AMD: ZEN

Not much new info, but some new slides released.
Anandtech has a run down, Videocardz too.
anandtech.com/show/10578/amd-zen-microarchitecture-dual-schedulers-micro-op-cache-memory-hierarchy-revealed
videocardz.com/63242/amd-zen-stuff

Other urls found in this thread:

youtu.be/oQS8s7TOXsE?t=90
pcper.com/reviews/Processors/AMD-Zen-Architecture-and-Performance-Preview
arstechnica.com/gadgets/2016/08/amd-zen-performance-details-release-date/
anandtech.com/show/10578/amd-zen-microarchitecture-dual-schedulers-micro-op-cache-memory-hierarchy-revealed/2
anandtech.com/bench/product/1684?vs=288
ark.intel.com/products/94196/Intel-Core-i7-6900K-Processor-20M-Cache-up-to-3_70-GHz
anandtech.com/bench/product/697?vs=53
anandtech.com/bench/product/697?vs=142
twitter.com/SFWRedditGifs

...

>single core performance of a 2500k
Ducking dropped.

Pretty much not a single point of interest on this slide

AMD restating that 14nm LPP shows an increase in fmax at lower power vs 28nm HPP.
Already saw this in Polaris vs older GCN cards, but AMD's Carrizo is fabbed on the same 28nm HPP process, and it can hit north of 3.5ghz without issue. Bodes well for lower core count parts like their Raven Ridge APUs.

The point here about clock gating is worth reading into. Logic within each core can likely be gated, as well as a complete CCX. Gating off an entire CCX and pushing 4 cores to a really high pstate to consume the entire TDP would make an 8 core Summit Ridge slight more competitive in serial workloads.

Rehashing what they've already shown.
40% uplift over Excavator, future Zen+ slated to also bring a decent performance uplift.

Hey, who doesn't want this to be a success and there to be real competition in the CPU market?

After their misleading rx480 launch event I'm shelving my hype until this thing is released and benchmarked.

But I do want it to be a big success.

Compare to data taken on Zen from a linux kernal patch, visualized with a little guesswork from Dresdenboy.

Pretty spot on.

youtu.be/oQS8s7TOXsE?t=90
AMD claims clock to clock broadwell-e performance

Looks almost identical except for the sizes of some of the caches I think

Summit Ridge die shot briefly analyzed by Hans de Vries of Chip Architect

AMD made absolutely no such claim.
A 40% uplift over Excavator does not come near Broadwell in performance per clock. Being "competitive" does not mean equaling performance.
Don't spread bullshit like this.

Oh and the micro-op queue looks like it's arranged in a different way.

And for good measure here are the leads of the Zen design team.
The woman up front is Suzanne Plummer, the lead architect who Zen is named after.

That's actually quite good. Considering the fact that skylake has like 20% better IPC. Having access to multiple cores at a competitive price is due time.

"Using Blender to measure the performance of a rendering workload (a Zen CPU mockup of course), AMD ran an 8-core / 16-thread Zen processor at 3.0 GHz against an 8-core / 16-thread Broadwell-E processor at 3.0 GHz (likely a fixed clocked Core i7-6900K). The point of the demonstration was to showcase the IPC improvements of Zen and it worked: the render completed on the Zen platform a second or two faster than it did on the Intel Broadwell-E system." - pcper.com/reviews/Processors/AMD-Zen-Architecture-and-Performance-Preview

You were saying?.

Not my words watch the video, they only compare it to broadwell-e in a blender test not in a variety of tasks

>AMD made absolutely no such claim

Yes they did retard.

Learn what context is.
One very specific blender scene is not general performance, and should never be construed as such.
An average 40% IPC uplift over Excavator is the key metric, and this has never been revised. Performing slightly better in some metrics does not mean otherwise. It will show more of an improvement in some areas and less in others.

Mindlessly regurgitating this bullshit is exactly why so many people got butthurt over Bulldozer. Tech illiterate retarded children like you were everywhere online hyping it up to the moon.

Shee... it.

Intel Status: Ass-blasted.

No wonder they're laying people off.

...

>Context
Go sperg at AMD over their marketing material

Video says the rendering was also sped up 4x.

For both I mean, the second difference might not actually be that small in reality.

Might just be more marketing bs though

>Video says the rendering was also sped up 4x.

A... are you really that retarded?

You're supposed to compare when each finished. They sped it up because watching a scene render for 2 minutes is fucking boring.

AYYMD IS FINISHED & BANKRUPT

The problem isn't AMD marketing material. The problem is retards like you and this child: AMD says or shows one very specific thing, and you dumbshits extrapolate it to mean something entirely different. You just repeating
>Hurrr ZEN WIL MATCH BROADWELL INTEL IS FINISHED XDDD

Then in reality launch day comes, reviews go live, and Sup Forums is filled with threads full of benches showing AMD losing in tons of metrics. Because thats what is going to happen. The Zen core doesn't have a big FPU, they flat aren't going to touch intel's performance here. In virtually every video encoding bench used by sites like Anandtech and Tom's Summit Ridge isn't going to come out on top.

The difference would be an even better result for AMD since the sped up video shows Zen finishing 1 second faster than Intel.

>Because thats what is going to happen. The Zen core doesn't have a big FPU

Yeah, they only have twice as many as the current FX chips...

That's what I was trying to say. The video finishes fast for both but it was also sped up 4x so the 1-2 second difference could be more like 4-8 seconds which would be a lot more noticeable in real in real life.

Videos of benchmarks are highly questionable though and blender isn't necessarily indicative of real world performance though.

Supplemental chart from anand
Big change is AMD changing from write through to write back cache, the L3 is also no longer a victim cache. Its somewhat telling that they're still throwing more SRAM at low levels in the cores than intel is. If the core was tighter they wouldn't need the size, and could instead benefit through the increased speed of a smaller cache.

>The problem is retards like you
No, the problem is AMD, they should know that Tech media(TM) will claim AMD has Broadwell-E performance with that wording. They are doing the hyping and they will bleed for it.
arstechnica.com/gadgets/2016/08/amd-zen-performance-details-release-date/

Also useful.
cache bandwidth comparison between 5ghz Vishera and 4.5ghz Sandy Bridge-E.
Appropriately illustrates where a huge portion of Piledriver's IPC deficiency comes from. AMD's memory controller is also decidedly inferior, though it is dual channel vs quad here.

I already dread the months following the release. We will see 9001 gaming """benchmarks""", FPS counters of how proprietary non-free games, that are 100% irrelevant to anyone over the age of 12, """perform""".

>Zen delayed until 2017

AYYMD IS FINISHED & BANKRUPT

BLUNDER OF THE CENTURY

Hahahahaha !! :-D

kill yourself autist

There was no delay. Just Ars regurgitating a digitimes clickbait article.
Early volume end of the year 2016, full volume 2017. Thats what AMD has been consistently saying for well over a year now, restating that everything was right on track at Computex earlier this year.

AMD IS FINIShED AMD IS FINISHED THE PRODUCT IS COMING OUT LATER THAN EXPECTED AMD IS BANKRUPT DONE DEAD

Really there is a big chance for Zen to be a huge success.

People that want more than 4 cores now need to drop a fuckton of shekels on Intel shit but if Zen will be moderately compararable to the high end Intel CPU's then people as the cheap trash they are, will go for Zen because it will for sure be cheaper.

Win-win for me because i want 6+ cores but for the cheapest intel one i would have to fucking sell myself 5 times to buy it.

And thats without a Mobo.

man the L3 cache is really slow on Vishera. Is K10/10.5 any better?

...

The L3 in the Bulldozer family is a victim cache, its basically used only in select few ops.
I don't think K10 treats its L3 the same way.

Phenom II does make good use of its L3 cache as shown by the gains granted by overclocking it. Bulldozer and Piledriver get shit for overclocking their L3s.

He really did it

Blender is almost a fully integer execution, let AMD try the same with some particle simulation or protein and they'll lose due to the slower FPU

To be fair not many applications are that FPU heavy these days. Also AMD did put more focus into their FPUs so it is possible they'd still be very competitive if they don't match Intel's products.

based 14nm LPP

Funfact "low power plus" is an inane marketing name. It actually stands for laser produced plasma, the finer tuned light source used by the process compared to their hastily ramped up LPE.

>Damn Lisa!
>Back at it again with the based leadership.

wtf they are all older than my dad no wonder amd cpu suck

The absolute madman

>tfw haven't bought an AMD CPU in 10 years

jesus christ for once don't fuck this up, i'm counting on you pajeet.

I agree, most workloads are pretty mixed these days but leaning slightly to integer.

For any kind of complex math you need FP though.

Eh, interesting.
AMD has two schedulers for the int unit and FPU while Intel has one big complex one for both.
This might be another way to save power at the cost of diesize.

>Also AMD did put more focus into their FPUs so it is possible they'd still be very competitive if they don't match Intel's products.

I have no idea where you're pulling this from.
The Zen core is using 128bit FMACs, they still can't execute native 256bit ops. If AMD actually made significantly gains on FPU throughput then average IPC increase over Excavator would be much, much higher than 40%.

Haswell can be up to 80% faster than Excavator in X265 encoding.
TONS of workloads heavily leverage the FPU, especially in content creation.

Well I'm no expert on any of this but Zen has twice as many FPUs available compared to Bulldozer. Seems like having more FPUs available should help considerably shouldn't it?

OOOOOOOOHHHHHHHHHHHHH BOY they beat an underclocked

Zen design started well before Lisa. She has nothing to do with this.

They'll have more independent FPUs since they're moving away from the FlexFPU design, but per core they will not be significantly more powerful.
AMD is not competing with intel's cores on FPU throughput. They won't be close.
Steamroller and Excavator when clock normalized will still lose to a Phenom II in a lot of FPU metrics. Piledriver was embarrassingly far behind in perf/clock.

Thats completely wrong.
Work on the Zen core arch started around 2013-2014. Su has been at the company for years, and she was vetted specifically to be the CEO. Rory was a gap filler appointed interim to find a long term replacement.

Su is the architect behind AMD's financial restructuring and semi-custom business. Projects that were Rory's doing like SkyBridge were shitcanned. Zen was developed under Su's directive. Further refinements of the BD family had been planned under Rory and were canceled.

Two CPUs of the same clock speed with identical specs. The implication is that zen has comparable or better performance/clock than broadwell-e.

There is no implication.
Its explicitly showing one very specific scenario which is not indicative of performance in anything else. Its like looking at Fritz chess for branch prediction performance, or 7zip LZMA compression.

anandtech.com/show/10578/amd-zen-microarchitecture-dual-schedulers-micro-op-cache-memory-hierarchy-revealed/2


>Edit 7:18am: Actually, the slide above is being slightly evasive in its description. It doesn't say how many cores the L3 cache is stretched over, or if there is a common LLC between all cores in the chip. However, we have recieved information from a source (which can't be confirmed via public AMD documents) that states that Zen will feature two sets of 8MB L3 cache between two groups of four cores each, giving 16 MB of L3 total. This would means 2 MB/core, but it also implies that there is no last-level unified cache in silicon across all cores, which Intel has. The reasons behind something like this is typically to do with modularity, and being able to scale a core design from low core counts to high core counts. But it would still leave a Zen core with the same L3 cache per core as Intel.


Apparently the L3 cache in Zen spans a maximum of four cores? I wonder what affect this will have on performance.

Thats just a rumor going around based on people foolishly assuming you can tell how structures are connected by looking at front end logic.
Independent CCX without the ability to snoop each other's L3 would harm performance significantly in a server rack. There would be no reason to link these independent chips together with high bandwidth GMI links if that were the case, and we know that Opterons are exactly that. Dual and quad die MCMs linked together.

Theres a huge cache interface, absurdly huge, sitting right in the middle of the L3 on the Summit Ridge die.

I payed 500 for 5820k, mobo, and 16gb 2133mhz.

The Jaguar compute cluster is four cores with a 2MB block of L2 cache all connected by a relatively beefy cache interface that provides coherency between all cores.
The PS4 and Xbone bone use stock IP from AMD so their APUs feature two of these clusters. Sony and Microsoft both patched them together to connect the two units. The PS4 is basically just linked together on the back end, while the Xbone has slab of SRAM between the two units to provide a fast intermediary which actually improves multicore scaling over the PS4.
Point being you can't judge how an IC works and signals based on looking at the front end. Structures do not need to be physically near each other to be connected.

Theres the Xbone for comparison

Summit Ridge die here: The cache unit in the Zen CCX is easily identifiable. Each core has a private L2, the L3 is the largest part of the CCX by far, and the cache unit is the single largest structure within that slab.

If zen can clock at same speeds there's no sense to underclock the broadwell one, so they probably can't do it.

This is like comparing diesel vs gasoline engines limiting the gasoline rpm to 4000 cause the diesel one can't go higher.

what the frick! ...I hate AMD now

>I have no idea how anything works: The Post

AMD and intel CPUs have totally different turbo behaviors, they also have totally different TDPs. For an equal comparison they each have to run at a fixed frequency, turbo off.
Why on earth would they arbitrarily overclock the CPUs in their test setup to show a simple perf/clock comparison? Perf/clock is the same regardless of what the clock is. There is no reason to overclock them, and every reason to have each chip run at a low fixed frequency that is close to what they'd each be running at fully loaded under normal conditions.

There's a couple of reasons I can think of why they'd run them at a "low" frequency. One obvious one is that they don't want Intel to see how the chips perform and then lower the prices of their comparable chips ahead of their launch and ruin it. The final specs for a product that hasn't launched yet have to be secret.

If your chip can't clock as high as the competency a perf/clock comparison has no sense at all.

-Hey look my 3ghz cpu is 1% faster than the intel one in blender if i underclock it to 3ghz too.
-Oh that's great, can you clock it at the same speed that intel defaults?
-No we can't clock that high.
-what a pity

Finally I can get the performance of a 2500k at the price of what a 2500k is worth now, coupled with a RX 480 i can get the performance of a gtx 970!

Is barely catching up to old tech AMD's business model now?

>shill: the post

>truth: the post

what a time to be alive

It's also likely an engineering sample, which means clocks are not set to be 3Ghz.

Excavator matches the 2500K. This is much better than that.

I doubt we'll see bulldozer like clocks on zen chips

Nominal frequency on the i7 6950X is 3ghz. Thats the top end of Broadwell-E.
The 6850K touts a 3.6ghz base, the mid range of Broadwell-E with 8 cores.
Haswell-E's high end i7 5960X has a base frequency of 3ghz.


The engineering sample of the 8 core Summit Ridge showed a 3.2ghz turbo. Yet here you are, the tech illiterate shitskin shill, pretending like 400mhz would be an impossible feat for AMD when the nominal frequency for the i7E family's high end is right at 3ghz.

So whats your excuse for being such a dumb little faggot child?

>This is much better than that
knowing AMD and the Poolaris fiasco, it isnt.
I can already see posts on Sup Forums before its release
>ZEN HAS THE PERFORMANCE OF A I7 6950X AT $300 O SHIT WADDUP
>Actual benchmarks get released and its just a shittier Haswell/Sandy Bridge i7/i5
>I-Its just a lie by Intel shills! AMD will be great again!

>Excavator matches the 2500K

No it absolutely does not.
anandtech.com/bench/product/1684?vs=288

Sandy Bridge is 20%-50% faster than Excavator, and thats with Excavator having a frequency advantage. Clock normalized the disparity is even larger.

Don't talk out of your ass about something that can be so easily refuted.

>Nominal frequency on the i7 6950X is 3ghz. Thats the top end of Broadwell-E.
the video is showing a 8c/16t 6900k
ark.intel.com/products/94196/Intel-Core-i7-6900K-Processor-20M-Cache-up-to-3_70-GHz
Not a 10c/20t 6950x
>The 6850K touts
>i7 5960X
Are you talking of every intel cpu except the one in the demo for any reason?

As i said while intel can't clock as fast all bulldozer it seems that zen can't clock as high as intel.

>Amdrones on full damage control.

>implying ZEN, intels flagship and most modern silicon don't cap out at around 4.5 GHz
You don't know jack shit nigger

>ZEN, intels flagship
sure.

Are you such a tech illiterate retarded faggot that you can't even read?
The i7 6900k has a 3.2ghz base clock.
I intentionally referenced the 6850k because it stands out in the line for its relatively high base clock. Its base frequency, higher than the rest of the line, is still only a stone throw away from the clocks shown by Summit Ridge.

3.2ghz base
same as the Summit Ridge ES turbo
>b-b-b-b-but AMD can't reach the same clocks!

This is you desperately grasping at straws

Masterrace in the front, Rajeets in the back

Sure it can render shit but 90% of people arent going to be doing that

Where the single core performance at
if it cant beat my 4590 its worthless

assuming you can clock the Zen near the Broadwell-E chip, great, but they don't do that demo if that's likely

Summit Ridge is not a mainstream desktop chip. It is not positioned to compete against mainstream i5s and i7s.

It is an i7E competitor.

As long as there's a significant improvement over my Q6600, I don't really care how they compare to other CPUs.

Pretty much everything is an improvement over that. Are you kidding?

anandtech.com/bench/product/697?vs=53

>same as the Summit Ridge ES turbo
so now we have to compare intel base with Summit Ridge turbo?

So 3.2 turbo is better than 3.2 base and 4ghz turbo now?

>b-b-b-b-but AMD can't reach the same clocks!
They can't with zen.

Stop the hype so u don't get disappointed on release kid.

The marketing material is intentionally manipulative and vague in order to make it appear better than it actually is. In reality Zen will only match or outperform the I7-980 by a few %.

Yeah, but I didn't want to upgrade to something like the 8350, because I knew the socket was a dead end and I've never have the chance to upgrade the CPU without replacing the motherboard a second time.

>muh games

no intel port entanglement problem?

>being this laughably desperate
Its the frequency of a 95w engineering sample, vs a 140w retail product.
Grasp at more straws, shitskin shill.

anandtech.com/bench/product/697?vs=142
The FX 8350 isn't very far behind the i7 980X. Zen's IPC is around 64%~ higher than Piledriver's.
You shitposters really don't have command over any facts.

dont mind me faggots

They are clearly not diverse enough, no wonder AMD sucks.

>>b-b-b-b-but AMD have lower tdp.
nice strategy change.

>You shitposters really don't have command over any facts.
I saw a bench where zen is barely above an i7-980 but i doubt it'll be the case in most workloads.

Still you don't have to disrespect and insult others just because your brand loyalty.

I want to buy a zen processor if it's good enough but i've seen hype damaging amd image every release so if you want to go that way prepare to be disappointed.