New Zen microarchitecture details

[Sweepr]

Mobile Kabylake is being tested for months now, there's even entries from Samsung notebooks @ SiSoftware.

 

[Arachnotronic]

Knowing AMD? Past performance??AMD has a lot to prove in CPUs????? Were you not around in the a64 days?

That was 13 years ago.

 

[Arachnotronic]

I don't recall anyone saying AMD would flat-out beat Intel in perf/watt with Zen, just that they would get back to a competitive position.

AtenRa has said this.

 

[LTC8K6]

Seems to me that if you are designing for high clocks, it's possibly because you need high clocks...?

 

[Kenmitch]

Thanks for the info. Not sure if I'd buy one but would really like to see how it performs. Might actually be a decent deal combined with a microcenter bundle if performance is up to snuff.

 

[moonbogg]

AMD will be lucky to match sandy bridge. This has been my Hit N Run ZINGER!!

 

[Dresdenboy]

I started a multi-answer but thought, that this would be overkill..

Primarily to elevate low yields AMD is separating the CPU and GPU dies.

If DT and server share CPU and GPU dies, this would reduce some of the high FinFET related costs.

 

[Dresdenboy]

@dresdenboy at a high level, how does zen -so far- compare with *dozer uarch

It should be a significant improvement. I might put them side by side into a table together with some Intel uarchs, as IDC asked for. A short comparison:
Zen vs. XV as ST(MT for BD):
Decode: 4 + uOp$ vs. 4(4)
ALU: 4 vs. 2(4)
AGU: 2 vs. 2(4)
FMAC: 2 vs. 2
FMUL+FADD : 4 vs. 2
L1 D$: 32kB vs. 32kB (64kB)
L1 I$: 32kB? + uOp$ + more ITLBs vs. 96kB
Branch prediction: likely better in Zen, and checkpointing reduces effective branch misprediction penalty
L2$: 512k (faster) vs. 1MB (slow)
power efficiency (process normalized): likely goes up

ST will see a jump from wider uarch, while MT depends on power efficiency.

 

[Dresdenboy]

OTOH we don't know how many FO4 delay does Zen has, so it's hard to judge Zen to have high frequency design as bulldozer, I bet it to be little bit higher delay than K8 but lower than bulldozer.

This FO4 number would tell a lot. For ARM and the 14nm/16nm implemented variants I have some estimations and statements (e.g. from Broadcom), allowing for some filling in the holes between designs at different processes. And although many parts of Zen will also be used in K12, there is not much of a relation between these two worlds.

Some timing constrained solutions described in patents could serve for any estimations. But there likely is only little ROI in finding that out with still a lot of error margin left. So far I think, Zen will use a typical value of 20-25 FO4 delays per stage. BD was 17 FO4 (12+5) design.

To estimate clock frequencies for given voltages, we also need to know, which standard cell libs they use, how their thresholds are for inserting LVT transistors, etc.

 

[Dresdenboy]

Thanks, Dresdenboy, for posting your Zen overview in these sharktel infested waters. It's refreshing to read informative posts about highly technical stuff. Engineering stuff that seems to trigger schadenfreude emotional responses in some.

Someday I may understand the fanboy/tribalism that exists around impersonal commercial brands. Especially in the GPU and CPU areas. Probably more in the GPU market, where the target demographic appear to be more easily influenced by social engineering than the actual hardware engineering.

Has modern marketing become so successful that consumers must belong to a brand in order to belong?

You (and the others) are welcome. Well, we are humans after all, provided with a limbic system. Marketing is a free market's tool to exploit human weaknesses. And somehow a lot of people - at different degrees - love to see fights and competitions, one winning over another, just to enjoy personal BIRGing. It could well be the case, that without emotions as motivating drivers, the tech stuff would be too dry and boring for most people.

 

[AtenRa]

AtenRa has said this.

I have said AMD "could aim for better perf/watt"

Edit:

This is the post.

95W TDP will give you higher yields = lower Chip cost = lower package/heat-sink cost = higher profit.

Also, OEMs will use 95W TDP SKUs for workstations and high end Gaming systems. Broader market penetration = higher volumes = higher profit.

edit.

One more thing, 95W TDP will need cheaper motherboard designs.

Edit 2.

And as i have said before, i strongly believe that AMD targets ZEN for highest perf/watt possible and not highest performance. So at default, 8 Core 16 Thread 95W TDP ZEN may come slower than same Broadwell-E but it could have way higher Perf/Watt than the 140W TDP Intel SKU.

 

[Dresdenboy]

You need more than atomics to support HW transactional memory or do you even need them at all ?

Both are related to the concept of synchronization but their somewhat orthogonal. With atomics you need locks for synchronizing access to a region of memory. With HW transactional memory we now use "transactions" instead to define the critical sections of the code that must have restricted access by some threads. One special attribute unique to transactions is that they have the capability to restore the previous state just before beginning the transaction when there is an access conflict between threads by aborting the current transaction ...

I didn't see much what this "atomic" refers to. But according to patents, the checkpointing capability is thought as a mechanism for implementing transactional memory.

 

[Vesku]

To be comparable, we need an AMD rep in this forum repeating that IPC will be higher (maybe meaning the module), while the uarch already showed narrower cores. On 09/04/29 I already cited "2 ALUs and 2 AGUs" per core from the patents. I think these "official" claims by JF where misleading, causing a cognitive dissonance which led to erratic behavior of forum members.

Ok, so you mean, it's comparable. Based on what?


This was the base for applying Gauss Copula models to the financial markets leading to the subprime mortgage crisis.

But as NTBMK said, it's good to have some salt in a discussion without new data points regarding clock frequency and power.

Only comparable part is reinforcing that it's silly to get invested in an unreleased product, doesn't matter whose brand it is.

I thought AMD's Bulldozer was ambitious and they'd need to have a design that was 30-40% (before CMT penalty) better than Phenom II, core to core at same power draw, in order to be a clear improvement from the Phenom II six core CPUs and create a viable product (6-8 core Nehalem like SKUs vs 2-4 core Sandybridge). I was concerned about their choice to design for higher clocks rather than beefier cores but let JFAMD's statements sway me enough to at least get a Bulldozer capable motherboard when I purchased a Phenom II x6. Shame the only compelling upgrade when I hit my refresh/rebuild window was to an IvyBridge i5. I was basically out $20-30 for thinking maybe Bulldozer would be OK. Even someone who bought a full Phenom II system from scratch or a new motherboard for an existing system for the purpose of dropping in a Bulldozer on launch was probably out of pocket ~$50 + 1-2 hours time to sell. With Zen there isn't even that level of pre-launch investment available.

 

[NTMBK]

I was basically out $20-30 for thinking maybe Bulldozer would be OK. Even someone who bought a full Phenom II system from scratch for the purpose of dropping in a Bulldozer on launch was probably out of pocket ~$50 + 1-2 hours time to sell off the motherboard and CPU and move to Sandybridge. With Zen there isn't even that level of pre-launch investment available.

Only until Bristol Ridge launches; then we can start buying "Zen capable" AM4 motherboards, and start the whole cycle again. :thumbsup:

 

[Vesku]

Only until Bristol Ridge launches; then we can start buying "Zen capable" AM4 motherboards, and start the whole cycle again. :thumbsup:

Will those be marketed as Zen ready? At least that still removes the motherboard only "upgrade ready" siren call. Also AMD employees are yet to publicly man any hype trains.

 

[ShintaiDK]

One can only hope the people fooled last time learned to wait this time around. Selling a rebranded Carrizo first on AM4 does sound questionable in that regard and a rerun of the rebranded chipsets sold as new Bulldozer ready boards. Get people locked in and hope they stay.

 

[SK10H]

Oh, my ignore list getting longer. Feel sorry about this forum. I have foresaw this thread would be screwed up because it's some good news related to AMD. I wouldn't reply until some valuable posts are here.

Me - Not enough post to have ignore list feature, wasting my bandwidth. Too many trolling on a microarchitecture thread. :thumbsup:

 

[guskline]

Dresdenboy Thank you!

 

[Azuma Hazuki]

People are ignoring something here: in a market society competition is essential. Try and imagine where we'd be if Intel were well and truly unopposed--and if you think the US government would actually follow through with their threats of breaking Intel up you're nuts.

Regarding AMD, I pretty much am an AMD fangirl but am also realistic about it. As things stand, Intel has an overall superior lineup at the top, wins most of the mid range (i5-i7), and ironically is a rather poor value under $120 or so (but that's not where the hype is). My motto is "use the right tool for the right job." At this point that's almost always Intel for anything more than basic home use or very light gaming; Zen may shift this equilibrium somewhat.

Wait For The Benchmarks (TM).

 

[DrMrLordX]

Thank you for the updated diagram. So . . . is anyone worried that AMD may be limiting Zen's performance in certain applications by continuing to use 128-bit FMACs?

Is Zen going to have to split 256-bit AVX/AVX2 instructions in hardware like XV?

 

[William Gaatjes]

Thank you Dresdenboy for making the cpu subforum interesting again. The endless Intel vs AMD rivalry is incredibly boring. I hope David Kanter from realworldtech will come up with an extensive in depth article too. Keep up the good work. In depth information is more interesting.

 

[LTC8K6]

I'm an Intel fan from way back. I do build AMD systems and don't really have any complaints about the AMD systems I have built.

I'm rooting for Zen primarily because I want AMD to give Intel a push.

 

[ThatBuzzkiller]

I didn't see much what this "atomic" refers to. But according to patents, the checkpointing capability is thought as a mechanism for implementing transactional memory.

Show me the patent ...

Atomic operations are just synchronization primitives meant to prevent race conditions from happening ...

 

[sm625]

People are ignoring something here: in a market society competition is essential. Try and imagine where we'd be if Intel were well and truly unopposed--

I know what would happen. Intel would release a new chip called the F35, and sell it for $400 billion.

 

[raghu78]

One can only hope the people fooled last time learned to wait this time around. Selling a rebranded Carrizo first on AM4 does sound questionable in that regard and a rerun of the rebranded chipsets sold as new Bulldozer ready boards. Get people locked in and hope they stay.

Other than saying Zen will suck and Intel will rule what do you in any AMD CPU thread. Get a life.