bottleneck by slow geometry ? How many polygon per clock ? You Know ? here from AMD Slide :
Raya Smith said:
Source
Yeah but vega is more compute GPU than gaming GPU.GP102 is 100% gaming GPU.Vega is like Kepler...(compute/gaming hybrid)
Vega have tons of things that dont need for gaming and only cost space and transistors.I think it will be slightly faster than GP104.It will be probalby badly bottleneck by slow geometry or only 64rops(most likely with only 4x shader engines)
New geometry engine will probably need optimization in game engine so it will be most of the time geometry bottleneck(around Gp106 geometry power like polaris but on 1500Mhz)
Also only 4x geometry procesors are confirmed so it will be just like Hawaii/FIJI only 4x shader engine.
Geometry throughput slide : Data base on AMD engineering design of Vega.Radeon R9 Fury X has 4 geometry engines and a peak of 4 polygons per clock.Vega is designed to handle up to 11 polygons per clock with 4 geometry engines.This represents an increase of 2.6x
Raya Smith said:But with Vega however, it looks like that limitation has finally gone away. AMD is teasing that Vega offers an improved load balancing mechanism, which pretty much directly hints that AMD can now efficiently distribute work over more than 4 engines. If so, this would represent a significant change in how the GCN architecture works under the hood, as work distribution is very much all about the “plumbing” of a GPU. Of the few details we do have here, AMD has told us that they are now capable of looking across draw calls and instances, to better split up work between the engines.
Vega GPU designs have 2048 bit memory bus, so if there is no 16 GB stack for HBM2 - you will be limited to two 8 GB stacks.
The stacks for HBM2 are 4GB. Nobody is making 8GB yet. And 16GB isn't happening.
And yet a HPC version with 16G's have allready been leaked.
16GB cards are supposed to happen at some time. But that's 8GB stacks and its only for HPC.
16GB stacks isn't even on roadmaps.
When Techreport, an who´s owner works for AMD says its 1070-1080 performance for Vega I trust it over your guesses.
Ok, so the point you're making, between the lines, is that vega is not gonna compete in the titan class? Or am I reading you wrong?
If Techreport and the leaks so far is true, its going to have a hard time even competing with a 1080.
But there is still 4-5 months till release or so. But we all know how the Polaris demos ended.
Only when primitive shaders are in use.If not it will have similat geometry performance as rx480.
Huh, how can you know that?
The new programmable geometry pipeline on Vega will offer up to 2x the peak throughput per clock compared to previous generations by utilizing a new “primitive shader.” This new shader combines the functions of vertex and geometry shader and, as AMD told it to me, “with the right knowledge” you can discard game based primitives at an incredible rate. This right knowledge though is the crucial component – it is something that has to be coded for directly and isn’t something that AMD or Vega will be able to do behind the scenes.
If so why use HBM over the latest DDR5X? You get all that expense, interposer complexity, memory size restrictions and (if I remember correctly) more latency, for a little higher bandwidth. Seems the worst of both worlds - either go DDR5X and have something that's cheap and flexible, or do HBM properly with 4 stacks and the appropriately huge performance.SO it has the same amount of memory bandwidth than a Fury X.
Only when primitive shaders are in use.If not it will have similat geometry performance as rx480.
which is 50% slower than GTX1060 btw.
http://www.hardware.fr/articles/953-7/performances-theoriques-geometrie.html
If so why use HBM over the latest DDR5X? You get all that expense, interposer complexity, memory size restrictions and (if I remember correctly) more latency, for a little higher bandwidth. Seems the worst of both worlds - either go DDR5X and have something that's cheap and flexible, or do HBM properly with 4 stacks and the appropriately huge performance.
Yeah but vega is more compute GPU than gaming GPU.GP102 is 100% gaming GPU.Vega is like Kepler...(compute/gaming hybrid)
Vega have tons of things that dont need for gaming and only cost space and transistors.
That statement never said that you need to do anything to hit the 2x peak triangle rate. Unless you want to imply that a mere doubling of throughput which is slower then GP102 is "an incredible rate". all it says is:Its explained on a few sites:
https://www.pcper.com/reviews/Graph...w-Redesigned-Memory-Architecture/Primitive-Sh
There can easily be two features here, one with processing more then one triangle per engine per clock and the other allowing for more intelligent discard.AMD told it to me, “with the right knowledge” you can discard game based primitives at an incredible rate.
The stacks for HBM2 are 4GB. Nobody is making 8GB yet. And 16GB isn't happening.
Only when primitive shaders are in use.If not it will have similat geometry performance as rx480.
which is 50% slower than GTX1060 btw.
http://www.hardware.fr/articles/953-7/performances-theoriques-geometrie.html
Yup, just like NV lost during GeForce 5, 6 & 7 generations. During the last 3 generations, Tahiti 7970Ghz smashed 680/770 and since November 2014 has been trading blows with a GTX780. Kepler will join the horrid GeForce 5 & 7 as the trifecta of NV architectural failures. Kepler was nothing more than marketing. The architecture had 0 legs for next gen games.
R9 290/290X easily leveled 780Ti during its generation and while Hawaii is able to play 2015-2016 modern games just fine, the 780Ti disaster is having trouble keeping up with an R9 380X/R9 470 in modern titles:
Fury X is now outperforming 980Ti at 1440p/4K:
https://www.techpowerup.com/reviews/Zotac/GeForce_GTX_1080_Amp_Extreme/29.html
Despite you constantly downplaying RX 480 before and post launch, it's an overall faster card than the 1060 as of now:
If Vega 10 underdelivers and ends up at 1070/1080 performance, at least let's hope it doesn't cost $600-700. Although looking at Fury X vs. 1070 1440p/4K benchmarks, 1070 level as representative of avg. performance can be discarded completely as an illogical prediction.
If so why use HBM over the latest DDR5X? You get all that expense, interposer complexity, memory size restrictions and (if I remember correctly) more latency, for a little higher bandwidth. Seems the worst of both worlds - either go DDR5X and have something that's cheap and flexible, or do HBM properly with 4 stacks and the appropriately huge performance.
wow, so 480 has 50% slower geometry performance than 1060, yet it now outperforms the 1060 overall across games in actual use considering DX11, DX12, Vulcan, etc?
Should we then not consider this a better performance design with Polaris compared to its Pascal competitor? Oh wait! You're going to tell me that 40-50W difference is now the most important metric!