My guess is that support for merged shader stages will land in Vulkan and DX soon as AMD was also pursuing API integration. If these features were made available through a common interface or in a way that requires only a minimal amount of vendor-specific changes to engine code, I could imagine it being used quite readily.
AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion
My guess is that support for merged shader stages will land in Vulkan and DX soon as AMD was also pursuing API integration. If these features were made available through a common interface or in a way that requires only a minimal amount of vendor-specific changes to engine code, I could imagine it being used quite readily.
Rootax
Veteran
When I talked about that with Sebastian Aaltonen (claybook) on twitter months ago, he told me there is no api available, no way to use primitive shaders on Vega right now. So the devs can't "implement it".
Ah, right, must have mixed it with the Wolfenstein's use of intrisincs to achieve similar function as primitive shaders (was it intrisincs or just regular compute?)
I'd bet on this being the case. With both AMD and NVIDIA having this supported, it would make sense to implement in API's.
Rootax
Veteran
On claybook he uses compute if my memory is right, so I asked him, does your compute code work better than primitive shaders to discard things, and the response followed...
Just compute shaders culling geometry:
https://techreport.com/news/33153/radeon-rx-vega-primitive-shaders-will-need-api-support
It's also an ingame setting which you can toggle on/off on all GPUs.
Next-generation Geometry for Vega is history
https://www.hardwareluxx.de/index.p...n-geometry-fuer-vega-offenbar-geschichte.html
Last edited:
Even if both have merged shaders, what each vendor merges appears to be different, and their initial implementations don't look like they agree on what gets culled, where culling happens, or what their shaders plug into.
AMD's initial offering for primitive shaders has them pulling the usual data from the same index buffer, feeding into the same tessellation stage, conservative triangle culling in the primitive shader/VS phase in front of the rasterizer (albeit a binning one).
Nvidia' method has a task shader pulling an object list, doing LOD and coarse culling, launching varying numbers of mesh shaders with no mention of the tessellation block, running mesh shaders (one recommendation was to let the fixed-function path handle triangle culling), and feeding into the rasterizer (possibly not binning?).
AMD's method directly plugs into the existing path, while Nvidia's may be to the side as there are recommendations for using tessellation when optimal, and there are normal call types still available despite the new shaders being considered separate from the existing types.
AMD's initial offering for primitive shaders has them pulling the usual data from the same index buffer, feeding into the same tessellation stage, conservative triangle culling in the primitive shader/VS phase in front of the rasterizer (albeit a binning one).
Nvidia' method has a task shader pulling an object list, doing LOD and coarse culling, launching varying numbers of mesh shaders with no mention of the tessellation block, running mesh shaders (one recommendation was to let the fixed-function path handle triangle culling), and feeding into the rasterizer (possibly not binning?).
AMD's method directly plugs into the existing path, while Nvidia's may be to the side as there are recommendations for using tessellation when optimal, and there are normal call types still available despite the new shaders being considered separate from the existing types.
...for Linux. Could be dead in the water for Windows too, but not necessarily.
Rootax
Veteran
If it's not there yet, imo we can assume it's dead too.
Same source as this and a couple of weeks old.
We can also assume that AMD doesn't actually lie in their slides which means that it at least has worked on Windows since there's a benchmark of it. So the hardware side works at least on certain cases
So does anybody think there is going to be a Vega128 die? (Using infinity fabric 2.0 to unify the memory?)
AMD (under Dr Su) is targeting each market aggressively and designing their (modular) hardware to fit each segment. As such, the Radeon RX chip is different from their Radeon Instinct chips, etc. And if you look at what AMD is doing with their SOC they are selling, & the tech laying underneath it all, it is AMD's Infinity Fabric. And Dr Su (and Raj) said it would get more important in the future (7nm release). Rightfully, Dr Su fired Raj because Vega didn't suite the gaming market. (Not enough of the chip die was dedicated to immediate end-user performance.) Lesson learned.
So how big is 7nm Vega? Will there be two difference chips for each end of the GPU market..?
AMD (under Dr Su) is targeting each market aggressively and designing their (modular) hardware to fit each segment. As such, the Radeon RX chip is different from their Radeon Instinct chips, etc. And if you look at what AMD is doing with their SOC they are selling, & the tech laying underneath it all, it is AMD's Infinity Fabric. And Dr Su (and Raj) said it would get more important in the future (7nm release). Rightfully, Dr Su fired Raj because Vega didn't suite the gaming market. (Not enough of the chip die was dedicated to immediate end-user performance.) Lesson learned.
So how big is 7nm Vega? Will there be two difference chips for each end of the GPU market..?
WTF did I just read?
Instinct uses exact same chips as RX
Raj wasn't fired.
Vegas problem(s) isn't "not enough die dedicated to immediate end-user performance".
WTF did I just read?
Instinct uses exact same chips as RX
Raj wasn't fired.
Vegas problem(s) isn't "not enough die dedicated to immediate end-user performance".
What happened with Raja Koduri is murky at best. I agree with everything else.
I'm still sure the whole "Sony-Navi" is just a miscommunication or misunderstanding in some unofficial conversation someone overheard or simply false rumor.
AMD does their architectures for themselves, the big semi-custom clients like Sony and MS give their input on what they would like to see in their next semi-customs (like apparently putting more ACEs with more threads in was Sonys idea but it was implemented for GCN as a whole by AMD) and possibly do even part of the engineering for their particular chips. Only way Navi could be "made for Sony" would be MS saying they want Vega-based instead (or next gen after Navi but that's probably too far in the future so unlikely) and even then Navi would be AMD first, Sony second.