New Zen microarchitecture details

[The Stilt]

Don't know if this is worth anything, but AMD updated their cover photo on Facebook.

Carrizo / Bristol Ridge.

 

[The Stilt]

If the final clocks for RX 480 are indeed as low as 1266MHz, it quite honestly doesn't look too good for 14nm LPP

If AMD isn't able to push the clocks significantly higher (1500MHz+) without blowing the power draw through the ceiling, that's extremely alarming IMO.

AMD partners have shipped Tahiti and Bonaire based cards clocked to 1150 - 1200MHz from the box, and these GPUs were built on TSMC 28nm HP process... ...Which is thought inferior to GlobalFoundries 28nm HPP for example. Not to mention that Bonaire for example is able to clock close to 1400MHz on air cooling.

So if AMD is really unable to ship a ~232mm² GPU clocked higher than 1266MHz, then my fears about the 14nm LPP have materialized even worse than I anticipated :'(

Meanwhile nVidia was able to clock a 35% larger (232 vs. 314mm²) GPU 50% higher (1266MHz vs. 1898MHz), while only having 20% higher TDP D:

What is happening at AMD

 

[24601]

If the final clocks for RX 480 are indeed as low as 1266MHz, it quite honestly doesn't look too good for 14nm LPP

If AMD isn't able to push the clocks significantly higher (1500MHz+) without blowing the power draw through the ceiling, that's extremely alarming IMO.

AMD partners have shipped Tahiti and Bonaire based cards clocked to 1150 - 1200MHz from the box, and these GPUs were built on TSMC 28nm HP process... ...Which is thought inferior to GlobalFoundries 28nm HPP for example. Not to mention that Bonaire for example is able to clock close to 1400MHz on air cooling.

So if AMD is really unable to ship a ~232mm² GPU clocked higher than 1266MHz, then my fears about the 14nm LPP have materialized even worse than I anticipated :'(

Meanwhile nVidia was able to clock a 35% larger (232 vs. 314mm²) GPU 50% higher (1266MHz vs. 1898MHz), while only having 20% higher TDP D:

What is happening at AMD

AIBs are saying GP104 can clock to 2.4 ghz.

The lack of anyone interested in Glofo for much of anything ever should have been people's first clue.

 

[ShintaiDK]

If the final clocks for RX 480 are indeed as low as 1266MHz, it quite honestly doesn't look too good for 14nm LPP

If AMD isn't able to push the clocks significantly higher (1500MHz+) without blowing the power draw through the ceiling, that's extremely alarming IMO.

AMD partners have shipped Tahiti and Bonaire based cards clocked to 1150 - 1200MHz from the box, and these GPUs were built on TSMC 28nm HP process... ...Which is thought inferior to GlobalFoundries 28nm HPP for example. Not to mention that Bonaire for example is able to clock close to 1400MHz on air cooling.

So if AMD is really unable to ship a ~232mm² GPU clocked higher than 1266MHz, then my fears about the 14nm LPP have materialized even worse than I anticipated :'(

Meanwhile nVidia was able to clock a 35% larger (232 vs. 314mm²) GPU 50% higher (1266MHz vs. 1898MHz), while only having 20% higher TDP D:

What is happening at AMD

This is what some of us said for a long time. But the hype crowd was in a denial. 14LPP just isn't good compared to TSMC 16FF+ or Intel 14nm. Well, none rivals Intels 14nm. And the result is 199$ 150W 4GB card with ~Hawaii performance because its all they got while Nvidia sits on anything performance related and up. Zen requires more than luck at this point.

 

[plopke]

isn't that for the many R480 threads , must have my daily read on how you guys interpret 40% xD. Which is quiet entertaining to read,yes I also hope for zen (). Will we get it , who knows. Starting to wonder if they left the desktop market all together and only will return with zen.

 

[itsmydamnation]

This is what some of us said for a long time. But the hype crowd was in a denial. 14LPP just isn't good compared to TSMC 16FF+ or Intel 14nm. Well, none rivals Intels 14nm. And the result is 199$ 150W 4GB card with ~Hawaii performance because its all they got while Nvidia sits on anything performance related and up. Zen requires more than luck at this point.

good to see the DERP DERP cuz AMD has arrived in the thread. Like normal not one it insight or piece of evidence to support your conclusion.

You both realize that Polaris can dynamically boost its clocks on a PER CU basis above its nominal clock, based on CU occupancy. So any clock ceiling has to factor in being able to sustain that boost state while delivering it within a sub 150 watt target. We also have to see if there is a 480x as only the 480 was talked about.

also Shintai in your expert opinion where should the 2nd smallest of 4 GPU soc's sit, where should each SOC begin and end in terms of price range and performance.

You can also say what does NV have in the laptop space, nothing, be great seeing a GP104 throttle after 3 seconds worth of game play.

If AMD isn't able to push the clocks significantly higher (1500MHz+) without blowing the power draw through the ceiling, that's extremely alarming IMO.

keep ignoring kyro and mongoose, CPU cores that directly contradict your constant process complaints both clocking very high by mobile SOC standards especially considering there width.

So if AMD is really unable to ship a ~232mm² GPU clocked higher than 1266MHz, then my fears about the 14nm LPP have materialized even worse than I anticipated

see above ( ie dynamically clocks higher), Also there is more then just clocks, AMD has specifically noted at higher occupancy and utilization per CU. Also consider that this is 480 not a 480X, it could very well be a situation like tonga where Apple get the best chips and we get the dregs :'(.

look at 390x power consumption, gaming is around 300-350watts to similar performance at whats is "leaked" at 115 to 135watts. So only a ~2.6 times increase in efficiency....

As said before there are 2 SOC above Polaris 10 for AMD

way to much 1+1 = 13 by people here.

 

[The Stilt]

You both realize that Polaris can dynamically boost its clocks on a PER CU basis above its nominal clock, based on CU occupancy.

What's your source for the new boosting algorithm in GCN4?

 

[SpaceBeer]

I'm not sure "low" Polaris clocks are only 14 LPP related, it also has something with GCN. Comparing last generation GCNs vs Kepler and Maxwell, which were made using the same process (28nm TSMC), it is obvious TDP and clocks depend more on architecture. And we know for sure that small increases in clock lead to much larger increases in TDP and power consumption (eg. Nano to Fury X, M295X to 380X - latter have 20% larger clocks, and 60% larger TDP). So even if GF 14LPP is not as good as TSMC 16FF+, it doesn't mean Zen won't reach large frequences.

 

[beginner99]

This is my main grief with building a rig right now. And I need a new rig very soon.
If Zen brings a somewhat competetive 8 core mainstream -> Cannonlake will likely bring 6 cores mainstream.
That equals that any quad you're building today is gonna be 2 cores short of a full deck in a years time.

Cannonlake design is almost certainly pretty much complete now. If they go with 6 cores, it's already decided, if not Zen being a success can not change that. I highly doubt Intel will go 6-core in mainstream yet.

What's your source for the new boosting algorithm in GCN4?

This was discussed in V+G forum few weeks back. It's from AMD patents. We however don't now if that stuff is in Polaris. Maybe it will only be used in Vega and later or not at all. But if it is used, yes, P10 could offer much higher performance in certain scenarios than one would expect from CU count.

I'm not sure "low" Polaris clocks are only 14 LPP related, it also has something with GCN. Comparing last generation GCNs vs Kepler and Maxwell, which were made using the same process (28nm TSMC), it is obvious TDP and clocks depend more on architecture. And we know for sure that small increases in clock lead to much larger increases in TDP and power consumption (eg. Nano to Fury X, M295X to 380X - latter have 20% larger clocks, and 60% larger TDP). So even if GF 14LPP is not as good as TSMC 16FF+, it doesn't mean Zen won't reach large frequences.

Exactly. Maxwell already clocked much higher than GCN on same process. And Pascal is a Maxwell die shrink more or less. Also FinFet say 40% higher clock at same power level. Hawaii is basically factory OCed. Power use gets way better at lower clocks (and voltage). At 900 mhz it's much more efficient. So 900 mhz * 1.4 = 1.26 Ghz, exactly what leaks indicate for P10. This also means there is potential for custom cards with much higher clocks (and power use). It could probably reach GTX 1070 levels of performance albeit at higher power usage.

 

[JimmiG]

I think he confused '40% IPC' and '40% performance'.

IPC. It needs to have a 60% higher IPC than their previous gen in order to match Intel on IPC. Provided the CPU runs at a higher clock speed like 4.7 / 5 GHz base/boost vs Intel who like to stay at around 4 GHz, it doesn't have to be quite as large. However for enthusiasts, who overclock their Intel CPUs to 4.4 - 4.8 GHz, it would have to be pretty close.

 

[el etro]

Nvidia TDP is set pratically at the Typical Board Power level, while AMD TDP is set at the Maximum Board Power level. P10 ACP is likely much less than this.


Pitcairn itself is rated at 175W TDP, yet it consumes roughly half of this: http://www.techpowerup.com/gpudb/339/radeon-hd-7870-ghz-edition

 

[Arachnotronic]

Nvidia TDP is set pratically at the Typical Board Power level, while AMD TDP is set at the Maximum Board Power level. P10 ACP is likely much less than this.

EDIT: Whoops, maybe not. Nevermind.

 

[Dresdenboy]

This is the decoded data matrix code: SM0035049AD6ZPAA000121

You could split it into SM0035049AD6 ZPAA000121 or SM0035049 AD6ZPAA000121.
SM=Summit Ridge?
00=?
35=clock frequency?

Sharper pic of the chip at Computerbase:

http://www.computerbase.de/2016-06/amd-zen-gezeigt/

I read:
"ZPAA000121"

This is an early ES, not comparable to the usually distributed ones marked with "ZD.." or "2D.." for desktop.

"ZP" could mean "Zeppelin" (which is assumed to be 16C, but who knows), or also simply "Zen Processor". "ZP" has also been used in some AMD patches.

"AA" or after "ZPA" the "A0" might stand for the stepping, and "121 is just a serial number, I think.

 

[ShintaiDK]

Nvidia TDP is set pratically at the Typical Board Power level, while AMD TDP is set at the Maximum Board Power level. P10 ACP is likely much less than this.


Pitcairn itself is rated at 175W TDP, yet it consumes roughly half of this: http://www.techpowerup.com/gpudb/339/radeon-hd-7870-ghz-edition

No...

 

[sirmo]

If the final clocks for RX 480 are indeed as low as 1266MHz, it quite honestly doesn't look too good for 14nm LPP

That's more than 30% of gain over Tonga (@970Mhz stock). And the previous cards didn't have much OC overhead. We don't know about these, but judging by the power efficiency claims there might be plenty of room this time.

 

[sirmo]

No...

Is this while thermally throttling?

 

[forgotmypasswrd]

This is the decoded data matrix code: SM0035049AD6ZPAA000121

You could split it into SM0035049AD6 ZPAA000121 or SM0035049 AD6ZPAA000121.
SM=Summit Ridge?
00=?
35=clock frequency?

are you sure that's correct? It seems to be missing 1 character. But I guess they might've changed it. The number after SM(in this case 00) should've been the clock frequency.

 

[el etro]

No...

Pretty much a worst possible case for the card.

 

[The Stilt]

This is the decoded data matrix code: SM0035049AD6ZPAA000121

You could split it into SM0035049AD6 ZPAA000121 or SM0035049 AD6ZPAA000121.
SM=Summit Ridge?
00=?
35=clock frequency?

Decoding 2D barcodes from a screenshot feels quite autistic to me, but this information doesn't reveal anything new

First ten characters is the prototype number (no frequencies encoded). In the last eight characters the first two tell which design it is (ZP = Zeppelin). The remaining six is just a serial number.

I've got several SM0024317x / GVxxxxxx prototypes and they are Godavari APUs as the "GV" indicates.

 

[sweetusernames]

If the final clocks for RX 480 are indeed as low as 1266MHz, it quite honestly doesn't look too good for 14nm LPP

If AMD isn't able to push the clocks significantly higher (1500MHz+) without blowing the power draw through the ceiling, that's extremely alarming IMO.

AMD partners have shipped Tahiti and Bonaire based cards clocked to 1150 - 1200MHz from the box, and these GPUs were built on TSMC 28nm HP process... ...Which is thought inferior to GlobalFoundries 28nm HPP for example. Not to mention that Bonaire for example is able to clock close to 1400MHz on air cooling.

So if AMD is really unable to ship a ~232mm² GPU clocked higher than 1266MHz, then my fears about the 14nm LPP have materialized even worse than I anticipated :'(

Meanwhile nVidia was able to clock a 35% larger (232 vs. 314mm²) GPU 50% higher (1266MHz vs. 1898MHz), while only having 20% higher TDP D:

What is happening at AMD

https://youtu.be/xtely2GDxhU?t=2503

yeah, it's really insane how high nVidia managed to clock Pascal.

hopefully this won't become GloFo disaster, I really want to see Zen do well.

 

[Erenhardt]

No...

Yes...
7870 has 190W TDP (typical board power)

Uses 103 watts roughly half of the tdp.

Meanwhile the gtx1080 you posted which has 180W tdp isn't shy of breaking the TDP and gobble more than 180watts. Even when it is throttling it's less than 10% below TDP.

 

[ShintaiDK]

Pretty much a worst possible case for the card.

Typical gaming the worst possible case? Should we talk about the 350W 390X usage in AOTS?

Lets be honest, you tried using the same BS we heard for ages about AMD CPUs with their GPUs. It doesn't hold water.

 

[ShintaiDK]

Yes...
7870 has 190W TDP (typical board power)

The 7870 got a TDP of 130W. The 7870Ghz 175W.

Meanwhile the gtx1080 you posted which has 180W tdp isn't shy of breaking the TDP and gobble more than 180watts. Even when it is throttling it's less than 10% below TDP.

Sorry to smash your dreams. The retail card with retail driver doesn't throttle

I assume you refer to peak power of 186W against its 180W TDP.

R9 285 uses 200W, its 190W TDP.
R9 290 uses 263W, its 250W TDP.
R9 290X uses 294W, its 250W TDP.
Fury X uses 280W, its 275W TDP.

You picked a really bad argument.

And for the CPUs we know that AMD states its TDP below usage. Throttlemonster APUs, 140W FX chips etc.

Being locked to 14LPP is just making it worse for AMD, and it will affect Zen CPUs and APUs negatively. The WSA is a huge burden they shouldn't carry.

 

[Erenhardt]

The 7870 got a TDP of 130W. The 7870Ghz 175W.


Every 7870 is a Ghz. 7870 had baseclock at 1000mhz. You are mixing things up.

Sorry to smash your dreams. The retail card with retail driver doesn't throttle

You mean all the reviews are lying?

I assume you refer to peak power of 186W against its 180W TDP.

R9 285 uses 200W, its 190W TDP.
R9 290 uses 263W, its 250W TDP.
R9 290X uses 294W, its 250W TDP.
Fury X uses 280W, its 275W TDP.

You picked a really bad argument.

290/x is 300watt tdp (unofficial) per anandtech.

But I have to agree that latest amd cards have their tdp much closer to the real power usage than the old series.

Anyway that is offtopic. It is hard to draw the line between CPU and GPU process scaling.

 

[monstercameron]

Im surprised by all this clock speed talk after the age of the MHz myth had long past.