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

Maybe when we get the ISA for RDNA 2 we'll find out what an "accelerated ray query (box or triangle)" instruction actually looks like.
Indeed! Should not take that long and might explain... :)
One of the queries I have about BVHs is whether it's possible to configure multiple, independent, BVHs simultaneously and use them ad-hoc.
The split between top-level and bottom-level acceleration structures might provide some clues here.

I still lack DXR experience, but the concept of top / bottom level AS usually is:
Build high quality tree per model. (Usually only once and offline, but DXR does it on client CPU on model load.)
Build low quality top levels per frame over the set of active models. (DXR on GPU - even for thousands of sub trees this should give no perf problems.)
DXR seems fine here? It only lacks the option to cache bottom AS to disk, which could be added.

One thing is quite ironic to me:
When many people said 'realtime RT is not possible because building AS takes too long', i laughed because i knew the top / bottom split idea had already solved this problem since centuries.
Then when RTX came out, it took me quite some time to realize the problem is suddenly back for open world games.
 
The slide matches the officially stated primitive rate so I'm inclined to believe it. It's curious that the XSX is stated as having one primitive unit per SA though. If the PS5 is the same that would mean the XSX has higher primitive throughput than the 6800 and the PS5 matches the 6900XT!
Navi21 can rasterize 4 primitives/clock as can XSX and PS5. Navi21 will be better at tessellation and some other circumstances though, such as reaching its peak pre-cull rate more often. How did you conclude XSX has a higher throughput that the 6800?
 
AMD to Produce Reference Design RX 6800 XT and RX 6800 Only Until "Early 2021"
AMD is expected to manufacture its reference-design Radeon RX 6800 XT and RX 6800 graphics cards only "through early 2021," according to a tweet by Scott Herkelman, Corporate VP and GM for the AMD Radeon brand, in response to a question by Daniel Rohrpasser. The "MBA" (made by AMD) reference-design cards will undergo production runs only until early 2021, beyond which the company will rely entirely on sales of custom-design boards by its add-in-board (AIB) partners to sell the RX 6800 and RX 6800 XT.
https://www.techpowerup.com/274887/...-rx-6800-xt-and-rx-6800-only-until-early-2021
 
Navi21 can rasterize 4 primitives/clock as can XSX and PS5. Navi21 will be better at tessellation and some other circumstances though, such as reaching its peak pre-cull rate more often. How did you conclude XSX has a higher throughput that the 6800?

The vanilla 6800 has one less SE so one less rasterizer. But higher clocks, and it not known if pre-cull numbers are the same.
 
Navi21 can rasterize 4 primitives/clock as can XSX and PS5. Navi21 will be better at tessellation and some other circumstances though, such as reaching its peak pre-cull rate more often. How did you conclude XSX has a higher throughput that the 6800?

Congrats for the GPU and the console, I was afraid for the consoles after Vega 7 release but happy after RDNA release and extatic after RDNA 2 presentation.
 
Last edited:
Performance figures do not look like double the rasterizers. Thanks to very high clocks in those synthies (slightly north of 2500 MHz), it's on par with 2080 Ti and it's six rasterizers, but slightly less than RTX 3080.

Thank you for Testing CarstenS.
Hmm Strange, here is a quote from a programmer answering my questions on twitter, maybe you have to activate the new pipilen, so the old byond3d suite test is now not working anymore? :


 
Last edited:
The vanilla 6800 has one less SE so one less rasterizer. But higher clocks, and it not known if pre-cull numbers are the same.
I hadn't realized an entire SE was harvested for the 6800. That would reduce the triangle rasterization rate, but Navi21 has enough pre-cull launch capability per SE that it's limited by the Geometry Processor and not the number of SEs in this case.
 
Performance figures do not look like double the rasterizers. Thanks to very high clocks in those synthies (slightly north of 2500 MHz), it's on par with 2080 Ti and it's six rasterizers, but slightly less than RTX 3080.
Which review has these synthetic benchmarks?
 
Performance figures do not look like double the rasterizers. Thanks to very high clocks in those synthies (slightly north of 2500 MHz), it's on par with 2080 Ti and it's six rasterizers, but slightly less than RTX 3080.

@CarstenS Can you maybe post the beyond3d polygon information from 3090 and 6800xt ? This will be very nice. When you reffer the values was this with culling or withouth culling? I can remember that you stated that N10 was better then the 2080ti without culling.

------------------------------------------------------------------------------------------------------------------

Navi21 can rasterize 4 primitives/clock as can XSX and PS5. Navi21 will be better at tessellation and some other circumstances though, such as reaching its peak pre-cull rate more often. How did you conclude XSX has a higher throughput that the 6800?

What is confusing is that Navi21 has indeed 8 Scan/converter (Rasterizer) when you look into the linux and macos driver. That make it strange. Normaly it should rasterize 8 primitives/clock.
Linux:
https://www.pcgamer.com/a-linux-update-may-have-let-slip-amd-big-navis-mammoth-core/
MaxOs:


--------------------------------------------------------------------------------------------------------------

The 288 GPix/s AMD is quoting are Fermi-style.

To clarify: I'm seeing numbers that align extremely close with 4 rasterizers á 16 pixels/clk, only at the end of the pipeline you get 308G depth writes/clk and 612 depth rejects/clk.

This is also strange. Why do you make a backend of 128rops when the frontend can only deliver 1/2 of the pixel of the backend?
 
Last edited:
Congrats for the GPU and the console, I was afraid for the consoles after Vega 7 release but happy after RDNA release and exctatic after RDNA 2 presentation.

Their still behind NV for the most, 6800XT is somewhere between 3070 and 3080 in normal rendering. In RT its very much behind its not even a discussion. DLSS is a whole different matter, as is BW/4k performance which is kinda why people get next gen products in this range i assume. With this were talking about a 6800xt thats north of 20TF's worth of raw power with IC/full rdna2, over double of what console got.
That lands them in the mid range GPU segment and even less so for RT and DLSS, AI/tensor aside.

Like someone else noted, its abit like zen2, AMD comes closer, but still not there. When 2013 consoles launched, they where more competitive on all fronts (GCN vs kepler/7950 vs 670 etc) then today. So yes, i'd had liked to see even more GPU in them for sure, tech is moving very fast still.
Same for Zen2, feels they missed the boat somewhat there, Zen3 is a rather large IPC improvement. You cant have too much CPU this day and age with 120fps and more madness. 16GB will be OK considering SSD's, but same there, had they had more ram, the SSD's could have loaded even more to VRAM at once. I know where talking about consoles which means compromises, but since its being taken up id share my opinion (dream). The 2001 xbox was basically on par with the most powerfull GPU at launch, same for 360 and PS2. Today we dont get that anymore :p
 
Is NVIDIA still bringing more FE's? Since FE is the only 3070 ever sold at MSRP, AIB cards are way higher, and it's FE was so limited only few select NVIDIA storefronts had any to sell (not betting my head on this, but I think it was mentioned somewhere only two European NVIDIA stores had any)
AMD has confirmed to produce references 'till sometime in Q1/21 and is expected to re-stock their own store too (which sells at MSRP guaranteed)
 
3070 brutally killed:

https://www.techspot.com/review/2146-amd-radeon-6800/

8GB of cost-free memory.

They use a good selection of the latest games in their benchmarks. Generally speaking it seems that games releasing on the new consoles perform relatively better on RDNA2 than "last gen" games, potentially to the point of the 6800XT being faster than the 3080 on average, at least at 1440p and very possibly at by future preferred resolution of 3440x1440p. I expect that trend to grow stronger over time as games come to take more advantage of RDNA2.

That combined with the additional VRAM and cheaper price are almost enough to seal the deal for me to go with a 6800XT. However RT and DLSS are the two big wild cards.

RT - I know many will say it doesn't matter that much but if the current trend holds then a very high proportion of games released on the next gen consoles (aka, pretty much all games) will feature some form of RT from now on. So in a way, looking at non RT performance today is like looking at performance without Antialiasing at the start of the last generation. i.e. largely irrelevant. And in that measure, if looks like Ampere may still have an advantage even when developers start to specifically target RDNA's RT hardware in the consoles. Then again, given the performance hit it may be that it gradually falls out of use on consoles with the current run just being a dash to include it as check box feature. That will limit it's applicability on the PC and nullify much of Nvidia's advantage here.

DLSS - I have zero, I mean precisely zero confidence that AMD will release a MLSS solution that is comparable to DLSS in quality or performance any time soon. Probably not before I'm looking to get my next GPU. Meanwhile DLSS itself, while not widely used, is not really niche either, and does feature in quite a few big name titles including the must have CB 2077. And despite some claims to the contrary, I think that where it's available it's a no brainer to use and is at least comparable to native resolution. I'd previously expected Nvidia to release a game agnostic version of it which would have essentially been game over, but I'm doubting that to be the case now, so it's relevance to me hinges on it's take up.

So in conclusion, I'm torn and don't know which to get!
 
Navi21 can rasterize 4 primitives/clock as can XSX and PS5. Navi21 will be better at tessellation and some other circumstances though, such as reaching its peak pre-cull rate more often. How did you conclude XSX has a higher throughput that the 6800?

I was assuming it only features 3 active Shader Engines and so 3 primitive units. I know there's no confirmation of that yet but I think we have had 96 ROPS confirmed which would suggest the entire 4th SE might be disabled. Happy to be told I'm wrong though!
 
Much increased polygon rates notwithstanding: What do those have to do with the amount of rasterizers? They are tested separately.
I thought polygon rates means how many polygons can be transfert to pixel by the rasterizer (scan converter)? Am i wrong?
 
Their still behind NV for the most, 6800XT is somewhere between 3070 and 3080 in normal rendering.
Well yeah, around 5-8% slower in 4K, on par at 1440p and dusting it below that compared to a 3080 in rasterization (depending on the test suite it could also be fastr) can be also called "somewhere in between" but I think you took the meaning quite too far. Also ignoring lower power, cheaper, and with more VRAM than both competitors. I dunno about 10 Gbytes, but 8 Gigabytes of VRAM will soon be a limit.
 
What is confusing is that Navi21 has indeed 8 Scan/converter (Rasterizer) when you look into the linux and macos driver. That make it strange. Normaly it should rasterize 8 primitives/clock.
Linux:
https://www.pcgamer.com/a-linux-update-may-have-let-slip-amd-big-navis-mammoth-core/
I suspect there's some subtlety with those defines. The front end of the scan converter matters for triangle throughput, but in order to scale the pixel rate maybe the back end is implemented differently. ;)
 
Back
Top