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

So big discrepancy between AMD's internal pre-launch benchmarks and actual performance results from 17 reviews. No surprises that review dates and launch date coincided.

https://www.3dcenter.org/news/radeo...ltate-zur-ultrahd4k-performance-im-ueberblick
I just browsed quickly through those, only one site (Igor) used Ryzen 5000 -series CPU in their tests, everyone else used either Ryzen 3000 or Intel Core (all the way down to i7-8700K (also no pcie4 here, even if the difference is minimal)).
That can easily attribute (some of the) differences
 
Its a bit of a zen2 moment. Almost there but not quite. I'm looking forward to next years navi 3, but they could have beat the 3080 completely if they used gddr6x. Infinity cache mitigates, but doesn't match.

It does compete quite well at 1440p at least. Not everyone who buys these cards uses a 4k monitor. I am curious how much memory overclocking would have helped though. As per the reviews, they couldn't overclock memory yet. Higher core clocks would also increase infinity cache clocks. Could lead to a significant improvement.
 
What I'm not worried about though is the poor hardware raytracing performance. Hardware raytracing is good for coherent rays and that's about it anyway. In fact it's primarily good for sharp reflections and that's it period. Even shadows are questionably useful at best since you have so much content restriction with it (severely limited movement, environment detail, etc.) EG the Demon Souls remake doesn't use a whiff of hardware raytracing, and looks better than either Watchdogs Legion or Miles Morales. It feels typically weird to watch Digital Foundry do their "the difference is obvious" thing as at least half the time I'm narrowly squinting at the glass in a game hoping to see what's so obvious about it.

That being said it is a victory for Nvidia PR wise. This is what AMD gets for letting their rival control the narrative of what's "important" in GPU tech.

If you're struggling to notice the difference between raytraced reflections and cubemaps in the DF Legion videos then raytracing is definitely not for you. The difference is astounding.
 
It does compete quite well at 1440p at least. Not everyone who buys these cards uses a 4k monitor. I am curious how much memory overclocking would have helped though. As per the reviews, they couldn't overclock memory yet. Higher core clocks would also increase infinity cache clocks. Could lead to a significant improvement.

And this is not horrible at 4k and the card is 50 dollars less than 3080. This is not the huge RT disparity.
 
It does compete quite well at 1440p at least. Not everyone who buys these cards uses a 4k monitor. I am curious how much memory overclocking would have helped though. As per the reviews, they couldn't overclock memory yet. Higher core clocks would also increase infinity cache clocks. Could lead to a significant improvement.
Huh? There's no issues with memory overclocking other than the fact that the chips would probably clock higher than AMD allows (17.2 Gbps is max in Radeon Software I think)
Infinity Cache works it it's own clock domain, 1,4 GHz base 1,94 GHz Boost. It's possible that it's somehow tied to core clocks, but I haven't seen any confirmation of such.
AMD did state however, that the benefits from Infinity Cache grow the higher your core clocks are.
 
Huh? There's no issues with memory overclocking other than the fact that the chips would probably clock higher than AMD allows (17.2 Gbps is max in Radeon Software I think)
Infinity Cache works it it's own clock domain, 1,4 GHz base 1,94 GHz Boost. It's possible that it's somehow tied to core clocks, but I haven't seen any confirmation of such.
AMD did state however, that the benefits from Infinity Cache grow the higher your core clocks are.

Oh my mistake, I read through the section hurriedly on TPU and read it as they weren't able to OC the memory. Just re-read it and saw that they were in fact able to OC it to the max allowed (2140 Mhz). Do you happen to know what the benefits are purely from memory OC without any core OC?

Yea I saw that slide as well, they stated the potential benefits of the cache grows with clock speeds so it must be tied to clocks in some way.
 
Oh my mistake, I read through the section hurriedly on TPU and read it as they weren't able to OC the memory. Just re-read it and saw that they were in fact able to OC it to the max allowed (2140 Mhz). Do you happen to know what the benefits are purely from memory OC without any core OC?
No clue, but asked if our testers have spare time to check how the cards act when only core or mem is OCd and the other left at stock
edit: don't have tests to show, but apparently memory clocks do benefit more than gpu clocks, especially when you enable fast timings for the memory (yep, there's now different timings for gpu memory too)
 
I don't understand why people get angry about there not being enough stock. It's the same thing every time there is a major new product launch.

Limited manufacturing capabilities aside, it just makes no sense to delay a product launch for weeks/months and spend all that extra money just to satisfy the launch crowd.
 
As I understand it, DXR specifically obfuscates the data format of the BVH. It seems the intention here is that each IHV can optimise the data format to match the way the hardware works.

So, for example, perhaps inline ray tracing (DXR 1.1) is preferred on AMD and the BVH data format is optimised for that.

Sure, but the question is how much this obfuscation is necessary. Let's make a speculative example about 'Add RT support to UE5 geometry.'
And let's assume AMDs RT support only means additional instructions to intersect boxes and triangles (they never mention 'traversal' in their marketing).
If so, Epic can reuse their existing data structures, which probably have hierarchy and LOD, and eventually work for RT. With HW support for both consoles, plus AMD GPUs on PC, that's a nice option to have. And no RT API is necessary at all.

To support RTX, a lot of problems come up:
Streaming BVH is not possible. So they have to build BVH on CPU each time they load a model, with little control over performance costs.
LOD is not possible either. They can not just load the levels of hierarchy that are necessary. They have to use discrete LODs, and build BVH for each level, requiring a lot of extra memory, with little control over how much of it.
In practice this means on PC you need more memory and CPU cores to have RT.

We really want to know precisely what the situation is.
 
We really want to know precisely what the situation is.
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.

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.

It's amazing that after a couple of years devs still appear to be stuck reading the minds of the DXR architects. Devs being treated like 5 year olds. Sigh.
 
The "error margin" between the tests, so to say, is astonisingly big (it was big for zen3 too, but I explained it away with different memory settings, although it seems zen3 and skylake scaling wrt memory clock/latency is pretty much the same), if you compare, for example Igor's results with TPU results it might look like they were testing GPUs from different price brackets.

Also, clocks are all over the place - some reviewers claim they've got N21s boosting upwards of 2300, while others are saying it's barely above 2050. I wonder if performance difference is related to that or it's simply a monitoring bug (which always happens at new launches of ATi gpus for some unexplicable reason). Judging from early OC attempts by reviewers, it's either broken or they still don't understand how to OC AMD chips (basically no one among them really understood how to OC vegas properly, while it's easy to get about 25% boost with V56 (FS GS 20k->25k ish) if you actually observe and learn what happens when you move the sliders up and down in Wattman or whatever utility you use)
 
Back
Top