It would be really interesting if Navi 21 and Navi 22 end up having some minor differences that reflected XSX and PS5. Although could such differences ended up hidden by the drivers?
There was a pretty transparent difference in the GPUs that launched around the time of the consoles, if Cerny's claim that the wider set of ACEs in Hawaii was lifted from the PS4. Actual feature set differences for API-related functions may be papered over by drivers, but would AMD want to add that level of disruption?
Some differences, like varying machine-learning ISA extensions for sub-families have been generally relegated to the specific markets that cared about them.
There were some Sony patents that described dividing screen space up into areas to be rendered at different resolutions to accelerate FOVeated rendering. I recall this being done at the geometry processor level rather than the pixel shader level like VRS. I could see a more VR focused Sony perhaps wanting to spend time customising an offshoot of Geometry Engines to support this, rather than waiting to take a newer geometry processor whose changes they didn't value as highly.
That specific part of the pipeline would seem like a subset of Tier 2 VRS, omitting Tier one and several other options in Tier 2.
Mesh shaders seem like a fantastic concept to me, but then again I have no idea what Sony's capabilities are, and what they might have given up but gained instead.
Sony cited primitive shaders, which AMD has backtracked on since failing to introduce them with Vega and paring their claims significantly for RDNA1. DX12 has gone more in Nvidia's direction, and while primitive and mesh shaders are generally situated in similar parts of the pipeline, their features and emphasis vary.
What Cerny mentioned did seem like it included features that AMD said might be possible with primitive shaders at their introduction with Vega--or some future version of primitive shaders different from mesh shaders.
I'm aware that this is what Anandtech has been told, but it's not what's been happening in practice. These are the roadmaps shown during 2019:
Both Zen 3 and RDNA2 are releasing their first models before 2021.
As such, it's reasonable to expect RDNA3 on 5nm to release before 2022, even if AMD don't want to compromise on that.
In practice, AMD's been burned repeatedly by promising tighter time ranges and having to delay. This era of meeting promises came after a change in policy to be more conservative so that AMD would underpromise and overdeliver, as opposed to the opposite.
Zen 3's release date is November 5 2020, or one respin away from being pushed into 2021.
I'm confused about this talk about PS5 "Geometry Engine" as if it's unique in itself. Is this not a generic term?
Sony does have a history of calling something $Generic_term Engine. AMD's used the term geometry engine for some of its GPUs in the GCN era, but there were multiple such blocks in GPUs like Polaris. A singular block in an RDNA-era GPU sounds like it might be in the same place as what AMD calls the Geometry Processor.
Cerny mentioned primitive shaders in the context of that new functionality, but not primitive shaders as AMD ever demonstrated. It's not clear if they fully correspond to what mainline RDNA would support.