-
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.
-
We have made minor adjustments to how the search bar works on ResetEra. You can read about the changes here.
AMD Big Navi VRAM specs leak: 16GB Navi 21 ($499) and 12GB Navi 22
- Thread starter Deleted member 44371
- Start date
- Rumor
I'll wait to see what they have to offer, but Nvidia always had better GPUs and this is coming from someone who had mostly AMD gpus.
People seem to have the idea stuck in their head that every game needs to be maxed at 4k and run over 60fps to be considered a 4k card. Even today that's not even possible with the 3080. Linus showed that Crysis remastered runs like crap with everything cranked up at 4k.
The games won't be GPU bound for doubling the frame rate but they might be CPU bound as I don't think PCs will have CPUs with double the single-thread performance of PS5/SX
Wouldn't that be a bit too low to compete against the RTX 3070?
I based those numbers on the rumored die sizes of those RDNA 2 chips:
Navi 21 - 505mm2
Navi 22 - 340mm2 ~ like the XSX GPU
Navi 23 - 240mm2 ~ like the PS5 GPU
And that's an excellent illustration of why too. Often times "Ultra" is just, we are throwing everything at this with minimal or no optimization, good luck to you!.
With the Crysis remaster it's literally turning off the LOD system entirely. Something utterly unoptimized by it's very nature. You literally cannot discern all of the extra detail from the best LOD model from even a few meters away form the camera.
A lot of Ultra features only offer minimal improvements for a LOT of performance trade off. High is usually where it's at. Of course there are exceptions, but this is usually what happens.
This is objectively not true. There's but several generations where AMD/ATI had better raw performance, better power efficiency, or both. They haven't been on top of the performance scales since the 290X but that doesn't change the facts.
Which it is lagging behind on based on these specs even, GDDR6 has a lower clock speed and the memory bus is much narrower. As has been explained many times, it'll be a while before 10GB of VRAM capacity becomes an issue, much less so if you're rendering less than 4k.
Yeah I don't care about Ultra settings anymore, stopped caring about them years ago. I usually do High and sometimes Medium since I value framerate over graphics. Heck for me it's Framerate>Resolution>Graphics.
I know DLSS is a huge initial advantage but I'm not confident counting AMD out yet. Especially considering how rushed the Ampere launch was. Waiting isn't going to cost me anything. There will be other machine learning temporal AA solutions in the future. Nvidia isn't the only one with a huge stake in that venture. AMD's cards tend to age pretty well, with the 9 year old 7970 still capable of running Doom Eternal at 1080p/60fps. Strong Linux / Vulkan performance is a nice plus. And calling the match before the fight even started is super boring and makes for boring threads.
But if AMD can't beat the 3080 at a lower price then that's probably game over.
But if AMD can't beat the 3080 at a lower price then that's probably game over.
You are probably right, but this was always my impression, probably because Nvidia products were always more expensive.
AMD are giving the chips codenames based on which came first. The most powerful chip was developed first so it is Navi 21. Any less powerful variants came later so you have Navi 22 and 23.
Even with the knowledge that AMD has planned to undercut Ampere via pricing (whatever the performance ends up being, all evidence shows they will be more power efficient GPUs), $499 seems too absurdly good to be true if this rumor reflects actual performance.
Intel does the same thing with its CPUs despite them only being better in (some) games, and only a few percentages at that, while losing soundly in all other categories.
Last edited:
499 isn't the top card so i wonder what they price that one at 600 maybe?
Pretty sure the 6900XT is the top-end card, aka. Navi 21.
I am skeptical of that being the price, in any case, especially if they're using HBM2. That's pretty expensive compared to GDDR6, last I recall.
True, so long as you are pushing grpahics settings and resolution you should comfortbaly beat out a console. As for CPu performence, I would not say never, over the next 7 years.
I see a lot of people on the Internet (not necessarily in this thread) cheer on AMD and hope they bring out incredible new products, and then immediately follow that up with "so it will force Nvidia to lower prices. I would never buy an AMD card personally" or some variation of that thought. Oftentimes when pushed they will cite bad drivers as the reason why.
It seems to me that AMD has a severe image problem that they need to overcome with regards to their software. Even if they were somehow able to deliver a compelling card this time, it appears to me that a lot of people would not give an actual purchase a second glance. Do you think that is mostly hot air and people would buy an AMD GPU if it were to somehow best Nvidia on price/performance?
It seems to me that AMD has a severe image problem that they need to overcome with regards to their software. Even if they were somehow able to deliver a compelling card this time, it appears to me that a lot of people would not give an actual purchase a second glance. Do you think that is mostly hot air and people would buy an AMD GPU if it were to somehow best Nvidia on price/performance?
The last rumor about this said 499 is the cut down navi 21 idk how tweaktown came to the conclusion it was the 6900xt, it seems to not be the case
Maybe? It depends on how the drivers work out this time and if supply issues remain for Ampere come October/November. The 5700XT is arguably the best AMD launch since the 400/500 RX series, and even it took a few months to work out the driver issues (I haven't had any issues with it but there's plenty of stories of people suffering black screens or having to switch from HDMI to DisplayPort and vice-versa as one or the other was having compatibility). To say their GPU division has an image problem is an understatement, but the same could have been said for their CPUs during the Dozer era. Personally, i'm going for whichever one gives me the best bang-for-buck at the $1,000 CAD price point (Don't care for raytracing and i'm assuming AMD will have their own DLSS equivalent). Assuming this performance is reflective of Navi 21, the main issue I can see for consumers is that the quality partner cards (Sapphire, PowerColor) might not be out until 2021.
Do you have a link for the latter i.e. cutdown vs. the full Navi 21?
I will look for it but rogame the person the article is based on even says he's not sure which variant those vram makeups are for
I think they only have a chance to compete at pure rasterization performances. The thing peoples don't quite grasp yet with RDNA 2 architecture is that rasterization, RT and ML all compete for the operations on the CUs. Nvidia's solution is to have RT & ML dedicated.
The thing also with DirectML is that it provides an API for implementing and running neural network models. It won't provide the models.
DirectML gives a way to stack conv2d layers with ReLU layers to form a net and that's it. You have to find a model that achieves good upsampling and they will have to provide the weights that allows the AI to learn. You'll get as good of an upscaler as the model allows, and it can easily go from DLSS 1.0 (horrible) to unlikely DLSS 2.0 (god-tier). But it's a recipe not many peoples in this industry are familiar with and Nvidia as been at the top of the AI game with an head start of 13 years now with almost all deep learning being done on CUDA cores in the industry/universities and it still took them like 2 years to find this recipe. They have been in close collaboration with Microsoft since 2017 on directML and most of the DX12u ML's logic are straight out of Nvidia papers, directML is RTX codified for an API.
Also since directML is an API regulated by Microsoft, it makes more sense that Microsoft will provide a base model for developers to ML their games for upscaling than AMD providing a solution, because the advantage of the API is to remove the need to write vendor-specific code paths. DirectML will be hardware agnostic, in fact even a CPU can run directML, any DX12 GPU in fact, i can't even recall the GPU series' name since it's so far back in support. RDNA2 of course has accelerated ML over these older GPUs.
So yea..Microsoft will probably be the one proposing a model, but when? It might take a long time. I'm pretty sure AMD will have a checkerboard + AI sharpness (ala FidelityFX) upscaling in the meantime, but i seriously doubt they'll invest a lot of R&D and time on a model that will ... be vendor-specific and before Microsoft gives an universal model? What would be the point?
Last edited:
AMD is going to have to pull a Ryzen with RDNA2. Be competitive and a decent amount cheaper on this gen. If they are only competitive and similar in price, no one will switch. Well not many.
AMD really need to go big to try and gain market share. Nvidia already have watch dogs legion and cyber punk locked up with rtx features.
OP
OP
We've seen AMD cards with vram advantages get smoked by Nvidia cards in actual performance in the past. I'm gonna need to see a lot more...
They're actually making a 6900? I thought they would stop at 6800 this gen, just for...reasons.
The numbers are just codenames. Generally the first on a new tens digit is the full version and later numbers are cut-down versions.
Navi 14 is smaller than Navi 10, etc.
The numbers are just codenames. Generally the first on a new tens digit is the full version and later numbers are cut-down versions.
Navi 14 is smaller than Navi 10, etc.
Depends on how high its clocks will be really.
No. RDNA architecture doesn't have such option.
Tweaktown seems to be underplaying these as the source tweet says
Also I though Navi 21 was expected to be at least 384 bit (and even that was cutting close), wouldn't 256 be a disappointment?
But yeah 499 for ~3080 at least sounds nice if it pans out. Still over a month to learn the truth :(
Also I though Navi 21 was expected to be at least 384 bit (and even that was cutting close), wouldn't 256 be a disappointment?
But yeah 499 for ~3080 at least sounds nice if it pans out. Still over a month to learn the truth :(
Console TFLOPs count does not indicate this. No reason to believe that RDNA 2 from consoles to PC will drastically differ in architecture.
In fact, it seems even AIB for the 3080 did not know about teh 2x shader cores as some of them had half the core count on their leaked specs before the Nvidia conference. It was a wild card.
See my post earlier.
Makes no sense for AMD to compete against Microsoft for a directML model for upscaling.
I'm not too worried about 10Gb but in general I cant agree. It sneaks up on you. Like a 6GB card right now IMO would be problematic.
One of the best purchasing decisions I ever made was to buy a HD 4890 with 1GB RAM rather than the 512 GB everybody online told me was fine at the time (by the time you need more the card will be too weak anyway, was the common argument). The 1GB FB made the card age so much better. Hell I tried to dig it out not too long ago, I had no GPU (sold mine to miners for a nice price) and needed something to play Destiny 2 PC, it would have still been great for that! turned out I forgot I had actually thrown it away :(
My mindset is never buy a card anywhere close to RAM limited. Again though, subjectively, 10GB feels ok. But then again there is next gen consoles...
If I was spending $700 on a card I'd probably want more than 10GB, but that's me. Some people buy cards a lot more often.
This would be a legit worry if DLSS was available by default in all games, by the time this becomes the norm I'm sure AMD will have a similar option.
The thing is that Nvidia isn't just offering a video card. They are offering services specific to their hardware like gsync, DLSS, Broadcast, Reflex, Ansel, etc as well as game ready drivers day one on big releases. They also partner with many big games to add Nvidia specific features like RT and DLSS. If you look at the latest Steam survey it's very clear that AMD offering comparable hardware at a comparable price isn't working for them. AMD needs to either step up their services or offer significantly better performance at a much better price point then Nvidia to make any head way. As long as Nvidia is seen as the premium brand most consumers won't see a reason to buy a comparable AMD card over it's Nvidia counter part.
Last edited:
Maybe at the top end. In general AMD is never behind in price/performance (usually they are a little better at it, because people prefer Nvidia brand and need an incentive to buy AMD). See EG, 5700XT currently. Well, in pre 3000 series landscape...
256-bit bus for Navi 21 ?????????
Yeah if true it's a little bad.
They are rumored to have some super mondo 128MB on GPU cache breakthrough tech to mitigate that, but still.
IF true...
Bit bus isn't everything. And if they found ways to get more out of it with making the gpu cheaper then im all for it.
The VRAM amounts are enticing, but those memory bus sizes do seem a tad on the small size. Hopefully AMD really puts the pedal to the medal on RDNA2. They really need to be competing with the RTX 3080 and not just the RTX 3070.