47 Photos and videos
1/10
My RDNA5 predictions based on AMD patents and research papers.
See other pinned post for more in-depth info or my threads here for specific info 👇
@RedGamingTech @RGamerB_ @3DCenter_org @blueisviolet @oliemack @hms1193 @SebAaltonen
@RedSwirl @nietras1 @compusemble @opinali
1
5
53
10,905
1/6 Intro:
Writing a follow-up in case you want to go even deeper. Been following and analyzing AMD GPU patents for the last ~9 months.
I've threads on X about this. If you want more info and analysis then I've linked multiple posts from Reddit and the Anandtech Forums below:
Early list of novel GPU patents. Most are used in PS6/RDNA5 (see @Kepler_L2 in Reddit comments):
reddit.com/r/GamingLeaksAndR…
@Mareeswj @dylan522p @SKundojjala @RedSwirl @GameGPU_com @3DCenter_org @RGamerB_ @anthony256 @theclub386 @mzuhair123 @carygolomb @pilerks1 @RedGamingTech
1
1
11
3,519
1/10
My RDNA5 predictions based on AMD patents and research papers.
See other pinned post for more in-depth info or my threads here for specific info 👇
@RedGamingTech @RGamerB_ @3DCenter_org @blueisviolet @oliemack @hms1193 @SebAaltonen
@RedSwirl @nietras1 @compusemble @opinali
1
5
53
10,905
A19 Pro's second gen dynamic caching is incredible.
RT occupancy A18 Pro 43.63% -> A19 Pro 66.60% 🤯 Geekerwan's review: youtube.com/watch?v=Y9SwluJ9…
Cyberpunk 2077 RT Overdrive
36.3% (PT) and 40.8% (RT)
SER off vs on: 30.4% -> 44.2%
C&C: chipsandcheese.com/p/shader-…
@SebAaltonen 1/2
2
7
1,001
IMPORTANT READ THIS:
I mixed up occupancy and coherence. This is not the same thing.
Apple test is for kernel occupancy and RT unit utilization, SER test (C&C) is coherence.
Also not the same test, RT workloads can vary widely, so can't be compared.
Sorry for any confusion. I
1
4
232
1/10
My RDNA5 predictions based on AMD patents and research papers.
See other pinned post for more in-depth info or my threads here for specific info 👇
@RedGamingTech @RGamerB_ @3DCenter_org @blueisviolet @oliemack @hms1193 @SebAaltonen
@RedSwirl @nietras1 @compusemble @opinali
1
5
53
10,905
9/10
The incredibly fast and cachemem efficient RT leaf node acceleration design enabled by the parallel low-precision fixed point ray/tri prefiltering pipeline: x.com/MrMPFR/status/19654147…
A comprehensive overhaul of the leaf nodes accell based on these 6 AMD patents. This will speed up ray/tri intersections massively. Seems like AMD is trying to do more with less.
See ALT descriptions for each photo provided 1/5 @opinali @blueisviolet
ALT Low precision parallel ray/tri calc using INT units instead of FP.
ALT DGF format leveraged for low precision/INT triangle prefiltering
ALT DGF format in-depth description
ALT Shared ray/edge testing speedup and could be leveraged with DGF block mass parallel testing for further speedups.
1
4
973
10/10
Can only reiterate what I've said here and elsewhere. Based on everything I'l seen and read so far, all publicly available, RDNA5 is probably AMD's Maxwell or Conroe moment.
It's also an incredibly forward looking µarch that changes everything and wipes the slate clean.
1
13
731
1/11
The Apex of RDNA5, AT0, will prob shatter performance expectations and annihilate the 5090.
@RedGamingTech @RGamerB_ @3DCenter_org @blueisviolet @AMDGPU_ @oliemack @carygolomb @dylan522p @hms1193 @SebAaltonen @RedSwirl @treasureh8nter @firstadopter @nietras1 @3kliksphilip
1/4
Prev RDNA5 L2 thread missed some things and the implementation could actually go even further.
TL;DR: Chiplet compatible by default. Everything except the GCP could be Shader Engine local: HWS (WGS), ACE (ADC), Geometry Processor (GE), and DMA.
Let's tackle them one by one:
2
3
1,175
10/11
Even if Rubin is amazing, RDNA5 is something else entirely, and only properly addressed with clean slate µarch.
So a post-Raster and RT paradigm shift is required to move the goal post and counteract the AT0 threat:
HW Accel for Neural shading, GauRast, DLSS5 Frame warp...
1
2
401
11/11
This probably the most exciting the GPU market has been since the 2000s.
I can't wait to see RDNA5 in action and see what NVIDIA does to counteract it.
Nomatter what we'll probably be blown away by the offerings from both companies so 2027-2028 can't come soon enough.
1
4
280