AMD: Speculation, Rumors, and Discussion (Archive)

[lanek]

I've been wondering about something. AMD potentially has more to gain from the next node transition (next gen) than Nvidia. Considering Nvidia have reduced transistor count for DP in order to boost performance for SP and AMD hasn't done the same, or at least not to nearly the same extent. There's also low hanging fruit there for better color compression.

It makes me wonder if AMD will continue to go with a lot of DP performance for the top end, or if they'll follow in the footsteps of Nvidia and reduce that in order to boost performance in games and things that don't require DP?

Then again with a new node transition and more transistors to play with, will Nvidia go back to higher DP performance for the professional market?

Regards,
SB

Is Nvidia will continue this trend ? I think the DP question was only for Maxwell, so next arch will certainly see DP performance coming back.

This decision of remove completely DP and other computing stuff, was only due to the problem both AMD and nvidia have encounter, when 22nm have been abandonned by TSMC. They have need to make some choice, and cut where they can.

 

[Samwell]

I've been wondering about something. AMD potentially has more to gain from the next node transition (next gen) than Nvidia. Considering Nvidia have reduced transistor count for DP in order to boost performance for SP and AMD hasn't done the same, or at least not to nearly the same extent. There's also low hanging fruit there for better color compression.

It makes me wonder if AMD will continue to go with a lot of DP performance for the top end, or if they'll follow in the footsteps of Nvidia and reduce that in order to boost performance in games and things that don't require DP?

Then again with a new node transition and more transistors to play with, will Nvidia go back to higher DP performance for the professional market?

Regards,
SB

I don't get your point. AMD has done the same as Nvidia. Maxwell is low DP, but Fiji is also low DP. So where is the difference?
Both will go with high DP at the Top-End and lower DP in the lower Tiers as it's not needed there.

 

[Kaotik]

Then again with a new node transition and more transistors to play with, will Nvidia go back to higher DP performance for the professional market?

Regards,
SB

Yes, they will, they've already confirmed it.

 

[Silent_Buddha]

Yes, they will, they've already confirmed it.

So it'll be back to a more level-ish playing field for the next node. With both focusing on DP. AMD were competing just fine WRT performance/power (and perf/mm^2) up until Nvidia surprised them by ditching DP performance.

Of course, AMD's personnel have been gutted (relatively speaking) since then, so it remains to be seen if they can challenge Nvidia on the upcoming node transition.

Regards,
SB

 

[silent_guy]

So it'll be back to a more level-ish playing field for the next node. With both focusing on DP. AMD were competing just fine WRT performance/power (and perf/mm^2) up until Nvidia surprised them by ditching DP performance.

I don't think DP explains the difference the perf/W increase of Maxwell at all. There are architectural changes that are a much more logical explanation. (And it's not as if Fiji is a DP screamer either.) The lack of DP on gm200 can be justified by the fact that this was the only way to be sufficiently faster than gm204 without going far beyond 600mm2.

That said: I don't think we'll be able to do an apples to apples comparison because I expect AMD to pay more attention to perf/W for 16nm than they did for 28nm.

So if, for 16nm, thing are closer in terms of perf/W between both architectures, those claiming that it's due to DP can still claim to be right...

 

[Alexko]

DP probably plays a role (otherwise, why remove it?) but I think it's quite clearly very minor, otherwise GCN incarnations with slow DP would be more efficient than they are.

 

[silent_guy]

DP probably plays a role (otherwise, why remove it?) ...

Like I wrote: you remove it because you're limited by die size, and the desire to have a top dog that has at least 50% more resources than the second in rank (though Fiji didn't pass even that bar.)

 

[Alexko]

Like I wrote: you remove it because you're limited by die size, and the desire to have a top dog that has at least 50% more resources than the second in rank (though Fiji didn't pass even that bar.)

True but it's also usually absent from mid-range chips.

 

[Ryan Smith]

Since AMD has all of the subtlety of a boot to the head, let the speculation on Polaris begin.

"Starry skies in Sonoma next week. Should have an excellent view of Polaris"

"Polaris is 2.5 times brighter today than when Ptolemy observed it in 169 A.D"

https://twitter.com/GChip/status/669637153748484096

 

[Newguy]

2.5x the performance or perf/w of the Fury X confirmed.

 

[Silent_Buddha]

2.5x the performance or perf/w of the Fury X confirmed.

Let's hope that it's that and not that it's going to take 1,847 years until it is released.

Regards,
SB

 

[Alexko]

Is there some sort of AMD event in Sonoma next week? I haven't heard of anything.

 

[Grall]

Well f--k me; the very WEEK I buy/transition to two 390Xes, the rumors start flying about next gen!

 

[gkar1]

Since AMD has all of the subtlety of a boot to the head, let the speculation on Polaris begin.

"Starry skies in Sonoma next week. Should have an excellent view of Polaris"

"Polaris is 2.5 times brighter today than when Ptolemy observed it in 169 A.D"

https://twitter.com/GChip/status/669637153748484096

This is altogether odd, we know Ptolemy died in 168A.D. so he never even saw Polaris and the star was actually discovered in the late 18th century by Herschel.

It's probably a reference to the dual Fiji card cause the case can be made for the dual card to be 2.5 times faster if you compare it to the Nano

 

[AlNom]

a) $169
b) 169mm^2
c) 169TFLOPs
d) 169 man years R&D etc
e) 16.9nm
f) 16:9
g) 169 sales projection
h) 16 September 2016

 

[Voxilla]

I don't get your point. AMD has done the same as Nvidia. Maxwell is low DP, but Fiji is also low DP. So where is the difference?
Both will go with high DP at the Top-End and lower DP in the lower Tiers as it's not needed there.

It's still pretty amazing that AMD manages 2.5 TFlop/s DP with Haiti, about twice Nvidia top end Tesla.
They have 1/2 the single precision rate!
https://www.amd.com/Documents/firepro-s9150-datasheet.pdf

 

[DieH@rd]

2.5x the performance or perf/w of the Fury X confirmed.

I would like 2.5x perf performance of the R9 Nano.

 

[Deleted member 2197]

It's still pretty amazing that AMD manages 2.5 TFlop/s DP with Haiti, about twice Nvidia top end Tesla.
They have 1/2 the single precision rate!

I think correct Tesla numbers are:

Up to 2.91 TFlops double precision performance and Up to 8.74 TFlops single precision performance on the Tesla K80 GPU Accelerator

.
http://www.nvidia.com/object/why-choose-tesla.html

 

[Grall]

c) 169TFLOPs

I'm all for this one!
...Excuse me, I need to be alone for a while with this tube of hand lotion...

 

[Ryan Smith]

It's probably a reference to the dual Fiji card cause the case can be made for the dual card to be 2.5 times faster if you compare it to the Nano

That's entirely possible since the card hasn't launched yet. It's a new codename, so I'm just spitballing here on which thread to put it in. (Though 2.5x Nano would be a really odd comparison)