AMD Radeon RDNA2 Navi (RX 6500, 6600, 6700, 6800, 6900 XT)

I have no miracle view where i can see whats behind anonym names. But i have checked his webpage, so i think he is right Sorry but for me this is a little bit confusing.

@OlegSH
In the twitter post i wrote that also the "polygon" thrououtput went up 80% but the front end look simmilar like navi10 with 4 Rasterizer. Also it is confirmed by @3dcgi that it is 4 rasterizer
So how will convert 80% more polygons into pixels when the rasterizer count keeps the same?

 

It can have more HW triangle cull units per rasterizer or it can simply cull more primitives in Primitive Shaders due to more CUs in navi21

 

@OlegSH the cull units are the same like Navi10 when you belive in the slides of AMD with 8 poly/clock. So it should be realy have to do with primitive shaders.

It will be nice if @CarstenS can give use some Beyond3d Suite values about the polygon output. Will be intresting to see how id compared to ampere with its 7 GPC / Rasterizer

 

Did we ever really understand this?

PARALLEL MICROPOLYGON RASTERIZERS - Advanced Micro Devices, Inc. (freepatentsonline.com)

Can anyone find a tessellation benchmark comparison of RDNA 1 and 2? Preferably with the driver's "balanced tessellation" setting disabled so that there's no restriction on tessellation factor?

 

I ran some tests comparing a 3090 to a 6900 for a tesselated height map landscape with ~30 million triangles on screen.
When viewed from above looking down, with all triangles front facing: 10 GTris/s versus 8 GTris/s for 3090 versus 6900
Wen viewed from ground level, with about half triangles front facing: 12 GTris/s versus 17 GTris/s

 

So the situation is reversed here? The 6900 becomes faster?

 

Yes, rasterizer and cull rate seems to be the same on the 3090, 7 per clock.
On the 6900 cull rate is twice as rasterizer rate 8 cull (and 4 rasterize) per clock
(also clock on the 6900 is about 25% higher as 3090)

 

This is strange for Nvidia. They have Culling in Polymorp Engine which is 0.5 triangsl/clock. I think 3090 have a lot more polymorh engines?

 

Nvidia does have a higher culling rate than rasterization so there must be something about the workload that drops the 3090 from its peak rate.

 

I think NV varies the rate of geometry processing per SM over different SKUs, both for TDP reasons as well as product segmentation. For example, GM204 rate is 2 clocks per triangle, while its bigger sibling GM200 is slowed down to 3 cycles. These rates could be different for the professional models even further.

 

Almost as good as my golden sample Reference 6800XT + EK + 480mm EK + 360mm cheap rad
I don't know why not a single reviewer managed to hit over 2710MHz on GPU core yet, especially on 6900XT's as they have slightly higher power budgets and voltage allowances ... Need to buy one just to test, but not paying stupid price for it, so have to wait

 

The good news.

The bad news:
https://wccftech.com/tsmc-using-water-tankers-for-chip-production-as-5nm-plant-faces-rationing/.
TSMC is running out of water.

 

Radeon RX 6700 XT would have a starting price of 479 USD and see better availability (guru3d.com)

 

40 CU / 96 MB Infinity Cache confirmed on stream

 

AMD is bringing Smart Access Memory’s frame rate boosts to Ryzen 3000 processors

++

 

It has actually been available for quite a while already, even though officially it wasn't supported before.

 

I think this was the first time AMD talked about ray tracing in more than a passing comment inside a slide. They seem pretty confident of their Resident Evil 8 implementation.

 

Thumbs pressed for better availability than other ... desireable models.

 

My guess is "4K" (Dynamic resolution) with RT shadows and very low setting RT Global Illumination...the hardware is not "magical" and we kinda already know it's clear limitations.