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

With no volume right now, it's complicated to gain anything. They don't take advantage of nVidia paper launch (worse, I've seen people shop for a 6800xt, but leaving with a 3080 because they start to be available here and there).

True, supply from both IHVs is woefully inadequate to meet the current demand. My comments should be taken as prognostication over the life of the products in question.
 
Get the 50 dollars more 6800XT, a north of 20TF GPU.
I prefer to get a custom 6800 than a reference(if that's even possible) 6800XT



RX 6800 is the better card (except rt & dlss), but I would wait for reviews on the 3070 Ti. They say based on lessor GA102 chip, priced the same as 6800 but will have 10GB.
Yeah the problem is 10GB is not that much better than 8...If it were 12 that would be something. I wanted to get my new card on December for CP and the 3070TI looks to be for the next year.
 
I'm going to give it the old college try on the 25th when they said the aib cards are coming. If not then it will be a 3080 or 6800xt whatever one i can find first.

The majority of my gaming in 2021 and 2022 will most likely be tradtional rendering but more and more will most likely start to be ray traced so i think i can take the chance on the 6800xt.

however i hope they have a dlss like solution to show off soon. I think that was the biggest mistake of the launch
 
so for scan converter you get: 4x2x1 = 8 scan converter

at macos10 the leaker did a failure by transfaring the data, he confirmed macos 11 is right. Its the same like in the linux driver.

rasterizer.jpg

8 scan converters doesn't match 8 raster units, as there are 4 on the diagram, so 4 scan converters.

Since there are 4 SEs for Navi21, num_sc_per_se = 1 looks to be the correct entry which gives 4 SCs, rather than the num_sc_per_sh = 1, which gives 8 SCs.

And what are 4 Packers per Scan Converter?
 
however i hope they have a dlss like solution to show off soon. I think that was the biggest mistake of the launch

I wouldn't describe it as a mistake. I'm sure they would have loved to show something if it were possible. The reality is that MLSS is extremely hard to do. God knows how long Nvidia with all their expertise in ML and the tensor cores to rely on were working on DLSS before it launched. And even then it took another year before it was useable.
 
8 scan converters doesn't match 8 raster units, as there are 4 on the diagram, so 4 scan converters.

Since there are 4 SEs for Navi21, num_sc_per_se = 1 looks to be the correct entry which gives 4 SCs, rather than the num_sc_per_sh = 1, which gives 8 SCs.

And what are 4 Packers per Scan Converter?

no it’s not that’s the funny thing in all drivers you have 8 scan converter.... Linux driver have the Same values. 4 was e mistake from the leaker. Also navi10 should have 2 rasterizer if we compare.

maybe diagram is wrong?
 
no it’s not that’s the funny thing in all drivers you have 8 scan converter.... Linux driver have the Same values. 4 was e mistake from the leaker.
Well, @3dcgi mentioned Navi21 rasterises 4 triangles per clock, so how do you get that with 8 scan converters?
 
Full flexibility possibly comes with an hefty price tag since BVH traversal can be highly irregular and doesn't map well to SIMD cores.


There's no contradiction. It means that while FF HW is performing some ray/box or ray/triangle intersections (which are part of the traversal process) the shader cores can be doing something else, like shading (same as texture mapping). Also they might be devoting some threads to traversal only while others are responsible for shading. Since traversal is partially done in SW there are many ways to skin that cat..

To add, and while the BVH might have a flexible structure you're still limited to box tests. Which is severely disappointing. They could've put in support for sphere, tetahedron, etc. Faster testing or faster refits, both have advantages so you're not limiting your characters models to T-Pose (Miles Morales joke har har).

Not that I think it matters much. Even on Nvidia (3080) just shadows rt in Rise of the Tomb Raider drops performance by 25%. For what, slightly better shadows? Raytracing wasn't worth the performance cost on Nvidia already, and it's even worse with AMD.

PS: If there's a silver lining for AMD it's that next year's refresh has opportunity for a few easy gains. Clock the cards up to what they could do assuming yields get better, and slot in 18-20gbps GDDR6 to see 4k improvements. Otherwise, given the choice between a 6800xt and a 3080 (assuming both are in stock) I'd go for the latter. RT might be rather dead endish and vastly too costly, but it's still there and combined with 4k performance for $50 more? Well.
 
@j^aws
It make no sense. Rasterising now polygons which are 6x6 pixels dimension make no sense when more and more polygons getting smaller than one pixel.
 
I wouldn't describe it as a mistake. I'm sure they would have loved to show something if it were possible. The reality is that MLSS is extremely hard to do. God knows how long Nvidia with all their expertise in ML and the tensor cores to rely on were working on DLSS before it launched. And even then it took another year before it was useable.
and like i said , i think its a mistake they didn't have anything ready even a preview working with one of their partner games or in conjunction with microsoft
 
I think the 128 is a number for retiring pixels, i.e. RB+ throughput.
I've just realised something that is perhaps obvious to others: VRS turns a single rasterised (EDIT: and shaded) fragment into 2 or 4 pixels in the render target.
 
Last edited:
@Frenetic Pony Most of the games released with DXR do not support DXR 1.1, which has some significant changes. I expect there to be some changes in how RT is used and how it performs going forward, especially as console devs dig into it and it's not just relegated to niche hardware. I'm sure the performance hit will remain large, but adoption should be pretty widespread.

I'm actually curious to see if there are different performance deltas between AMD and Nvidia as more games adopt DXR 1.1, which AMD seemed to be heavily involved in.
 
@j^aws

that is the question!? Something is strange. @CarstenS said that there are no improvements between navi10 and navi21 in rasterisation, but the driver says different. (This was wrong information from me sorry.

I thought always that scan converter is always a rasterizer, maybe the definition is wrong. Maybe scan converter + other units is a rasterizer. But what are the other units?
 
Last edited:
@j^aws

that is the question!? Something is strange. @CarstenS said that there are no improvements between navi10 and navi21 in rasterisation, but the driver says different. I thought always that scan converter is always a rasterizer, maybe the definition is wrong. Maybe scan converter + other units is a rasterizer. But what are the other units?
I'm thinking some algorithm with 2 scan converters that can be overlapped and their difference is a smaller triangle, and as small as 1 fragment scalling coverage up to 32 fragments per raster unit per triangle. Just speculating.
 
Back
Top