New Zen microarchitecture details

[Atari2600]

Again... What other explanation do you have? Have you ever idea of what the FO4 is? Enlighten me, please...

What other explanation do I have for an AMD CPU of an approximately similar approach on an approximately equivalent process to its Intel counterparts achieving approximately similar clock speeds?

Nothing magical - just that its a fairly converged solution in both design and build which will result in fairly converged results.


Fixating on a measurement of inverter delay, which just happens to also be used as a unit of nomalisation means nothing. You've no idea of the exact correlation of a normalised F04 to actual results from a given process, nor do you have any means of comparing across uArchitectures. For all we know, the FO4 of Zen and of Haswell could be closer than the end frequencies of both products - the difference resulting mostly from the process and not the design.

Your broadly right in pointing at FO4 as a means of measuring design speed - but its too complicated to simplify to one parameter. You should know that.

 

[krumme]

but but but.... but..... but....... FO4.....

You clearly gotta be carefull about post like this. I mean 10 minutes after this post and the first 3.6 base rumors came up. I will say you were a bit unluckly and normally ones errors will be hidden in time

 

[Redentor]

That is step F3...

According to canard PC, they have step F4 and is 3.6/4.0 https://twitter.com/CPCHardware/status/817044837358780416 (I suppose still ES)...

http://semiaccurate.com/forums/showpost.php?p=268286&postcount=3372 LOL, 1 year ago

 

[The Stilt]

3.9/3.6

Nice! This sample should already close the gap or exceed it with the 6900k vs the fixed 3.4GHz Ryzen at new horizon.

Extremely impressive if true, especially at 95W. By true I mean that it really sustains 3.6GHz when all cores are stressed properly (not talking about Prime95 either).

I'm starting to feel that Intel might have just "dropped the soap in the jail shower"

 

[bjt2]

What other explanation do I have for an AMD CPU of an approximately similar approach on an approximately equivalent process to its Intel counterparts achieving approximately similar clock speeds?

Nothing magical - just that its a fairly converged solution in both design and build which will result in fairly converged results.


Fixating on a measurement of inverter delay, which just happens to also be used as a unit of nomalisation means nothing. You've no idea of the exact correlation of a normalised F04 to actual results from a given process, nor do you have any means of comparing across uArchitectures. For all we know, the FO4 of Zen and of Haswell could be closer than the end frequencies of both products - the difference resulting mostly from the process and not the design.

Your broadly right in pointing at FO4 as a means of measuring design speed - but its too complicated to simplify to one parameter. You should know that.

Base clock of 6900K is 3.2GHz and has 140W TDP.
Zen ES has 3.6/4.0 and rising at 95W...
And INTEL process is allegedly better... Or not? This is that you are saying?
We saw already this on the 28nm BULK: FX 8370E @95W is 4.0 base and 4.3 turbo. Way more than 6900K... 28nm is magical? No. It's the FO4.

 

[krumme]

Your broadly right in pointing at FO4 as a means of measuring design speed - but its too complicated to simplify to one parameter. You should know that.

Ofcource its a simplification.
I wouldnt expect AMD to get the same results as eg Intel.
As Papermaster said there was 300 people working on the core. Now look at Intel. Is it a magnitude or more higher?
There is a lot of hard hand made work that goes into getting those high freq results. It demands man power to get into every little detail and get the most of it - the complexity you mention.
So personally for me its way beyond expectations what they do here. If 3.6 is the base in 95w tdp 8c for topend its crazy imo. But lets see. I am still not convinced.

 

[Atari2600]

You clearly gotta be carefull about post like this. I mean 10 minutes after this post and the first 3.6 base rumors came up. I will say you were a bit unluckly and normally ones errors will be hidden in time

Its quite annoying when something so complicated is reduced to a single number when simplifying so far means its essentially meaningless.

For instance, I'd be completely misleading you by saying the only reason Mercedes won the F1 world championship this year is because they had Mercedes engines. [In effect saying the drivers, tyres, aerodynamics and chassis played no part whatsoever - even if its reckoned the Mercedes engine was the best in the field by a small amount.]

 

[Atari2600]

No. It's the FO4.

FO4, as you use it, is a measurement, not a feature.

 

[Atari2600]

As Papermaster said there was 300 people working on the core. Now look at Intel. Is it a magnitude or more higher?

More probably.

Zen is a win for those (like myself) that would say good Engineers led by brilliant Engineers will produce a better product than any Engineers led by managers*.

*with their MBAs, slick powerpoints and f**k all actual knowledge.

 

[bjt2]

Obviously also the process and the final actual design counts. But if on a worse process you see way higher frequencies, can't be other than the FO4. Then the other details gives you +-200MHz... But here there is not doubt: +400MHz base with 45W less is only possible with FO4. Or the INTEL process that is shitty...

 

[inf64]

3.8Ghz base just in time to fend off SKL-X?

 

[bjt2]

FO4, as you use it, is a measurement, not a feature.

Relative FO4 (the one i was referring) is a design parameter, because you know perfectly the relative FO4 of each circuit you choose to include in your design.
Absolute FO4 is a measure.

 

[bjt2]

3.8Ghz base just in time to fend off SKL-X?

I have not lost hopes of a 4GHz base... But here all deemed this impossible...

 

[swilli89]

Extremely impressive if true, especially at 95W. By true I mean that it really sustains 3.6GHz when all cores are stressed properly (not talking about Prime95 either).

I'm starting to feel that Intel might have just "dropped the soap in the jail shower"

You do start to get that feeling..

 

[krumme]

Its quite annoying when something so complicated is reduced to a single number when simplifying so far means its essentially meaningless.

For instance, I'd be completely misleading you by saying the only reason Mercedes won the F1 world championship this year is because they had Mercedes engines. [In effect saying the drivers, tyres, aerodynamics and chassis played no part whatsoever - even if its reckoned the Mercedes engine was the best in the field by a small amount.]

I agree and i understand what you mean. And it would have been better if we had more info and ways to talk about it. The FO4 was repeated so much a few nearly got a brain damage.

I always assumed that the advantage Intel got from beeing able to design to the process, and handle the integration here really tight, would give them better results than AMD eg for getting those higher freq. I am not so sure today of this integration benefit.

Its a handfull of years since i worked in a huge technology based company and to be fair what i watched there reflect that its difficult to get those integration benefit between production and product design. The feedback from production to production technology weaken the information, and from here to product design its weakened again. Its like the info going back to design is just the basic that is really needed.

The basic problem of this for Intel is the entire vertical integration benefit and with the need to fill a lot of expensive fabs thats a serious challenge to their business model.

Btw: Lets have a look at F1 and the low budget HAAS team next year. Perhaps you dont have to make your own gearbox, suspension and even chassis to get next to Ferrari. Williams might complain and an old Massa and the fonded rookie might get the problems worse. Lets see of the integration benefit here

 

[inf64]

I'm not sure people realize how fast this 3.6/4Ghz part will be . It will be brutal across the board and likely good deal cheaper than any 8C intel part.

 

[raghu78]

3.8Ghz base just in time to fend off SKL-X?

I would be damn happy with the current clocks 3.6 Ghz base / 4.0 Ghz turbo we are seeing on the F4 stepping. AMD can refine the CPU further in H2 2017 with another stepping to probably hit higher clocks. I am excited for what Zen+ can achieve on a very mature 14nm process in 2018. Maybe 4 Ghz base clocks on octa core ain't too far away. If AMD keeps up the pace we are likely to see Zen+ at 4 Ghz base clocks in 2018. Now that would be a blast.

 

[Atari2600]

Perhaps you dont have to make your own gearbox, suspension and even chassis to get next to Ferrari. Williams might complain

Others have taken entire drivetrains before... Sauber spring to mind - not sure if they used the rear suspension or not.


Customer chassis was shot down a few years back... erm, actually probably quite a few years back now - maybe 8-10 yrs due to fears from the midfielders of them being marginalised by last years front running chassis run by customers.

 

[iBoMbY]

I'm not sure people realize how fast this 3.6/4Ghz part will be . It will be brutal across the board and likely good deal cheaper than any 8C intel part.

Yes, it currently looks like AMD is going to blow Intel out of the water in 2017. That would be truly amazing, if it all turns out to be true, and AMD can deliver the CPU in numbers. A real game changer after a long drought.

 

[hojnikb]

Well that sux. No overclocking on budget chipsets.

 

[Doom2pro]

Well that sux. No overclocking on budget chipsets.

At least the built in overclocking is there.

 

[laamanaator]

Well that sux. No overclocking on budget chipsets.

Except that it doesn't suck. Those boards which are fitted with an A320 chipsets are also going to have extremely poor VRMs. I'd be amased if an A320 board is able to support 8c ryzen without throttling because of high VRM temps. Disabling overclocking on the low end boards is going to save someones house from burning down.

 

[rvborgh]

i am really liking how the clock speeds are starting to pick up on the samples. Anyone know how quickly they can respin samples? and how many spins are typically "in process" at the same time? Anyhow... very impressive AMD. Looking forward to buying a bunch.

 

[scannall]

I think Zen will be a really great product. But I also think some people are getting a little wild in their expectations. Guys, leave a little on the table for now. Wait until it's out in the wild, and on shelves and tested. If you leave a little on the table you get to be pleasantly surprised if it does better than that.

 

[Zstream]

B350 is all I need!