New Zen microarchitecture details

[Doom2pro]

That s right, the serie of 6 is likely for the IGP power plane..

Dont know for the GTX but D-Pack have only three connectors on one side and a metalic part on the other side for soldering, wich is not the case here..

GTX cards come with many VRM designs, the one I was referring to has three SO-8 packages per phase, aligned exactly like this OEM board.

Here is an example of a GTX with three SO-8 packages per phase:

And here is a D-PAK design for comparison:

And here is a blown up view of the AM4 OEM VRM:

SO-8?

It's too bad the images are blurry...

 

[sirmo]

Those D-Pack components are MOSFETs, they are basically rolling their own discrete VRMs.

Datasheet for the one in the second picture n-channel MOSFET (PDF): http://www.unikc.com.cn/UploadFile/products/2015713145449-UNIKCP0903BDL_Ver1.1_20130315.pdf

 

[Dresdenboy]

That's an interesting discussion!

If you're so much after identifying these parts, you might try a deconvolution tool as a sharpener. This is one of the best methods for motion blurred pictures. With even more time at hand, one might stack cut out images of the individual components with subpixel accuracy to sample them to a higher res image.

Examples and some software: http://smartdeblur.net/index.html

 

[itsmydamnation]

Hmm guess I was mistaken on that point. I guess I could see it not outperforming XV by all that much on a core vs. module basis then, which is a bit disappointing.

If you consider that on average for FP workloads something like 40% of ops have a memory operation ( either load or store) and you can only load 2x128bit and store 1x128 the likelyhood that you would be able to sustain anywhere near 4xFMA a cycle is so remote its not funny.

really the FPU is probably designed to be able to do a sustained peak around 2 ops a cycle. the idea of 4 ports is about reducing latency of non FMA adds and muls nothing more,when your goal is much improved IPC* this looks to be a good design choice.

IPC meaning increase in performance from low ILP workloads.

 

[The Stilt]

Good clocks for A12-9800. I wonder if there's an A12-9850K model which is around 4Ghz.

A12-9800 is supposed to be the fastest model. AFAIK it will remain so, i.e no "K" models coming, at least for now.

 

[BigDaveX]

Carrizo really wasn't optimized for high clockspeeds, so unless AMD have done some major tweaking for Bristol Ridge, there'd probably be no point to making an overclocker-friendly version.

 

[dark zero]

Carrizo really wasn't optimized for high clockspeeds, so unless AMD have done some major tweaking for Bristol Ridge, there'd probably be no point to making an overclocker-friendly version.

BR ended to be a Richland with steroids. Also and interesting enough is pending if they added back 2 Mb L2 cache. That could make finally end into a Phenom II levels and even Nehalem levels.

 

[The Stilt]

Also and interesting enough is pending if they added back 2 Mb L2 cache.

How can it be "pending" when AMD has released the die shots for "both" Carrizo and Bristol Ridge?

Carrizo on top, 50% opacity with better alignment this time.

The outer region of the picture (from BR die shot) show parts of eight different dies in total. The center part is the only essential one.

 

[MrTeal]

That's an interesting discussion!

If you're so much after identifying these parts, you might try a deconvolution tool as a sharpener. This is one of the best methods for motion blurred pictures. With even more time at hand, one might stack cut out images of the individual components with subpixel accuracy to sample them to a higher res image.

Examples and some software: http://smartdeblur.net/index.html

It looks like a pretty standard single highside and dual lowside VRM design, so four phases for the CPU core. Performance won't be up there with a good driver+fet integrated solution, but it should be fine at 95W if it's done well.

 

[CHADBOGA]

7 months left in the year and they already shown that they have silicon that is clocking up to expectations. If anything I'm more confident now.

An update on Zen that may cause you to re-assess.

http://forums.anandtech.com/showthread.php?t=2476153

 

[Dresdenboy]

http://www.jagatreview.com/2016/06/computex-2016-amd-pamerkan-wujud-prosesor-zen/

 

[Abwx]

http://www.jagatreview.com/2016/06/computex-2016-amd-pamerkan-wujud-prosesor-zen/

Hardware.fr are also at Computex and they are talking of the schedule :

Lisa Su a précisé que ses plus gros clients recevraient les premiers échantillons d'ici quelques semaines et que la montée en puissance de la production débuterait au 3ème trimestre, à priori à temps pour un lancement vers la fin de l'année.

http://www.hardware.fr/news/14653/computex-computex-zen-qui-presente-zen.html

Demo of the RX480 was made with a system using Zen...

 

[Doom2pro]

http://www.jagatreview.com/2016/06/computex-2016-amd-pamerkan-wujud-prosesor-zen/

Hello... AM4 has a belly button... Pin density is pretty high already... I foresee this square shrinking over time

 

[dark zero]

http://www.hardware.fr/news/14653/computex-computex-zen-qui-presente-zen.html

Demo of the RX480 was made with a system using Zen...

Ok this is interesting... the Doom demo had similar performance from a system that uses Haswell E as CPU and a GTX 980 as GPU (and considering that drivers are unpolished it might get better)
So Zen might be as powerful as a Broadwell in MT and a little better than HW in ST.

 

[KTE]

I already predicted its a Q1 2017 mass launch, and I don't see anything changing my expectations.

This '40% IPC' is shown against Vishera most recently. Realistically, I can't see it being IPC in general either and it's gonna come back to bite AMD hard if this product isn't some magic gold dust.

Anyway samples to partners very soon... Leaks

Sent from HTC 10

 

[inf64]

I already predicted its a Q1 2017 mass launch, and I don't see anything changing my expectations.

This '40% IPC' is shown against Vishera most recently.

Source please? IPC was always shown vs EX core (and core vs core not 1C/2T vs 1M/2T). EX core is around 15% faster than PD core on average, bar some +/- corner cases.
So 40% over EX core is around 60% over PD core. Then there is SMT bonus in MT workloads where PD module has 15-20% penalty.

 

[KTE]

Why AMD Is Not In A Hurry To Release Zen-Based Processors
http://seekingalpha.com/article/3980601?source=ansh

I still remember this:
"Suzanne Plummer, Zen team leader went on record to say this in September 2015: “This is the first time in a very long time that we engineers have been given the total freedom to build a processor from scratch and do the best we can do. It is a multi-year project with a really large team. It’s like a marathon effort with some sprints in the middle. The team is working very hard, but they can see the finish line. I guarantee that it will deliver a huge improvement in performance and (low) power consumption over the previous generation.”

Source please? IPC was always shown vs EX core (and core vs core not 1C/2T vs 1M/2T). EX core is around 15% faster than PD core on average, bar some +/- corner cases.
So 40% over EX core is around 60% over PD core. Then there is SMT bonus in MT workloads where PD module has 15-20% penalty.

Sorry, my mistake there.

Not 40%...

But Zen comparison against Vishera as the last generation core. Like:

And even when AMD talked about the process improvement, they referenced PD.

Sent from HTC 10

 

[inf64]

Sorry, my mistake there.

Not 40%...

But Zen comparison against Vishera as the last generation core. Like:

And even when AMD talked about the process improvement, they referenced PD.

Well yes, performance vise(when you count in the clocks) 3.2Ghz base clock 8C/16T Zen should be around 2-2.2x MT performance of 8350. ST should be around 40-45% better, stock VS stock provided Zen can reach around 3.7Ghz in Turbo mode. This should be enough for Zen to roughly match 3Ghz 8C/16T Haswell-E or higher clocked hex core Broadwell-E parts.

 

[ShintaiDK]

Well yes, performance vise(when you count in the clocks) 3.2Ghz base clock 8C/16T Zen should be around 2-2.2x MT performance of 8350. ST should be around 40-45% better, stock VS stock provided Zen can reach around 3.7Ghz in Turbo mode. This should be enough for Zen to roughly match 3Ghz 8C/16T Haswell-E or higher clocked hex core Broadwell-E parts.

Even AMDs own most super optimistic cherry picked prediction didn't give that.

 

[The Stilt]

3.2GHz 8C/16T providing 2 - 2.2x MT performance of a FX-8350 would mean that the IPC has increased by 100 - 120%. Either you expect Zen to be > 73% faster than Excavator, or Excavator to be 42.85% faster than Piledriver. Either way, both figures differ significantly from AMDs own figures / expectations: PD to XV = 15.5% (average), XV to Zen <= 40%. That's with 25% SMT yield, which is pretty optimistic IMO.

With those specs Zen would have significantly higher IPC than Haswell / Broadwell / Skylake and most likely Kaby Lake.

 

[Arachnotronic]

3.2GHz 8C/16T providing 2 - 2.2x MT performance of a FX-8350 would mean that the IPC has increased by 100 - 120%. Either you expect Zen to be > 73% faster than Excavator, or Excavator to be 42.85% faster than Piledriver. Either way, both figures differ significantly from AMDs own figures / expectations: PD to XV = 15.5% (average), XV to Zen <= 40%. That's with 25% SMT yield, which is pretty optimistic IMO.

With those specs Zen would have significantly higher IPC than Haswell / Broadwell / Skylake and most likely Kaby Lake.

If I were in charge of making the marketing slide for AMD, I wouldn't compare Summit Ridge to FX-8350, I would compare it to FX-8150 for maximum effect -- which is probably what they did.

The Stilt, back on May 17 you said that the official word is now "up to 40%" more IPC than XV. This makes sense because if AMD were able to hit, say, +50% in some scenarios they would be shouting it from the rooftops.

Zen is going to be a lot better than the Bulldozer-derived stuff that's currently in the market, but the dreams of it being a Skylake killer just seem too optimistic.

 

[The Stilt]

If I were in charge of making the marketing slide for AMD, I wouldn't compare Summit Ridge to FX-8350, I would compare it to FX-8150 for maximum effect -- which is probably what they did.

The Stilt, back on May 17 you said that the official word is now "up to 40%" more IPC than XV. This makes sense because if AMD were able to hit, say, +50% in some scenarios they would be shouting it from the rooftops.

Zen is going to be a lot better than the Bulldozer-derived stuff that's currently in the market, but the dreams of it being a Skylake killer just seem too optimistic.

Since AMD said "over a current generation core" multiple times in different occasions, I don't think it can really be the Orochi Rev. B (Bulldozer) they were comparing Zen to. The original Bulldozer is no longer a current generation core, since it has been out of production for a good while now.

Also I don't expect that AMD would even compare the IPC of their own cores between each other in that manner (directly). I believe the Orochi chart illustrated performance-per-watt figures than anything else.

As as I said before, I believe AMD compared 95W Orochi Rev. C (Vishera, e.g. FX-8370E) to 95W Zeppelin in that chart. At least that's what I would personally do if I had to come up with a representive chart to show some impressive figures, while the actual IPC is still behind the competition. In that scenario much of the merit would belong to the smaller process node and not necessarily the design itself.

 

[Arachnotronic]

Since AMD said "over a current generation core" multiple times in different occasions, I don't think it can really be the Orochi Rev. B (Bulldozer) they were comparing Zen to. The original Bulldozer is no longer a current generation core, since it has been out of production for a good while now.

Interesting, but I would argue that they never mentioned "current generation core" on that particular slide.

Also I don't expect that AMD would even compare the IPC of their own cores between each other in that manner (directly). I believe the Orochi chart illustrated performance-per-watt figures than anything else.

Interesting.

As as I said before, I believe AMD compared 95W Orochi Rev. C (Vishera, e.g. FX-8370E) to 95W Zeppelin in that chart. At least that's what I would personally do if I had to come up with a representive chart to show some impressive figures, while the actual IPC is still behind the competition. In that scenario much of the merit would come from the smaller process node and not necessarily from the design itself.

That would make sense given that FX-8370E is a 95W part and so is Summit Ridge. I like your thinking.

 

[sm625]

Why AMD Is Not In A Hurry To Release Zen-Based Processors
http://seekingalpha.com/article/3980601?source=ansh

From the article:

The current inventory of 14-nm Atom processors, not Apollo Lake, makes it undesirable for AMD to release entry-level Zen processors now. Intel can quickly dump these Atoms on its dozens of PC vendor partners to immediately diminish the Zen's chances of commercial success.

What does atom garbage have to do with Zen? Intel has been dumping atom garbage for years. What good has it done them?

 

[Arachnotronic]

From the article:



What does atom garbage have to do with Zen? Intel has been dumping atom garbage for years. What good has it done them?

I don't think this is a good article.