New Zen microarchitecture details

[krumme]

Is there any expiry date on that wsa btw? Lol

 

[Dresdenboy]

Is there any expiry date on that wsa btw? Lol

The Wafer Supply Agreement will be in effect no longer than May 2, 2024.

http://www.wikinvest.com/stock/Advanced_Micro_Devices_(AMD)/Wafer_Supply_Agreement

Note that AMD's purchases very likely are not enough to cover all of the best process' related fixed and variable costs (because of using older processes, too), so that GF further needs to license or develop competitive (vs TSMC) processes without simply offering something bad to AMD.

I feel that my glooming was bad.... that is worse.... consider that if AMD quits the market the prices will likely skyrocket to new levels... they are literally abusing the customers....

There are other options.

 

[KTE]

I left the raw data at bottom of graph. The theorized zen would be avg 25% faster than a 6700k in MT, not 10-15.

10-15% is my prediction...

Did you know, Intels old SMT paper "Power-Sensitive Multithreaded Architecture" showed 1C SMT giving 1.4x performance gain.

Who would have thought (!)

The part we don't hear often is that 1.4x performance (IPC) through SMT carries a 1.2x power penalty. Simply more utilization.

Let alone adding beefed up cores to the mix.

Sent from HTC 10

 

[Dresdenboy]

10-15% is my prediction...

Did you know, Intels old SMT paper "Power-Sensitive Multithreaded Architecture" showed 1C SMT giving 1.4x performance gain.

Who would have thought (!)

The part we don't hear often is that 1.4x performance (IPC) through SMT carries a 1.2x power penalty. Simply more utilization.

Let alone adding beefed up cores to the mix.

SMT might even result in lower performance under specific circumstances.

Do you compare the 40% IPC uplift with a SMT based uplift? Very interesting take. It will be interesting to see, how this may lead to useful predictions.

And of course, SMT will increase power consumption, as does a higher IPC @ iso frequency & iso process. All these obvious things are true. But this is not the whole truth.

 

[FlanK3r]

I'm complaining that obvious sarcasm isn't obvious.

Zen aka Summit Ridge was more than year ago planed at end of the 2016. And now is...? Q2 man You have few months time, so relax!

 

[Qwertilot]

Don't think he's the one who needs to relax

That such heavy sarcasm could be taken seriously by people is of course a rather depressing statement on the level of some of the discussion here!

 

[looncraz]

Samsung isn't manufactoring anything for AMD on 14nm from the looks of it.

The big issue isn't if they make it on Glofo or Samsung. But that they dont make it on the better TSMC 16FF+.

Not sure I'd call 16FF+ superior to 14nm LPP. They're very similar in their known characteristics, but 14nm LPP has better density and (at least currently) yields.

In theory, 14nm LPP should have 15%+ faster switching performance while having less current tolerance, but whether or not this is an advantage for one over the other depends on the product design and the libraries.

So, really, the only details we have make 14nm LPP superior to 16nm FF+. You should be able to have higher clock speeds at lower power, reduced area (somewhat), and you have more process maturity - which is good for libraries, yields, and certain adoption costs.

 

[The Stilt]

Polaris GPUs should give us a slight idea about the characteristics of 14nm LPP. If AMD won't be able to ship even the smaller Polaris "11" > 1500MHz default clocks, then most likely it is the process to blaim. Bonaire ASICs have been shipped at 1200MHz by the ODMs and they overclock close to 1400MHz or even above on air.

All companies except AMD seem to be avoiding the 14nm LPE/P like the lep'ered one. Even Apple abandoned it after a single product?

 

[KTE]

SMT might even result in lower performance under specific circumstances.

Do you compare the 40% IPC uplift with a SMT based uplift? Very interesting take. It will be interesting to see, how this may lead to useful predictions.

For 1C running SPEC CPU workloads, yup. For 4C that'll equate to a 2x gain in throughput.

Personally, I think AMD should jump to TSMCs 10nm which is ramping year end.

I don't want them to make the F10h @ 65nm mistake all over again. I'm more worried about their clocks/power than the uarch performance.

I'm not comparing but just look at the problem Qualcomm's SD 820 is having, after the 810 saga. Samsung 14nm, two fast cores and two slow cores, billed and previewed as 1.3-2.0x performance at 30% less power. The reality? Paper stats mean very little.

The performance is rubbish after about 10 seconds under load. The devices (i.e. HTC 10) overheat big time -> causing thermal throttling -> causing lag and stuttering, migrating all the load to the slower Kryo cluster. The battery life is also 2007 level unless you primarily use Whatsapp and have a high standby time.

Sent from HTC 10

 

[itsmydamnation]

All companies except AMD seem to be avoiding the 14nm LPE/P like the lep'ered one

you mean like the entire high end android market .......

 

[ShintaiDK]

All companies except AMD seem to be avoiding the 14nm LPE/P like the lep'ered one. Even Apple abandoned it after a single product?

Qualcomm needed a devils pact to go 14LPP instead of 16FF+. Foolish company just for getting into the S7.

 

[Lepton87]

Qualcomm needed a devils pact to go 14LPP instead of 16FF+. Foolish company just for getting into the S7.

How foolish of them to have made just US$ 5.3 billion in profits last year. We don't need such foolish companies. With the way they are bleeding money we probably won't see them in 10 years. Good riddance.

 

[ShintaiDK]

How foolish of them to have made just US$ 5.3 billion in profits last year. We don't need such foolish companies. With the way they are bleeding money we probably won't see them in 10 years. Good riddance.

Their chip volume was down 19% YoY.

 

[Dresdenboy]

For 1C running SPEC CPU workloads, yup. For 4C that'll equate to a 2x gain in throughput.

Personally, I think AMD should jump to TSMCs 10nm which is ramping year end.

I don't want them to make the F10h @ 65nm mistake all over again. I'm more worried about their clocks/power than the uarch performance.

I'm not comparing but just look at the problem Qualcomm's SD 820 is having, after the 810 saga. Samsung 14nm, two fast cores and two slow cores, billed and previewed as 1.3-2.0x performance at 30% less power. The reality? Paper stats mean very little.

The performance is rubbish after about 10 seconds under load. The devices (i.e. HTC 10) overheat big time -> causing thermal throttling -> causing lag and stuttering, migrating all the load to the slower Kryo cluster. The battery life is also 2007 level unless you primarily use Whatsapp and have a high standby time.

Sent from HTC 10

Are you sure that this is just a matter of the used process? As it seems, there were also heat problems with older Snapdragons (20nm). 20nm vs. 14nm was never promised as a big performance jump at iso power (35% at best). S7 also doesn't seem to do that bad. Same for the S6 with big.LITTLE. Even the 14nm Apple A9 throttles less while running higher frequencies than its predecessor. So is there a systematic problem with the process or is it more related to the SoC and/or phone designs?

My sources:
http://www.notebookcheck.net/HTC-10-Smartphone-Review.165871.0.html
http://www.notebookcheck.net/Samsung-Galaxy-S7-Edge-Smartphone-Review.161628.0.html
http://www.notebookcheck.net/Samsung-Galaxy-S6-Edge-Smartphone-Review.140746.0.html
http://www.notebookcheck.net/HTC-One-M9-Smartphone-Review.142065.0.html
http://arstechnica.com/apple/2015/09/a-3d-touch-above-the-iphone-6s-and-6s-plus-reviewed/4/

 

[Dresdenboy]

Their chip volume was down 19% YoY.

Can you show the link to the process?

 

[Justinbaileyman]

Today is Computex day so where the hell is Zen or Bristol Ridge?? Usually there are shit loads of leaked info and picks or video's hours before the event. Damn it all this waiting is killing me.. Come on guys show us something for god sakes..

 

[Rickyyy369]

Today is Computex day so where the hell is Zen or Bristol Ridge?? Usually there are shit loads of leaked info and picks or video's hours before the event. Damn it all this waiting is killing me.. Come on guys show us something for god sakes..

I just want to see some AM4 boards.

 

[The Stilt]

Wouldn't be surprised if the "7th" gen. APUs initially lauch as mobile only (FP4) and AM4 stuff follow closer to the launch of Zen.

 

[ShintaiDK]

Can you show the link to the process?

You mean their financials?

 

[AtenRa]

Today is Computex day so where the hell is Zen or Bristol Ridge?? Usually there are shit loads of leaked info and picks or video's hours before the event. Damn it all this waiting is killing me.. Come on guys show us something for god sakes..

Watch the Live webcast on 5/31 at 10pm EST- Sign Up Now

http://www.amd.com/en-us/who-we-are/corporate-information/events/computex

 

[ET]

Wouldn't be surprised if the "7th" gen. APUs initially lauch as mobile only (FP4) and AM4 stuff follow closer to the launch of Zen.

I'd be surprised and disappointed. I think that AMD needs something half-modern in the desktop marketplace, and DDR4 plus future upgradability plus low power is a good start for having something with a bit of appeal.

The only reason I can see to not release it now on desktop would be if not enough are produced to satisfy mobile demand, and given that yields shouldn't be a problem I doubt that would be an issue.

 

[DrMrLordX]

There were those shipping manifests from last year that showed reference boards going out to OEMs. They could have been ready by now if they really wanted to launch AM4. Why are they holding back?

 

[dark zero]

I feel that we will see Bristol and Stoney Ridge together and some Zen tease.

 

[nenforcer]

Wouldn't be surprised if the "7th" gen. APUs initially lauch as mobile only (FP4) and AM4 stuff follow closer to the launch of Zen.

They already have haven't they? Wasn't that the Carrizo chips which were in some HP laptop (EliteBook 755 G2/3) products last year? Or was that 6th gen? The Bristol Ridge is just suppose to be Carrizo in desktop format still on .28nm is it not?

 

[Dresdenboy]

You mean their financials?

This would only be enough for oversimplified models, or to support an agenda. Which roles did management, design, efforts of other market participants, random effects, etc. play here?