** The AMD Navi Thread **

[TNA]

I think mGPU will make a comeback

Doubt it. mGPU is ****. Devs won't do the work needed to make it work on dx12. Until consoles have it, it will likely remain dead.

Never got day one drivers when I had my 295x2. I will personally likely never do it again.

What I can see that is more likely to come next is multi core gpu's on the same die.

 

[4K8KW10]

Shouldn't many small chips working as one make multi-card easier to do, though?

Chiplets, say hallo

I think mGPU will make a comeback, once DX12 gets more entrenched as an ecosystem. But that will only happen with the next console switch, so it's going to be a few years from now until we see it happen.

I think people overstate how bad CF/SLI is, plenty of great games work beautifully with it. I don't see why someone who would want to pay double for 20-30% more performance (1080 ti -> 2080 ti; V64 -> VII) is making such a better decision than someone who would rather get another card and enjoy >50% performance improvements in 1/2 - 2/3rds of relevant games.

It should have happened already but AMD is in debt to us all for not to.

 

[Yaayuh!]

I always found the GTX690 the most efficient and smoothest SLI solution i've ever used. I wouldn't mind seeing this revisited, must be a better to implement with DX12.

 

[D.P.]

I always found the GTX690 the most efficient and smoothest SLI solution i've ever used. I wouldn't mind seeing this revisited, must be a better to implement with DX12.

The issue is largely how modern game engined with deferred rendering and complex faked GI and visual trickery require s lot of inter- and intrs-frame data dependencies.

This is why MGPU started slumping for both AMD and Nvidia at the same time.


And anyone talking about c"making multi-gpus appear as one" or " just do what ryzen does" doesn't have a clue what they are talking about.


Interestingly, with RTX it is much easier to go down a multi-gpu solution as the algorithm and hardware are intrinsically infinitely parallelizeable without data interdependence. Each ray is entirely independent and can calculated on separate dies , the results are merely summed in an accumulation buffer.

 

[Troezar]

I'm really keen to see what Navi is like. I'm now running a 5830 and will be buying an AMD card to tide me over until AMD's next architecture. So either Vega 64, Radeo VII or hopefully the newer Navi. Fingers crossed it will perform well at 1440p.

 

[Calin Banc]

Interestingly, with RTX it is much easier to go down a multi-gpu solution as the algorithm and hardware are intrinsically infinitely parallelizeable without data interdependence. Each ray is entirely independent and can calculated on separate dies , the results are merely summed in an accumulation buffer.

Then you can have your standard GPU with dedicated RT chips glued or even your dedicated RT (chip) card with as many chips as required.
Of course, just as well you can have chips for phyiscs or AI.

So a chiplet design is not entirely unachievable. Just different.

 

[4K8KW10]

Ray-tracing doesn't make sense because using it you sacrifice your 4K60 framerate gaming to 2K30 gaming.

 

[B12B6]

Some people make you smile.
They hate AMD for all it is. Their products are garbage because they dont give 2080Ti performance for 300$ or they dont give you faster card than 1080Ti so why even bother release anything. Have you ever heard other people or other usercases? No. Even if AMD would bring faster and cheaper product than Nvidia has, they would still wait and see what Nvidia can bring to the table. Same people want AMD to die and at the same time they are mad as hell to AMD not bringin better products so Nvidia could not rob them bling and would have to lower prices. If they were to buy AMD the card would need to be a lot faster and significantly cheaper.

Im excited for Navi, I want ~Vega64 with lower powerconsumption and lower price. I actually want AMD not to bring high end chips to market, they wont sell. People will not change from Nvidia no matter what. I really hope there is no problem with powerconsumption on Navi, using 390 undervolted is just too much heath. 7870 and Gtx770 where way nicer to play with. Its only reason I dont plan to buy Vega at the moment even if they are reasonably cheap.

The issue is largely how modern game engined with deferred rendering and complex faked GI and visual trickery require s lot of inter- and intrs-frame data dependencies.

This is why MGPU started slumping for both AMD and Nvidia at the same time.

Not my area of experties in any way or shape but I dont think that is what the problem is. Making specific code for every game for such a small market is not just feasable. What I can remember 295/690 days the biggest problem is coherent memory. They have overcome it in servers but for gaming its way harder.

 

[TNA]

And anyone talking about c"making multi-gpus appear as one" or " just do what ryzen does" doesn't have a clue what they are talking about.

I heard it from the adored guy who mentioned it last year in one of his videos. But maybe he does not know what he is talking about. Lets see if his predictions will come true with Navi

He was talking about chiplets and having 4 small one's on one die as I recall. Care to enlighten us this won't be happening?

 

[ANDREW GREEN]

I heard it from the adored guy who mentioned it last year in one of his videos. But maybe he does not know what he is talking about. Lets see if his predictions will come true with Navi

He was talking about chiplets and having 4 small one's on one die as I recall. Care to enlighten us this won't be happening?

Was he talking about stacking on the same chipset which is maybe a bit different?

No I also don't know what I'm talking about

 

[Troezar]

Anyone remember how warning there was of Zen being so much better? Just wondering if all is quiet on Navi because it's very good or disappointing. I can see good reason to stay silent on both counts

Fingers crossed I can get 1080 performance for a reasonable price and similar for 9900K like CPU performance. I think with quality motherboard AMD could be >£300 cheaper with Navi and Ryzen 3.

 

[ricid]

I remember somewhere (anandtech?) reporting from some trade show or other an impression that AMD reps seemed very bullish about both their CPU and GPU products for the year, in a 'more than usual' sort of way. Obviously one person's interpretation of a vibe they got from some marketing types isn't exactly much to go on, but I'm hopeful it'll at least be less of a disappointment than the last few iterations.

 

[CuriousTomCat]

I thought polaris was much better than Nvidia's competing products.

 

[Calin Banc]

I heard it from the adored guy who mentioned it last year in one of his videos. But maybe he does not know what he is talking about. Lets see if his predictions will come true with Navi

He was talking about chiplets and having 4 small one's on one die as I recall. Care to enlighten us this won't be happening?

I believe he was talking about an APU style system, with an I/O chip plus CPU chiplet(s) and a smaller GPU. Also in his last video he was talking about a supercomputer with the CPU and GPU interconnected, but that's something different than gaming.

 

[nashathedog]

Not my area of experties in any way or shape but I dont think that is what the problem is. Making specific code for every game for such a small market is not just feasable. What I can remember 295/690 days the biggest problem is coherent memory. They have overcome it in servers but for gaming its way harder.

Both cards could have done with more memory, The 690 was a pair of 2gb 680's when it should have been a pair of 4gb 680's & the 295 was a pair of 4gb 290x's when it should have been a pair of 8gb 290x's but even if they had of been it wouldn't of made that big a difference as we saw when someone did their own dual 300 series Grenada graphics card (Powercolor?) that used a pair of 8gb 390 chips. As you know it's down to the additional support needed for individual games to take advantage of the tech.

Chiplets are definitely not out of the question as @D.P. implied, As already mentioned Turing shows us an example of how ways of doing things can be moved around & over to a secondary chip/component, I think there's a good chance that that's how Navi will be getting ray tracing support hence the apparent release delay which may be due to AMD changing their approach at the last minute after Turing's Ray-tracing was put out there by Nvidia. It may also prove to be a way Navi 10 can be faster than even the VII, in the same way an RTX2060 can beat a 1080ti when using the ray-tracing tech. If AMD are working on rushing out Navi with a ray tracing chiplet it may mean it's performance will be precariouis for quite a while as the driver team will have their work load spike in the early days.

 

[Rroff]

Turing shows us an example of how ways of doing things can be moved around & over to a secondary chip/component

This is something I think both AMD and nVidia will be working on generally with their architectures over the next little while towards architectures that can be uncoupled in an effective way from a monolithic design.

 

[Grim5]

Here’s to hoping Navi 10, err I mean PlayStation 5 GPU has hardware RT built in

This is something I think both AMD and nVidia will be working on generally with their architectures over the next little while towards architectures that can be uncoupled in an effective way from a monolithic design.

I’m not an engineer so I’m not sure what else can be done using chiplets type design but it’s certainly good for thought - perhaps in the future we could see GPUs with a core that takes care of the basic calculation functions with several other cores that handle fixed function calculations on their own to speed up the entire output.

If RTX is a good indication for what fixed function cores can do then it could be a much better way to extract performance out of cards - RTX cores Boosts it’s calculation types by 3 times over the fastest Pascal card - something that cannot be achieved by just using a single shader core for everything, even if that was manufactured on a 7nm process node, there is just no architecture or node shrink that has delivered 3 times performance.

 

[nashathedog]

This is something I think both AMD and nVidia will be working on generally with their architectures over the next little while towards architectures that can be uncoupled in an effective way from a monolithic design.

Exactly, They'll find a way to do it because they're almost at the end of the road with die shrinking & they need to continue moving forward in respects to the performance they offer, at the moment there's no alternative.

 

[Burnt_Custard]

I just wondering if Navi can beat the value for money their own Polaris series currently offers. A RX 570-590 for £100-200 is actually really good value. Vega 56 and 64 for £250-350 is not too shabby as well. So unless they offer something better than that (or same but with lower power requirements) then Navi may be underwhelming.

 

[4K8KW10]

Exactly, They'll find a way to do it because they're almost at the end of the road with die shrinking & they need to continue moving forward in respects to the performance they offer, at the moment there's no alternative.

Yeah, imagine future cards with the 3nm chiplets - please one graphics card with 10 chiplets, please; nope, I'd like one with 12 chiplets, please lol