Discussion Next Navi GPU specifications (Hypothetical!)

[mohit9206]

If amd launches big Navi in early 2020, most people will just wait a little more for Nvidia 7nm cards as those will be substantial upgrade. The time to launch big Navi is now.

 

[Yotsugi]

most people will just wait a little more for Nvidia 7nm cards

That's a lotta waiting.

The time to launch big Navi is now.

The time to launch big Navi is the exact one drawn on their internal roadmaps.

 

[mohit9206]

That's a lotta waiting.

The time to launch big Navi is the exact one drawn on their internal roadmaps.

Vega was a failure which a lot of people had high hopes for. People already are fed up of AMDs graphics products and their patience keeps wearing thin. 5700XT is a good product which people don't want to purchase because AMD makes terrible reference cards and again they need to wait until reference cards come out while the Super cards are already available. Its almost as if AMD doesn't want to compete or something. And then you make people wait even more for hi end cards which also will have terrible reference models and wait further for custom cards. Just keep waiting if one wants to purchase AMD. I want AMD to kick Nvidia's butt but i don't think it's going to happen anytime soon as AMDs graphics division is in complete shambles.

 

[Yotsugi]

Vega was a failure which a lot of people had high hopes for

How is that relevant in 2013+6 is beyond me.

People already are fed up of AMDs graphics products and their patience keeps wearing thin

People stopped buying them circa 2010, so they can be fed up with them however they want.

5700XT is a good product which people don't want to purchase because AMD makes terrible reference cards

AMD would gladly make you a 2.5 slot 3 fan reference board if you can somehow persuade AIBs to not gouge their eyes for that.

Super cards are already available

Super stuff is inferior in perf/$ so availability matters not.

And then you make people wait even more for hi end cards which also will have terrible reference models and wait further for custom cards

No one buys high-end stuff, it's entirely PR stunt segment of the overall dGPU market.

as AMDs graphics division is in complete shambles.

They have so much stuff in the pipeline, and their roadmaps are so agressive I'd suggest you to danny deleto this post about now.

 

[extide]

If amd launches big Navi in early 2020, most people will just wait a little more for Nvidia 7nm cards as those will be substantial upgrade. The time to launch big Navi is now.

Yeah, seriously, if AMD has a shot at the GPU crown, even if it is short lived, they need to take it. It would help AMD MASSIVELY in mindshare to have the top GPU, even for just a little while.

 

[Glo.]

Yeah, seriously, if AMD has a shot at the GPU crown, even if it is short lived, they need to take it. It would help AMD MASSIVELY in mindshare to have the top GPU, even for just a little while.

Is there enough financial incentive for AMD to design and release such product?

 

[maddie]

Is there enough financial incentive for AMD to design and release such product?

Overall or for that SKU only? A critical difference, as logic does not rule here.

 

[Yotsugi]

if AMD has a shot at the GPU crown, even if it is short lived, they need to take it

"No".

It would help AMD MASSIVELY in mindshare to have the top GPU, even for just a little while.

Why bother, x86 bros will start eradication the very next year.

 

[GodisanAtheist]

I wonder if the other units like ACEs, ROPs, TMUs etc will also be scaling with the additional SPs in the larger Navi to ensure it maintains the somewhat balanced ratios of the 5700 series.

This has been the most maddening thing about the 5700 launch, the utter lack of deep dive into the Navi arch and if AMD has worked in fixes for some of the major hardware level limitations of the GCN arch, the most important being scalabiity.

If kept all the same bottlenecks from the GCN arch era, then what good is another massive shader array without the front end to actually feed it?

 

[Yotsugi]

If kept all the same bottlenecks from the GCN arch era, then what good is another massive shader array without the front end to actually feed it?

It feeds itself.

 

[DiogoDX]

AMD will need a 80CU - 128ROPs - 512Bits or HBM - 16Gb vram monster to compete with Nvidia in 7nm (I'm expeting the 3080Ti to have >6K CUDA cores).

With hardware RT and VRS this will be >500mm GPU and the callenge will be to do this on 300W max.

 

[Glo.]

AMD will need a 80CU - 128ROPs - 512Bits or HBM - 16Gb vram monster to compete with Nvidia in 7nm (I'm expeting the 3080Ti to have >6K CUDA cores).

With hardware RT and VRS this will be >500mm GPU and the callenge will be to do this on 300W max.

By the time it arrives we have around 1.5 years of time to enjoy current generation of GPUs.

 

[Yotsugi]

AMD will need a 80CU - 128ROPs - 512Bits or HBM - 16Gb vram monster to compete with Nvidia in 7nm (I'm expeting the 3080Ti to have >6K CUDA cores).

With hardware RT and VRS this will be >500mm GPU and the callenge will be to do this on 300W max.

That's a lotta N7 xtors.
Who's gonna pay for these?

 

[Glo.]

That's a lotta N7 xtors.
Who's gonna pay for these?

NVIDIA!

 

[insertcarehere]

if AMD has a shot at the GPU crown, even if it is short lived, they need to take it.

"No".

It would help AMD MASSIVELY in mindshare to have the top GPU, even for just a little while.

Why bother, x86 bros will start eradication the very next year.

Because completely punting on the top-end and focusing on high value GPUs worked so well for AMD the last thousand times they tried it....

 

[GodisanAtheist]

Because completely punting on the top-end and focusing on high value GPUs worked so well for AMD the last thousand times they tried it....

-I think it's as simple as: if AMD has a competitive arch, they'll compete with it.

 

[darkswordsman17]

Because completely punting on the top-end and focusing on high value GPUs worked so well for AMD the last thousand times they tried it....

What "thousand times" did they do that? Radeon VII was this year. They sold prosumer and watercooled versions of Vega in 2017. They made FuryX and even a dual Fury card in 2015. Guess, what? None of those helped them in the consumer space (as clearly evidenced by your ridiculous post and the other people whining that AMD doesn't have some stupid expensive card for them not to buy and claiming that they don't even ever try). They're not going to spend the money to put out an even larger chip that won't sell enough to make it worth it, for gamers that aren't going to buy it.

Heck even Nvidia is getting serious pushback on their pricing. If you think AMD is going to fare better then you've must have suffered a serious brain injury in the past year.

Only when AMD/ATi had substantially better hardware and features did gamers actually support them enough to even consider that, and even then a lot of them still stuck to Nvidia. AMD would be a fool to try to go out of their way to cater to that market.

They'd be much better suited leveraging their capabilities to offer something Nvidia couldn't, like a high end console (where it leverages their console/semi-custom, CPU, and GPU development; put CPU modules with GPU module, use HBM2 as system memory and maybe buffer for some NAND that would have the OS and enough for a game to be loaded into when running, with options for PCIe 4 SSD for more storage; with a self-contained water cooling setup to be able to push clock speeds while handling that load). It'd be a complete system that should handily beat comparably spec'ed PC (even the CPU should see a boost as the HBM2 would obliterate memory bandwidth issues - imagine having GB of cache). It should be able to more directly translate the performance optimizations of consoles while leveraging the extra capabilities of higher end components and the software freedom that PC offers. They could even sell a version of that as workstations (that would push the performance of that market as well).

 

[Yotsugi]

Because completely punting on the top-end and focusing on high value GPUs worked so well for AMD the last thousand times they tried it....

GPUs is nothing more than a best TAM expansion for x86 people.
And they will cast the hammer of platform lock soon enough.

 

[lifeblood]

So I gotta ask the question, what specific chips are we talking about?

Early on it seemed the general consensus was there were three Navi dies, Navi 10, then Navi 12 or 14 depending on who you asked, and Navi 20. Navi 20 was supposed to be the big Navi which left 12/14 as low end. Now people are referencing Navi 12 as being targeted at the 2080ti which I thought was supposed to be Navi 20. Is 12 the new 20 or something?

If I’m understanding the current consensus correctly it is now (going from largest to smallest) Navi 20 12, 10, 14. Do I have it right or have I missed the bus? (for millennial's please substitute e-scooter for bus)

 

[Yotsugi]

If I’m understanding the current consensus correctly it is now (going from largest to smallest) Navi 20 12, 10, 14

10, 12, 14, 20 and 21.
Five Navi dies.

And the size is 14 > 10 > %redacted% > %redacted% > %redacted%.
Is that enough for you?

 

[Midwayman]

I'm still hoping they get the whole chiplet multi-gpu addressable as a single gpu thing worked out. Then scaling would be easy.

 

[Yotsugi]

I'm still hoping they get the whole chiplet multi-gpu addressable as a single gpu thing worked out. Then scaling would be easy.

Not in the next 5-7-maybe-more years.

 

[Det0x]

Not in the next 5-7-maybe-more years.

Depends what market you are looking for.. Arcturus could very well be a multi chiplet design for the Radeon Instinct pro market.

 

[Guru]

The new rdna architecture seems very capable, they are running only 40cu's on the RX 5700xt, that leaves them with up to 24cu's they can add, which is likely to reach the RTX 2080 super in performance or even beat it.

I mean the RX 5700 is 36cu's and lower clock speeds than the 5700xt, so with just 4cu's and a tad higher clock speeds the rx5700xt is 15% faster than the 5700, if its able to scale similarly the more cu's you add, then adding 4 more cu's should offer at least 10% performance without increasing clocks, lets say the scaling is lower the more cu's you add, so once you are up 8 its say 18% performance improvement, 16cu's 25% performance improvement, 24cu's 32% performance improvement, with an increase in clock speeds as well it can achieve the performance of the RTX 2080ti, which is about 35% faster than the rx5700xt.

 

[Yotsugi]

Depends what market you are looking for.. Arcturus could very well be a multi chiplet design for the Radeon Instinct pro market.

Crunching FP64 in vacuum != computer graphics.