AMD Navi 23 ‘NVIDIA Killer’ GPU Rumored to Support Hardware Ray Tracing, Coming Next Year

[TNA]

Based on what? Nvidia Hopper will MCM, Intel Xe is already MCM, so why would AMD - who were the first to utilise chiplets in CPU design - stay with a monolithic design when it's obvious that we're reaching a dead end with how far monolithic designs can go and still churn out performance uplifts gen on gen.

He just makes stuff up in his head and create threads all over the place to ask everyone what they think all the time.

He mentions them in the correct thread, no one replies so he goes and makes a dedicated thread for it. Lol

That said I don't think we are at a dead end just yet. But I want to see mcm come in asap so we can get those sweet price for performance improvements of the past again. At least for a while.

 

[LoadsaMoney]

That boy be *white*!!!! Think he's going for the same shade as the wall behind him...

 

[Wrinkly]

Will laugh if many of us get AMD high end and it outperforms 3080 for less cash before many of the Nvidia diehards even get their cards lol... will definitely highlight how much of shambles Ampere launch has been.

Looking very plausable. There is a good chance AMD can beat the 3080 in raster performance, but I don't see them getting close to Ampere with RT.

I can see the top AMD card having to do between quarter and half resolution RT on PC depending on resolution. Maybe good enough for fast paced games.

 

[Grim5]

I can see the top AMD card having to do between quarter and half resolution RT on PC depending on resolution. Maybe good enough for fast paced games.

That would be ironic. Mocking DLSS but then AMD ray tracing runs at 480p

 

[Wrinkly]

That would be ironic. Mocking DLSS but then AMD ray tracing runs at 480p

Poor Volta

 

[Grim5]

Spotted these Ray Tracing output specs. Peak numbers, Tri-Rate (green bars) is the important one.

 

[james.miller]

Looking very plausable. There is a good chance AMD can beat the 3080 in raster performance, but I don't see them getting close to Ampere with RT.

I can see the top AMD card having to do between quarter and half resolution RT on PC depending on resolution. Maybe good enough for fast paced games.

Well, you just have to look at what we know so far. well, what we assume to be fairly accurate lol.

- it's a smaller die the than 3080 (unless AMD are hiding a MONSTER from everybody)
- it has no dedicated RT hardware so there will be bigger performance penalties.
- that smaller die also has to house that massive cache it supposedly has, leaving even less silicon to play with.

AMD would have to come up with something special to leap two generations of nVidia RTX performance out of the gate, AND do so with no dedicated hardware. That doesnt look very likely.

 

[Nablaoperator]

Tri-Rate (green bars) is the important one.

Care to explain?

 

[Soapysage]

Spotted these Ray Tracing output specs. Peak numbers, Tri-Rate (green bars) is the important one.

That goes along with what a few others have said, AMD's version of RT is very fast at low/medium quality levels, the ray-box rate, so 'performance' mode as per what Nvidia call it, would have very little hit on fps on RDNA2, however for 'quality' mode, increasing everything up to ultra RT, would cause a bigger hit and drop it down to between 3070 and 3080 levels.

 

[Nablaoperator]

- it's a smaller die the than 3080 (unless AMD are hiding a MONSTER from everybody)
- it has no dedicated RT hardware so there will be bigger performance penalties.
- that smaller die also has to house that massive cache it supposedly has, leaving even less silicon to play with.

- 20% of the 3080 die area is affected by gangrene.. AMD just needs a chip with 22 billion or more transistors to beat that in pure grunt
- dedicated RT hardware exists.. in fact the caching capacity in TMU has been increased.. also AMD can cheat RT effects for greater performance as the approach is more flexible
- the guy who invented the shared cache architecture is talking about 22% IPC increase (its existence on navi is still a speculation)
Overall, i believe AMD has the best approach here. Atleast the silicon will not be gathering dust in non RT games cuz it will be working in mixed mode

 

[five8five]

No, let it speed up. The carnage will be more spectacular that way if AMD fail to deliver

It was before your time here, but the Polaris Hype Train was awesome. Many remember it fondly. Humbug made it very memorable

The Polaris Hype Train was epic. I believe humbug still has the engine driver cosplay uniform in back of the wardrobe. Gibbo told him to never ever put it back on but you never know

 

[Nablaoperator]

That would be ironic. Mocking DLSS but then AMD ray tracing runs at 480p

Do you know if DLSS is applied at the start of the rendering pipeline?
Cuz those comparisons arent making sense..

 

[james.miller]

- 20% of the 3080 die area is affected by gangrene.. AMD just needs a chip with 22 billion or more transistors to beat that in pure grunt
- dedicated RT hardware exists.. in fact the caching capacity in TMU has been increased.. also AMD can cheat RT effects for greater performance as the approach is more flexible
- the guy who invented the shared cache architecture is talking about 22% IPC increase (its existence on navi is still a speculation)
Overall, i believe AMD has the best approach here. Atleast the silicon will not be gathering dust in non RT games cuz it will be working in mixed mode

there's no hardware for denoising is there? it's a hybrid approach. Whereas nVidia use the Tensor cores for that job, AMD are banking on existing hardware and using lower precision rapid-packed math to do the job.

To be clear, i believe a unified architecture is the 'best' way, but it's not a panacea - there will always be drawbacks.

 

[Nablaoperator]

No, let it speed up. The carnage will be more spectacular that way if AMD fail to deliver

It was before your time here, but the Polaris Hype Train was awesome. Many remember it fondly. Humbug made it very memorable

lol.. doesn't look like you're selling insurance here

 

[Soapysage]

- 20% of the 3080 die area is affected by gangrene.. AMD just needs a chip with 22 billion or more transistors to beat that in pure grunt
- dedicated RT hardware exists.. in fact the caching capacity in TMU has been increased.. also AMD can cheat RT effects for greater performance as the approach is more flexible
- the guy who invented the shared cache architecture is talking about 22% IPC increase (its existence on navi is still a speculation)
Overall, i believe AMD has the best approach here. Atleast the silicon will not be gathering dust in non RT games cuz it will be working in mixed mode

Plus regarding die size, TSMC 7nm is denser than Samsung 8nm. So although it's smaller, it could have more transistors, using Navi 10 as an example, if Navi 21 were to be the same density as Navi 10, it would have about 22 billion transistors, less than the 3080's 28 billion, however if we take note of Nvidia's A100 die, it also uses TSMC 7nm but a more mature version, if RDNA2 was to be closer to it's density, Navi 21 with die size of 536mm2 could have 35 billion transistors. So until we know for sure, we can keep speculating all day long.

 

[KentMan]

Spotted these Ray Tracing output specs. Peak numbers, Tri-Rate (green bars) is the important one.

So if by some weird quirk that graph was actually true, you are mocking AMD's first attempt at Ray tracing was actually inbetween their 3070 and 3080, and better than Nvidias first attempt? lol so its ok as Nvidia get a pass on their first attempt but AMD would be ridiculed for it?

To be perfectly honest if AMD got between 3070 and 3080 Ray Tracing performance on their first attempt i'd call that a win, and id be exceedingly worried if i was an Nvidia owner with all their "Dedicated Ray Tracing Hardware" especially given the same Hardware AMD will be using will be in the PS5 and XboxSX, meaning game devs will be more inclined to offer RT that works across all hardware rather than once again Nvidia blackbox implementation of it...

Some people here are so blinded by Nvidias marketing they simply cannot see the wood for the tree's.

If you want to drive something you have to look at adoption rate, if im a game Dev and im making a game for XboxSX, PS5 and PC, and i want to use Ray Tracing, im not even going to consider RTX at first as it simply will not work on XboxSX and PS5, so im going to go with their implementation, and potentially bolt on RTX features for the PC port, all the while those without RTX hardware are still going to get a decent implementation if they have RDNA2 hardware on hand.

 

[james.miller]

you are mocking AMD's first attempt at Ray tracing was actually inbetween their 3070 and 3080, and better than Nvidias first attempt? lol so its ok as Nvidia get a pass on their first attempt but AMD would be ridiculed for it?

Wait what have i missed, who's mocking AMD? and dont you remember RTX being slated when it debuted?

 

[KentMan]

Wait what have i missed, who's mocking AMD?

@Grim5 if you look at his history of posts on this thread you'll see hes been harping on about Nvidias supreme performance with DLSS and RTX in this thread for ages, even bringing up some random bar chart that was leaked on the net a few days ago and is more than likely fake.

If you read my post you would see i was refering to that exact bar graph, if he is going to use that as a stick to beat AMD with yet again, he is again making himself look foolish, as according to that graph, AMD's Ray Tracing perf on this their first attempt is superior to Nvidias lol...

 

[melmac]

I read something recently that AMD is also going to use TSMC 6nm, which is basically 7nm EUV, not to be confused with 7nm+.

https://www.anandtech.com/show/1422...echnology-7-nm-with-higher-transistor-density

https://www.gizchina.com/2020/08/21/tsmc-6nm-process-is-currently-in-mass-production/

i'll see if i can find the link.

In TSMC 7nm+ and 7nm EUV are the same thing. First generation 7nm is just 7nm, second generation 7nm with EUV is called 7nm+

I doubt AMD will be using 6nm for RDNA 2. It takes a year or more before for companies to start using new nodes. But that's not the issue. 6nm is based on first generation 7nm. So Companies who have chip designs for 7nm will be able to use 6nm without much problem. But chips designed for N7+ would need a redesign. And going from the Anand tech article, 7nm+ seems to be better than 6nm.

Here is the relevant quote from the Anand tech article you linked to.

Essentially, N6 allows to shrink die sizes of designs developed using N7 design rules by around 15% while using the familiar IP for additional cost savings.

Navi 1 = N7
RDNA 2 = N7+

 

[KentMan]

In TSMC 7nm+ and 7nm EUV are the same thing. First generation 7nm is just 7nm, second generation 7nm with EUV is called 7nm+

I doubt AMD will be using 6nm for RDNA 2. It takes a year or more before for companies to start using new nodes. But that's not the issue. 6nm is based on first generation 7nm. So Companies who have chip designs for 7nm will be able to use 6nm without much problem. But chips designed for N7+ would need a redesign. And going from the Anand tech article, 7nm+ seems to be better than 6nm.

Here is the relevant quote from the Anand tech article you linked to.



Navi 1 = N7
RDNA 2 = N7+

Yep pretty sure i read somewhere though that AMD had some designs coming on 6nm (7nm EUV) as well, wouldnt be surprised if they refresh something on that, maybe APU's?