They are, especially if you can use DLSS without compromising the IQ or sacrificing too much in the video settings.
IMO the 3080 is more a really good, high/ultra settings 1440p card.
-
Ever wanted an RSS feed of all your favorite gaming news sites? Go check out our new Gaming Headlines feed! Read more about it here.
AMD Big Navi VRAM specs leak: 16GB Navi 21 ($499) and 12GB Navi 22
- Thread starter Deleted member 44371
- Start date
- Rumor
IMO the 3080 is more a really good, high/ultra settings 1440p card.
They need to be competing with the 3080 and the 3090. While the latter is more of a workstation card than a gaming GPU, especially if those recent leaks are a best-case scenario vs. the 3080, them not having a Pro Vega II equivalent with Big Navi for OEMs to adopt will result in them ceding a large chunk of the market to NVIDIA.
I'm talking about actual performance differences in games benchmarked, which is all that matters, and 3080 vs 2080 Ti. This is based on one source though:
GeForce RTX 3080 Ray Tracing & DLSS Performance, RTX ON vs OFF
AU Viewer MSI B550 & AMD Ryzen Deal: http://msi.gm/PCCGTomahawk3600DealMSI B550 Motherboards: http://msi.gm/B550TomahawkMotherboardJoin us on Patreon: https:...
youtu.be
"the gap between RTX on and RTX off (between 3080 and 2080 Ti) is just 10%". The performance improvement is so small because in some games, there is no difference (same percentage drop in performance by turning RT on) at all in the games they tested between a 3080 and 2080 Ti.
Sigh.
A model (software) which is running on DML would "compete against MS" in the same way a game running on DX12 would compete with DX12.
This has zero meaning for RT h/w as these games aren't limited by that.
That video has all of its conclusions wrong.
The point is, in games tested, the difference in RT performance/difference in performance impact from RT between Turing and Ampere is very small.
I do not think that is the case at all. Ampere handles RT better than Turing.
You merely have to look at the performance differential between the 2080 and 3080 in a fully rasterised game vs. a path traced one to see that.
Welp, I do say it's possible AMD has a model.. but why. The purpose of using Dx12u from a developer's perspective is to call the instructions and be hardware agnostic. You think they would like to have to train an AI on an AMD model and then a model on Nvidia and then Intel...? Makes much more sense that Microsoft gives a model.
It would never be used. The game developers also would want the DirectML to be hardware agnostic
Performance impact comes from additional shading which is required to incorporate any RT calculations.
Shading is done as shading is usually done and has no relation to RT h/w.
When RT h/w is the limiting factor Ampere is considerably more faster than Turing in comparison to workloads where it's not.
So again, their conclusions in this video are all wrong.
It doesn't matter how you train the model for inferencing. You only need compatible and capable (in performance) h/w to run the latter, its vendor is irrelevant.
DML is h/w agnostic.
Again, don't confuse the API - which is DirectML - with the software which runs through it. DLSS is the software, NGX (which it runs on currently) and DML are APIs. To have something like DLSS AMD must develop this software - or use someone else's software instead which would run on DML which they will support with enough features and performance to be compatible with it. MS could make such software, sure. Will they though and how well will it run on GPUs without dedicated ML h/w is a different question.
Last edited:
I've heard that AIB partners can only launch with Navi 22 products this year, while AMD will exclusively launch the Navi 21 products.
Is the closest thing I've heard so far. Prices can fluctuate until they actually announce it, just like Nvidia.
Exactly, the only way to make this claim is to check a game that only uses raytracing.
https://www.eurogamer.net/articles/digitalfoundry-2020-nvidia-geforce-rtx-3080-review?page=6
So you have a 47% gain in RT performance here. Both the 2080Ti & 3080 have the same amount of RT cores and I think more or less the same effective clocks. Of course if you're going use a game where the GPU spends 80% on rasterization & 20% on RT then you're not going to see a lot of performance gain overall even from a 50% bump in RT performance.
Say you spend 80 seconds of a GPU's rendering time on a frame for the rasterization part & 20 seconds for the RT part. What happens when you double RT performance? Well, you spend 10 seconds instead of 20 seconds on it. Overall you'll spend 90 seconds on rendering the frame instead of 100 or in other words you gain about 10% overall performance. For some reason a lot of people are calling that 10% the gain in RT performance but that's just wrong. Weird that a big tech youtuber would make that mistake. Seem like a bad conclusion from HU.
RT (BVH testing) runs in parallel to shading even on Turing so it's even less than that.
You have two workloads running in parallel. One is finishing in 10 seconds and the other in 20. Your performance is limited by the second one.
Now you have new h/w where the first workload finishes in 3 seconds and the second one in 10. Will you see 3X gain or 2X gain?
I would like to see some more tests then. 10% difference in HU tests of actual relevant games is small.
Go look at the work Richard did for the RTX 3080 Review and compare its percentage uptick in a pure path traced game vs. a purely rasterised one. Sm for SM Ampere is quite a bit faster at doing RT workloads.
Fully rasterised/compute game:
Fully path traced/ray traced game:
Go look at the work Richard did for the RTX 3080 Review and compare its percentage uptick in a pure path traced game vs. a purely rasterised one. Sm for SM Ampere is quite a bit faster at doing RT workloads.
Fully rasterised/compute game:
Fully path traced/ray traced game:
Thanks interesting will watch the full vid.
The good thing for those of us that will stay in the midrange is that we at least know that AMD puts out good competition there. I'm quite confident that the waiting game will benefit me since competition may bring down the 3060 price, or put something better for the same price-range so I'll just wait. Excited anyway though
I am still so freaking confused when it comes to Navi 21 specs.
It is 80 CUs and 96 ROPs?
Or 80 CUs and 64 ROPs
Or not even 80 CUs?
It is 80 CUs and 96 ROPs?
Or 80 CUs and 64 ROPs
Or not even 80 CUs?
So if this card has a 256bit bus and 16GB GDDR6 RAM, that doesn't bode well for it being a 3080 competitor at all. It seems like something with as much power as the 3080 would be massively constrained by bandwidth that low.
So if these leaks are true, the 6900XT is either A: Not a 3080 competitor, or B: bandwidth starved compared to the 3080.
So if these leaks are true, the 6900XT is either A: Not a 3080 competitor, or B: bandwidth starved compared to the 3080.
Last edited:
It's possible that the difference in ROPs is the different Navi 21 models i.e. the RX 6900 and 6900XT.
thank you!Go look at the work Richard did for the RTX 3080 Review and compare its percentage uptick in a pure path traced game vs. a purely rasterised one. Sm for SM Ampere is quite a bit faster at doing RT workloads.
Fully rasterised/compute game:
Fully path traced/ray traced game:
According to the firmware data from AMD, N21 has 80 CUs which would be 5120 "shader cores".
https://twitter.com/_rogame/status/1289239501647171584
The amount of ROPs is even, 4 Render Backends per Shader-Engine, so either 64 or 128 ROPs.
It's likely that AMD changed the amount of ROPs per RB and is going with 128 ROPs.
I drew N21 with 2048-Bit HBM2(e), which fits to the amount of L2$ tiles (16 texture channel caches for 16 memory channels).
(On a side note, Arcturus doesn't make a lot of sense with 16)
Otherwise it would be a 256-Bit GDDR6-Interface, where rumours from RedGamingTech have it that "128MB Infinity Cache" would balance it out, which I'm not inclined to believe.
According to the firmware data from AMD, N21 has 80 CUs which would be 5120 "shader cores".
https://twitter.com/_rogame/status/1289239501647171584
The amount of ROPs is even, 4 Render Backends per Shader-Engine, so either 64 or 128 ROPs.
It's likely that AMD changed the amount of ROPs per RB and is going with 128 ROPs.
I drew N21 with 2048-Bit HBM2(e), which fits to the amount of L2$ tiles (16 texture channel caches for 16 memory channels).
(On a side note, Arcturus doesn't make a lot of sense with 16)
Otherwise it would be a 256-Bit GDDR6-Interface, where rumours from RedGamingTech have it that "128MB Infinity Cache" would balance it out, which I'm not inclined to believe.
Thanks!
I see, thank you also.
If the Navi 22 still has 40 CUs, how much of an improvement will it really have over the 5700xt?
I dont believe they'd have a step from 80CU's to 40 CU's with nothing in between. That would be ridiculous. But RDNA 2 could offer: Boost clock at least up to 2.2 ghz of PS5 (vs 1.9 of 5700XT)+if rumors are to be believed (salted), 50%+ IPC gains.
Basically who knows. IMO it could be anywhere between that and above the 3090.
Oh hey, November 11th is my birthday.
All these cards are probably gonna be priced higher than what I'm willing to spend though.
I see the RX 5600 XT coming up on Newegg as a deal tomorrow, if it hits $275ish I think I'll probably just nab that for now.
Last edited:
Yeah, the 5700xt was a midrange part, and that had really rough availability
Didn't Nvidia say they started mass masking the 3080/3090 in August? If AMD already started making the cards and it comes out mid-ish November, they will probably have a much larger stock. Not only that, with AMD being a different chip process, it should not be fighting for any resources with Nvidia.
Yeah, it will fight for resources with Apple, other AMD products (CPUs) and the rest of the industry instead. NV is pretty much alone as a big customer of Samsung's 8nm process.
There are indications that Navi 21 will be in very scarce supply this year, one of them being an apparent lack of AIB cards until 2021.
Also if AMD would have started production already they would've definitely launched it against Ampere, in any quantities. Having even a dozen cards on the market is better if it's able to compete - people would at least know this and hold on from purchasing 3080s.
I don't you should take the word from someone inside AMD lol
Also, while Nvidia's launch was a mess, it wasn't a paper launch. Just too much demand.
I do not think AMD's launch will be a paper one, but the same thing will happen.
Welp im going to assume we can add this tweet to this video retroactively at the end of next month
Get Good: AMD Radeon Marketing Friendly Fire & RDNA2 Competing vs. RTX 3000
Previously, we ran one like this talking about how Intel keeps hurting itself with its marketing. Now, it's time to talk about how AMD needs to reach more th...
www.youtube.com
I do not think there has been much stock of Ampere, depends on what constitutes a paper launch really.