Bondrewd
Veteran
It's not delayed.
Right on money chefe.
Jawed
Legend
It seems, according to this slide:
that there is native "fixed function primitive culling hardware to remove SW based culling overhead" which implies that AMD has gone "backwards" with NGG (primitive shaders), adding new hardware for primitive culling.
Is that the correct interpretation?
that there is native "fixed function primitive culling hardware to remove SW based culling overhead" which implies that AMD has gone "backwards" with NGG (primitive shaders), adding new hardware for primitive culling.
Is that the correct interpretation?
Bondrewd
Veteran
Well yeah, they do a prim unit per shader array and N31 has 50% more of them over N21.
It seems there's another slide deck, as seen here:
DegustatoR
Legend
So as I've said, 7900XT will be more or less equal to 4080 in "rasterization" and will loose heavily when any RT will come into play.
7900XTX will likely end up being between 4080 and 4090 in "rasterization" and will also loose to the former with RT albeit there will probably be titles where it won't (FC6, REVIII, etc)
D
Deleted member 2197
Guest
There's a lot of RDNA3 performance numbers going around. Is that tweet including FSR in the 7900XT comparison against the 4080?
Perf numbers by Quasar Zone are for maxed out C2077 at 4K, no RT, and DLSS/FSR are used.
When they tested 4080, they matched the 6950XT framerate to the one listed by AMD in their 7900 XT/XTX perf slide.
I‘m confused that they write in headline 12 prim/clk and in the text 24 prim/clk. I think normally it uses 12 prim/clk, when you use primitive shader and put it in your program you get 24 prim/clk?It seems, according to this slide:
that there is native "fixed function primitive culling hardware to remove SW based culling overhead" which implies that AMD has gone "backwards" with NGG (primitive shaders), adding new hardware for primitive culling.
Is that the correct interpretation?
Maybe @Rys can give a little bit more clearance about this? Maybe he writes a benchmark
Waiting 10 years for beyond3d suite 2.0I‘m also wondering about Nvidia and it’s 12 Rasterizer. I was thinking that Ada will have 8 Rasterizer/GPC
Last edited:
TopSpoiler
Regular
Jawed
Legend
"Primitives" and "vertices" in the text - they're not actually the same thing.
In a mesh there's approximately one triangle (primitive) per vertex, but when a mesh "wraps" around an object, going from front-facing to back-facing there are potentially some triangles that end up being viewed edge-on. They are degenerate (have 0 area measured in screen pixels), so should be culled, though they are not actually back-facing.
Another cause of culling within a mesh is when some triangles end-up entirely outside of the screen. The mesh in this case is near the edge and only parts of it fall over the edge. Only triangles whose full set of vertices are outside of the screen can be culled though, you can't cull a triangle which only has one vertex that's outside.
But anyway, none of this relates to the question I have, the apparent non-use of NGG because there's fixed function hardware to remove software culling overhead.
Was NGG ever used?
Jawed
Legend
Really don't know.
The reference to "software based culling overhead" is the key here and I can't tell what AMD is actually describing. This overhead has been "removed" in RDNA 3. Does that imply AMD has been using NGG, but with RDNA 3 that's been dropped in favour of fixed function hardware?
I think they mean that you had to write some special code if you want to use the new pipline in your software. Didn't we discussed here that there are special commands for NGG? And i think now you can use the normal Piplines (DX11, DX12 and Vulcan) withouth the command and the polyons will automaticly bean routet to NGG?
DegustatoR
Legend
Was NGG ever exposed in client side s/w as something you can program for? AFAIR all geometry went through NGG automatically starting with RDNA1.
Didn't we discussed thet here:
forum.beyond3d.com
Current Generation Hardware Speculation with a Technical Spin [post GDC 2020] [XBSX, PS5]
Nintendo did it with Switch. There are games (MK11?) that run higher clocks in general (maybe just in handheld mode), and they also added a boost to clocks specific to some games during loading screens. The former has an impact on battery life, the latter may have an impact, but it might also...
forum.beyond3d.com
DegustatoR
Legend
What you can do on PS5 rarely has any relevance to what you can do on PC.Didn't we discussed thet here:
Current Generation Hardware Speculation with a Technical Spin [post GDC 2020] [XBSX, PS5]
Nintendo did it with Switch. There are games (MK11?) that run higher clocks in general (maybe just in handheld mode), and they also added a boost to clocks specific to some games during loading screens. The former has an impact on battery life, the latter may have an impact, but it might also...
Jawed
Legend
What is NGG and shader culling on AMD RDNA GPUs?
NGG (Next Generation Geometry) is the technology that is responsible for any vertex and geometry processing in AMD RDNA GPUs. I decided to do a write-up about my experience implementing it in RADV, which is the Vulkan driver used by many Linux systems, including the Steam Deck. I will also talk...
timur.hu
I hadn't seen this article before today. It talks about RDNA, which in effect means RDNA 1 and 2.
These are referred to as hardware stages, which I suppose at the least implies that they are initiated/scheduled by dedicated hardware, even though they run as shaders.
It's notable that RDNA 2 has functionality there that's not present on RDNA 1. Which may be relevant to general conclusions regarding NGG culling being of no benefit on RDNA 1.
The original merge request for culling is here:
radv: Implement NGG culling (!10525) · Merge requests · Mesa / mesa · GitLab
NGG culling: Quick summary This MR implements NGG culling on top of the NGG lowering that was merged in MR
gitlab.freedesktop.org
which matches, as far as I can tell, the model of culling that was described for Vega (but never used).
This article refers explicitly to RDNA 3:
and I think in the speculation thread we discussed NGG as being the only option for RDNA 3 in light of code commits.
The mystery remains, what hardware interaction does NGG use and is that hardware being referenced in the slide I linked earlier? The only hint at hardware usage that I can find is calling a s_sendmsg, where space in the output buffer is requested for vertex/primitives to be stored. Messages, themselves, have to be buffered (queued) in order to be handled and maybe the slide is referring to the intense message traffic produced by NGG (independent of culling)? Such intense message traffic can be interpreted as "software overhead", I suppose and have been improved in RDNA 3 versus RDNA 2.
Note for @Digidi : sorry when I quoted the slide earlier I misquoted it, which is how you ended up talking about primitives and I corrected you saying vertices...
RDNA3 really has dedicated matrix accelerators after all!
That's really exciting!
That's really exciting!
Well of course, that's likely one reason why it doesn't support mesh shaders. RDNA1 is a dinosaur architecture.
Yes but no. They're there to support the WMMA-instructions, Dot2 (FP16, BF16, INT8) and Dot4 (INT4), but they still employ all the "normal ALUs" in Compute Unit, not just those accelerators.
Last edited:
Jawed
Legend
I think that conclusion is wrong. This slide is pretty explicit in its title, in my opinion:
saying that the 64 lanes of the vector unit can also function as the "matrix accelerator" with two variants of dot product instructions, either DOT2 or DOT4. Well, these aren't traditional dot product instructions, because they are accumulating variants.
This is factually the case according to one of the articles I've just linked, Mesh Shaders requires per-primitive output parameters.
Last edited:
Similar threads
