	D32	 	FP16-1	 	FP16-2	 	FP16-3	 	FP16-4	 	INT16-1	 	INT16-2	 	INT16-3	 	INT16-4	 	FP32-1	 	FP32-2	 	FP32-3	 	FP32-4	 	INT32-1	 	INT32-2	 	INT32-4
Vega FE
141.176		160.000		184.615		200.000		218.182		218.182		266.666		266.666		266.666		200.000		266.666		266.666		300.000		300.000		300.000		300.000		300.000		266.666		266.666		266.666		266.666		266.666		266.666		266.666		266.666		266.666		266.666
141.176		160.000		184.615		218.182		240.000		240.000		240.000		266.666		300.000		200.000		300.000		300.000		300.000		266.666		160.000		160.000		266.666		266.666		141.176		141.176		266.666		150.000		77.419		77.419		266.666		266.666		266.666
96.000		92.307		120.000		133.333		141.176		160.000		160.000		160.000		171.429		150.000		150.000		141.176		160.000		141.176		75.000		75.000		141.176		141.176		72.727		72.727		150.000		75.000		38.709		38.709		266.666		266.666		266.666
 
Fury X
240.000		240.000		266.666		266.666		240.000		240.000		240.000		240.000		266.666		200.000		266.666		266.666		266.666		266.666		266.666		266.666		266.666		240.000		266.666		218.182		266.666		266.666		240.000		240.000		266.666		240.000		240.000
218.182		218.182		218.182		266.666		240.000		200.000		218.182		240.000		240.000		184.615		266.666		266.666		266.666		266.666		133.333		133.333		266.666		266.666		126.315		133.333		266.666		133.333		64.864		66.666		266.666		266.666		266.666
133.333		133.333		133.333		133.333		133.333		133.333		133.333		133.333		126.315		126.315		126.315		133.333		133.333		133.333		66.666		66.666		133.333		133.333		66.666		66.666		133.333		66.666		33.333		32.876		266.666		266.666		240.000
I still find the results of this test rather inconclusive with modern hardware.
 
CarstenS said:
I still find the results of this test rather inconclusive with modern hardware.
Click to expand...
Was that with equivalent clocks on Fiji and Vega?

INT8 is interesting as it gets faster with samples. Still appears slower than Fiji on the simple stuff. I'd agree results seem inconclusive. Generally equivalent for many tests, but worse at compressed formats which seems odd.

Thanks for the test though.
 
Nope, default clocks and default power targets. Out of the box, if you will.

FWIW, the runtime on those tests is very very short. Probably ramp up of clocks from idle will play a large role, whether or not it's at 218 GT/s or 240 and 266. My guess.
 
It's also probably not using high precision timers. Meaning yeah ms up, ms down will make it inaccurate and it probably just wasn't intended for such high fillrates.
 
The high precision part I figured with all the step differences. I guess the clocks could be explained away with a small job that finishes in a single cycle opposed to sustained throughput. That being the case TMUs are likely still there or I'd expect more variation.
 
I was hoping someone could help explain PCGH's polygon throughput tests. In particular, what exactly does "50% Culled" and "100% Culled" mean in the context of the Beyond 3D Suite? I understand that triangle culling is the elimination of triangles that can't be seen, or are just to small to be rendered.

With my original understanding, PCGH's results for the 1080 and RX580 made sense, because the 100% Culled test had a far higher throughput since the culled triangles were just being thrown out. You look at the Vega FE results, and it has the same throughput for 50% Culled and 100% Culled, which based on my original understanding indicated that it was not properly throwing out culled triangles.

However what threw a wrench in that idea for me, was that Fiji has a higher throughout with 50% Culled than it does 100% Culled. That leads me to believe I'm not understanding this process correctly. Could anyone help me out here?
 
AmericanLocomotive said:
With my original understanding, PCGH's results for the 1080 and RX580 made sense, because the 100% Culled test had a far higher throughput since the culled triangles were just being thrown out. You look at the Vega FE results, and it has the same throughput for 50% Culled and 100% Culled, which based on my original understanding indicated that it was not properly throwing out culled triangles.
Click to expand...
If it weren't properly suppressing triangles, at some point there is going to a be a problem since that would mean they weren't culled.
It's not required that the architecture be able to cull a triangle without taking throughput from the non-culled processing. Evaluating a triangle isn't free, and at least some evaluation has to happen in part of the pipeline all triangles go through. Some additional resources would need to be in place to allow more of the non-culled geometry to get through in addition to the culling.

However what threw a wrench in that idea for me, was that Fiji has a higher throughout with 50% Culled than it does 100% Culled. That leads me to believe I'm not understanding this process correctly. Could anyone help me out here?
Click to expand...

This sounds like you are going by the default selections, which is 100% a culled triangle strip test and a 50% culled triangle list.
The two types are not equivalent and architectures can treat them differently. The full set of tests seems to be showing that Vega is treating the various scenarios equally, which neither Fiji or Polaris do.
Vega and Polaris seem to be able to handle the strip without being constrained like Fiji is. Why Polaris has throughput indicating it can cull measurably more in addition to handling non-culled triangles, but Vega cannot is unclear.
 
Last edited:
AmericanLocomotive said:
However what threw a wrench in that idea for me, was that Fiji has a higher throughout with 50% Culled than it does 100% Culled. That leads me to believe I'm not understanding this process correctly. Could anyone help me out here?
Click to expand...
This is a misleading test because it mixes 50%/100% culled percentages with triangle strips and triangle lists.

AMD GPUs have traditionally had very poor performance when rendering triangle strips. GCN4 (Polaris) finally fixed this issue. Polaris peak performance is around 2x faster when rendering strips compared to previous GCN versions. Vega (GCN5) seems similar to Polaris (GCN4) in this regard. Polaris and Vega both also improve the performance of discarded triangles. A separate test case for both features would be needed to draw proper conclusions.
 
sebbbi said:
This is a misleading test because it mixes 50%/100% culled percentages with triangle strips and triangle lists.
Click to expand...
The full collection is in the dropdown menu on the left just above the diagram bars:
BNaF26q.png
 
While no-one had publicly stated they're making 8-Hi HBM2, Samsung has now confirmed they've increased their 8-Hi HBM2 production, which means they were doing them before too. They claim they're the only one doing 8-Hi HBM2 but it could just aswell refer to publicly available info instead of actual knowledge on wether Hynix is doing them too or not.
https://news.samsung.com/global/sam...hbm2-to-address-rapidly-growing-market-demand
 
"AMD is getting ready to show off the impossible at #SIGGRAPH2017. Sign up here for the spiciest AMD Capsaicin"

This can have many meaning but lets talk about the juicy one.

Any of you smart people think it is a good marketing strategy to show ur product inferior in everyway to a lower tier product that you said you were targeting for a month and then show "the impossible" and show people that it can actually compete against what it was mean to compete in first place?

I don't but I have no marketing degree so what could I know.
 
xEx said:
"AMD is getting ready to show off the impo. now ssible at #SIGGRAPH2017. Sign up here for the spiciest AMD Capsaicin"

This can have many meaning but lets talk about the juicy one.

Any of you smart people think it is a good marketing strategy to show ur product inferior in everyway to a lower tier product that you said you were targeting for a month and then show "the impossible" and show people that it can actually compete against what it was mean to compete in first place?

I don't but I have no marketing degree so what could I know.
Click to expand...

Well if people are expecting 1070 performance and they get 1080 or 1080ti performance i think people will be happy and you'd get alot of buzz closer to the realese. If that makes up for the month of bad buzz is anyones guess. Who knows maybe the high end vega will be faster than a 1080ti. Not much longer now
 
Rootax said:
BTW, "Big" Polaris was canned, or it was never in the book ?
Click to expand...
Good question.

IIRC, the unreleased chip in the Polaris time frame was Greenland. AMD was about to launch: Greenland, Baffin and Ellesmere. Later two were renamed to Polaris 11 and Polaris 10. However, both Baffin and Ellesmere are apparently old GFX8 based chips while Greenland has always been supposed to be GFX9 based.
 
There was never a big Polaris. Greenland was always mentioned as GFX9 and linked to HBM2, meaning Vega architecture.
The first time Greenland appeared was in a leaked slide by fudzilla.

Vpz54Bx.jpg


Two HBM2 links, 500GB/s. Plus I think the Greenland codename appears a bunch of times in Vega FE's drivers.
And looking at that quad-channel DDR4, it's linked to what we now know as a Threadripper through infinity fabric in a single MCM.
If it ever comes to life, this MCM would be as big as an EPYC CPU.
 
You must log in or register to reply here.
Back
Top