New AMD GPUs in 2018

supposedly 25% faster than Nvidias Turing in compute*
Click to expand...

* Until you get the Tensor units involved.

Turing will absolutely embarrass Vega at FP16 matrix operations (it's calling-card).
 
defaultluser said:
* Until you get the Tensor units involved.

Turing will absolutely embarrass Vega at FP16 matrix operations (it's calling-card).
Click to expand...

Yeah, it will be interesting to see the comparison..I would expect Nvidia to come out ahead considering the node advantage, the massive die, and the fact they have allocated dedicated HW for FP16 while VEGA uses traditional programmable shaders just like every other GPU that has come before the RTX cards.
 
They both use the regular shaders for most FP16 arithmetic. It's just when you do the 4x4 matrix multiply + add of FP16's that you can instead use tensor cores on Turing (and Volta) which is indeed different silicon, and much faster.
 
euskalzabe said:
Or not. Everything is pure speculation until we have products in hand. We don't even have RTXs yet.
Click to expand...

We have very clear white papers on turing for fp16. It’s the AMD card that is pure speculation.

Even if an AMD gaming card launched with 20TF if there are no architecture changes that increase IPC (you know what I mean) Turing will run laps around it even ignoring ~1/3 of the die (ray tracing cores and tensor cores.). nVidia already had a huge advantage in TF to FPS efficiency and they increased it even further with Turing.
 
Dayaks said:
We have very clear white papers on turing for fp16. It’s the AMD card that is pure speculation.

Even if an AMD gaming card launched with 20TF if there are no architecture changes that increase IPC (you know what I mean) Turing will run laps around it even ignoring ~1/3 of the die (ray tracing cores and tensor cores.). nVidia already had a huge advantage in TF to FPS efficiency and they increased it even further with Turing.
Click to expand...
And if the reports are correct the 336 mm2 would mean if they put 2 on 1 card which would not even match Nvidia die size would exactly run circles around Turing ?

I mean you can make a chip really really large but it does not mean that the potential from smaller chips somehow evaporates...
 
Pieter3dnow said:
And if the reports are correct the 336 mm2 would mean if they put 2 on 1 card which would not even match Nvidia die size would exactly run circles around Turing ?

I mean you can make a chip really really large but it does not mean that the potential from smaller chips somehow evaporates...
Click to expand...

From what I’ve seen nVidia has vastly higher FP16 capability per mm^2. Now we’re very, very overdue for an architecture improvement from AMD so hopefully they surprise us next year.
 
Dayaks said:
From what I’ve seen nVidia has vastly higher FP16 capability per mm^2. Now we’re very, very overdue for an architecture improvement from AMD so hopefully they surprise us next year.
Click to expand...

That statement is rather weird because you know it is Vega that is coming in 7nm so that is not something new. Navi is a product directed at middle segment so there is nothing new again even if they used different architectural features it is not going to come close.
And even in 2020 nothing is known about what exactly is coming but what is known that the changes in RTG personnel that took place in 2018 will produce something in 2021 that should be "better" competition wise (finally deal(t) with the problems of power usage when scaling).
 
Pieter3dnow said:
That statement is rather weird because you know it is Vega that is coming in 7nm so that is not something new. Navi is a product directed at middle segment so there is nothing new again even if they used different architectural features it is not going to come close.
And even in 2020 nothing is known about what exactly is coming but what is known that the changes in RTG personnel that took place in 2018 will produce something in 2021 that should be "better" competition wise (finally deal(t) with the problems of power usage when scaling).
Click to expand...

Maybe I read too far into this: “The new GPU supports intrinsic AI instructions and features four HBM2 8GB stacks running across a 4096-bit memory interface for a total of 32GB vRAM.”

If nothing is changing then AMD will have a tough time at FP16 and AI.

On the gaming side I think it’s more likely than not RTX could backfire and AMD might have an advantage.
 
Pieter3dnow said:
That statement is rather weird because you know it is Vega that is coming in 7nm so that is not something new. Navi is a product directed at middle segment so there is nothing new again even if they used different architectural features it is not going to come close.
And even in 2020 nothing is known about what exactly is coming but what is known that the changes in RTG personnel that took place in 2018 will produce something in 2021 that should be "better" competition wise (finally deal(t) with the problems of power usage when scaling).
Click to expand...

The new architecture has to replace GCN has been in the works much before the RTG changes. I can pretty much bet the next gen card is coming in 2020. 100 bucks. I know i will win lol! Next Gen was always planned after navi and it will be here in 2020.
 
NKD said:
The new architecture has to replace GCN has been in the works much before the RTG changes. I can pretty much bet the next gen card is coming in 2020. 100 bucks. I know i will win lol! Next Gen was always planned after navi and it will be here in 2020.
Click to expand...

I thought it still wasn't known weather Navi is GCN or next gen. Originally AMD said Navi wasn't coming until 2020. They have been dead silent for a while though, so who knows.
 
NKD said:
The new architecture has to replace GCN has been in the works much before the RTG changes. I can pretty much bet the next gen card is coming in 2020. 100 bucks. I know i will win lol! Next Gen was always planned after navi and it will be here in 2020.
Click to expand...
GCN is nothing more then how the components are placed on the die itself more then anything else. Posted some suggested info about Navi here:
https://hardforum.com/threads/we-ha...z-edition-redo.1952587/page-6#post-1043694631

GCN allows the compute units to function very well.
 
Last edited:
Pieter3dnow said:
I suggest you donate the R&D budget for next coming cards so they can beat Nvidia hands down..
Click to expand...

LOL while ryzen doing well i think eventually the money will start pouring in to R&D. They will just have spent almost 3 times the time Nvidia spends on a new architecture. So due to low R&D on the GPU side 2020 makes it like over 4 years for the next gen to come out. I think the rumor is they started this in 2014 or 2015 for the next gen architecture. That right there is low R&D result lol. That's why we have been getting rebadges. I honestly thin GCN can shine on 7nm as well. But it is very compute heavy card and it loses efficiency in gaming as the die gets bigger it seems.

They probably realized it way back then but they just have been really slow due to R&D so its just taking longer. But AMDs cpu situation should change all that though going forward. I think good times are ahead. Nvidia knows it too thats why they are cashing out on Turing now. Cuz they know AMD with bigger budget will likely come swinging in a few years under Lisa Su! I do like the silence from RTG now. That's how it should be, all work and then let the product talk! I was getting tired of Hype.
 
Pieter3dnow said:
I suggest you donate the R&D budget for next coming cards so they can beat Nvidia hands down..
Click to expand...

I thought that's what I was doing when I purchased my Fury card :) I would like nothing better than for AMD to beat Nvidia. Only reason why I have an Nvidia card now is because Vega was 6 months away and I needed a card ASAP.
 
I always wondered if they ever planned anymore work with Hawaii if Tahiti is what Polaris scaled to with refinements as to open the memory paths back to 384bit or 512bit and 8Gb DDR6 as Hawaii scales with clock speed as 1.4Ghz would be fast for the big island .
 
crazycrave said:
I always wondered if they ever planned anymore work with Hawaii if Tahiti is what Polaris scaled to with refinements as to open the memory paths back to 384bit or 512bit and 8Gb DDR6 as Hawaii scales with clock speed as 1.4Ghz would be fast for the big island .
Click to expand...

Hawaii uses to much power and other solutions available now prolly better.
 
crazycrave said:
I always wondered if they ever planned anymore work with Hawaii if Tahiti is what Polaris scaled to with refinements as to open the memory paths back to 384bit or 512bit and 8Gb DDR6 as Hawaii scales with clock speed as 1.4Ghz would be fast for the big island .
Click to expand...

Hawaii is long dead. I love that arch, I had the fastest cards here on [H]ard, and arguably the fastest non LN2 cards out there. My Golden sample would do 1.325Ghz with a mild voltage bump, and 1.35Ghz but the extra voltage was not worth it as there was zero improvement (memory restrictions at that point since I did not have Samsung ram which was a crime considering the golden die!!!)...My other 2 would do 1.3Ghz all day, but I ran them at 1250 IIRC in X-fire.

I would have loved to see the die I had with the updated 390 PCB layout and the fastest Samsung ram they shipped with..I would have been able to squeeze out another 5-10% performance..The speed of my 290s was the reason I skipped Fury/Fury-X, as they were just not fast enough for me.

I want to get my hands on a 7nm VEGA with the same PCB layout so I can slap a full cover gpu block on it on bad!
 
You must log in or register to reply here.
Back
Top