• Guest, The rules for the P & N subforum have been updated to prohibit "ad hominem" or personal attacks against other posters. See the full details in the post "Politics and News Rules & Guidelines."

New Zen microarchitecture details

Page 68 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Ancalagon44

Diamond Member
Feb 17, 2010
3,275
202
106
40% IPC increase is actually still a vague statement, because IPC depends upon the current workload being executed.

In other words, for the 40% IPC increase statement to be true, all that AMD needs to do is find a program that exhibits 40% increased IPC than the previous design. They could make it a completely contrived benchmark if they wanted to.
 

Vesku

Diamond Member
Aug 25, 2005
3,745
27
86
Hmm, seems like the 40% improvement claim of Lisa Su's presentation was versus Vishera FX series?
 

R0H1T

Platinum Member
Jan 12, 2013
2,566
142
106
Hmm, seems like the 40% improvement claim of Lisa Su's presentation was versus Vishera FX series?
The final Zen performance still depends on the clocks though. Intel for the second time in a row have gone on to max out their clocks on the same node, with DC & then Kabylake, so realistically Zen's second iteration or revision is where the optimal clocks may land with the subsequent version possibly maxing the clocks out. In other words we'll have to wait a bit, likely a year or so, to see the true potential of Zen & 14nm LPP wrt clocks & eventually performance relative to Intel.
 
Last edited:

Ancalagon44

Diamond Member
Feb 17, 2010
3,275
202
106
I reckon they will deliver the 40% more IPC as promised, in the majority of workloads, but at a cost of a 20% decrease in clocks.

I'm curious about the 8 core/16 thread chip though. It is quite a behemoth to sell to consumers. FX chip maybe?
 

SpaceBeer

Senior member
Apr 2, 2016
305
98
101
40% IPC increase is actually still a vague statement, because IPC depends upon the current workload being executed.

In other words, for the 40% IPC increase statement to be true, all that AMD needs to do is find a program that exhibits 40% increased IPC than the previous design. They could make it a completely contrived benchmark if they wanted to.
Something like this?
http://images.anandtech.com/doci/9483/01 - Gains over Sandy.png

Or maybe to compare them in gaming :)
http://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/10

So it seems it's not so important if Zen has same IPC like Skylake or Haswell (or even IB). At least not important for vast majority of PC users. And I'm sure price will be adjusted according to performance
 

CHADBOGA

Platinum Member
Mar 31, 2009
2,026
624
136
The replies so far showcase exactly why this practice is commonplace as it's a psychological trick that works even if people know that it's happening and exactly why it's happening.
Might be time for you to put up a female in your avatar. :D
 

sm625

Diamond Member
May 6, 2011
8,176
135
106
The only reason companies put females in leadership positions is because their focus groups show that females will shield the company from criticism when it crashes and burns taking everything with it.

No one puts females in leadership positions in US companies for any reason other than that.

If you see a female being put in a leadership position in the US you know the company is in pure damage control mode.
Donald, is that you?
 
Mar 10, 2006
11,719
2,003
126
So it seems it's not so important if Zen has same IPC like Skylake or Haswell (or even IB). At least not important for vast majority of PC users. And I'm sure price will be adjusted according to performance
Price always follows performance. The more powerful the chip (in relative terms), the more valuable it is, and therefore more willing people are to buy it.

That's why Intel can charge $1750 for a 6950X and have people buy it. It is the highest-core count desktop CPU out there, it has the highest multithreading performance out of the box, and it can overclock to ~4GHz (if AdamK47's sample is representative).

Anyway, tying this back to Zen, I think Summit Ridge will be priced at a slight discount to its performance. Intel is a much better known brand than AMD, so AMD has to overcome the mindshare deficit by offering greater "bang for the buck" so to speak.

If they have something akin to an 8 core BDW-E in performance, I can't see them offering it for $399, but for, like, $899 -- this undercuts Intel. If they have something more like a 6 core Broadwell-E, then I'd expect them to charge $350.

That being said, the problem that AMD faces is that Intel's margins on these (relatively low volume) parts are so high compared to the rest of their business, they have a lot of flexibility to adjust pricing down to match. And, on top of that, Intel has the brand/mind-share advantage.
 

ZGR

Golden Member
Oct 26, 2012
1,853
267
126
I agree with ~$900 for the top of the line 8c/16t Zen. I expect AMD to try to undercut Intel at every category, but not by much.

If Intel does lower their CPU prices, I would be staggered. But what about X99 prices vs AM4? It could be interesting.
 

dark zero

Platinum Member
Jun 2, 2015
2,542
100
91
I feel that AMD octa won't go past 600 dollars... Intel has nothing to do to prevent the drop of prices.
 
Mar 10, 2006
11,719
2,003
126
If Intel does lower their CPU prices, I would be staggered. But what about X99 prices vs AM4? It could be interesting.
The board prices are dominated by costs that are out of Intel's control. Intel could reduce the prices of its X99 chipsets, but that'll only go so far.
 

NTMBK

Diamond Member
Nov 14, 2011
8,843
1,921
136
I agree with ~$900 for the top of the line 8c/16t Zen. I expect AMD to try to undercut Intel at every category, but not by much.

If Intel does lower their CPU prices, I would be staggered. But what about X99 prices vs AM4? It could be interesting.
X99 is a quad channel platform with tonnes of PCIe lanes, making for a more complex board with lots of traces. AM4 is (apparently) dual channel with fewer PCIe lanes, so boards should be simpler and cheaper.
 

tential

Diamond Member
May 13, 2008
7,363
641
121
If Zen is only 8 and 6 core I don't think AMD will cut out the whole mainstream. I'm going with $500-600 for the 8 core and $300-400 for the 6 core.
 

The Stilt

Golden Member
Dec 5, 2015
1,709
3,057
106
I feel that AMD octa won't go past 600 dollars... Intel has nothing to do to prevent the drop of prices.
Let's say the fastest Summit Ridge SKU operates at 3.0GHz base and 3.5GHz boost frequency, and that the IPC is exactly 40% higher than on Excavator in every workload. The cheapest Broadwell-E (6800K) SKU sells for 445€ incl. tax in Europe, and it has 3.8GHz boost and 3.4GHz base frequency.

In that example Broadwell-E will have both higher IPC and higher frequency on it's side. The single threaded performance in standard FP benchmarks would be 25 - 30% higher. Despite having two cores less, 6800K would still tie with 8C/16T Zeppelin, if the both Broadwell-E and Zeppelin would have identical SMT efficiency. AMD has few months of real-world ("on silicon") experience on SMT, while Intel has well over 15 years of experience on it. I'll raise my (virtual) hat to AMD if they can match, let alone excel the SMT efficiency of Intel. Either scenario might technically be possible, yet extremely unlikely.

In order to match even the entry level Broadwell-E SKU in ST workloads (e.g. Cinebench R15) Zeppelin should either have boost frequency of ~ 4.5GHz or IPC improvement of < 80% over Excavator, or preferably both. In some workloads (such as HEVC) even < 80% IPC improvement wouldn't be enough, since Broadwell-E can have <100% higher IPC than Excavator.

A vast improvement no doubt, but I don't think AMD can ask (and get) more than 350&#8364; for any consumer Zeppelin model. Zeppelin will also have certain restrictions, which the X99 HEDT plaform for example doesn't have.
 
Last edited:

ZGR

Golden Member
Oct 26, 2012
1,853
267
126
X99 is a quad channel platform with tonnes of PCIe lanes, making for a more complex board with lots of traces. AM4 is (apparently) dual channel with fewer PCIe lanes, so boards should be simpler and cheaper.
I guess I'm only thinking of the "budget" X99 boards vs high end AM4. I'm wondering where the overlap will be price wise with the amount of features each board will get. I am confident we will see some premium AM4 motherboards out there that are not cheap.
 

Doom2pro

Senior member
Apr 2, 2016
587
619
106
The only reason companies put females in leadership positions is because their focus groups show that females will shield the company from criticism when it crashes and burns taking everything with it.

No one puts females in leadership positions in US companies for any reason other than that.

If you see a female being put in a leadership position in the US you know the company is in pure damage control mode.
Trump, is that you?


PLEASE, Enough derailing the thread with responding to this post above. Please stay on topic. -Admin DrPizza
 
Last edited by a moderator:

KTE

Senior member
May 26, 2016
478
130
76
Let's say the fastest Summit Ridge SKU operates at 3.0GHz base and 3.5GHz boost frequency, and that the IPC is exactly 40% higher than on Excavator in every workload. The cheapest Broadwell-E (6800K) SKU sells for 445€ incl. tax in Europe, and it has 3.8GHz boost and 3.4GHz base frequency.

In that example Broadwell-E will have both higher IPC and higher frequency on it's side. The single threaded performance in standard FP benchmarks would be 25 - 30% higher. Despite having two cores less, 6800K would still tie with 8C/16T Zeppelin, if the both Broadwell-E and Zeppelin would have identical SMT efficiency. AMD has few months of real-world ("on silicon") experience on SMT, while Intel has well over 15 years of experience on it. I'll raise my (virtual) hat to AMD if they can match, let alone excel the SMT efficiency of Intel. Either scenario might technically be possible, yet extremely unlikely.

In order to match even the entry level Broadwell-E SKU in ST workloads (e.g. Cinebench R15) Zeppelin should either have boost frequency of ~ 4.5GHz or IPC improvement of < 80% over Excavator, or preferably both. In some workloads (such as HEVC) even < 80% IPC improvement wouldn't be enough, since Broadwell-E can have <100% higher IPC than Excavator.

A vast improvement no doubt, but I don't think AMD can ask (and get) more than 350€ for any consumer Zeppelin model. Zeppelin will also have certain restrictions, which the X99 HEDT plaform for example doesn't have.
And you know very well that all this is well beyond the realms of possible ;)

25% IPC keeping similar frequencies AKA on average across the board ignoring outliers is alone a very high requirement.
 

krumme

Diamond Member
Oct 9, 2009
5,881
1,484
136
Let's say the fastest Summit Ridge SKU operates at 3.0GHz base and 3.5GHz boost frequency, and that the IPC is exactly 40% higher than on Excavator in every workload. The cheapest Broadwell-E (6800K) SKU sells for 445€ incl. tax in Europe, and it has 3.8GHz boost and 3.4GHz base frequency.

In that example Broadwell-E will have both higher IPC and higher frequency on it's side. The single threaded performance in standard FP benchmarks would be 25 - 30% higher. Despite having two cores less, 6800K would still tie with 8C/16T Zeppelin, if the both Broadwell-E and Zeppelin would have identical SMT efficiency. AMD has few months of real-world ("on silicon") experience on SMT, while Intel has well over 15 years of experience on it. I'll raise my (virtual) hat to AMD if they can match, let alone excel the SMT efficiency of Intel. Either scenario might technically be possible, yet extremely unlikely.

In order to match even the entry level Broadwell-E SKU in ST workloads (e.g. Cinebench R15) Zeppelin should either have boost frequency of ~ 4.5GHz or IPC improvement of < 80% over Excavator, or preferably both. In some workloads (such as HEVC) even < 80% IPC improvement wouldn't be enough, since Broadwell-E can have <100% higher IPC than Excavator.

A vast improvement no doubt, but I don't think AMD can ask (and get) more than 350€ for any consumer Zeppelin model. Zeppelin will also have certain restrictions, which the X99 HEDT plaform for example doesn't have.
Where did you pull those ipc numbers?

Some tractor fp test from the k7 days? :)
 

sm625

Diamond Member
May 6, 2011
8,176
135
106
I expect at least one 8C/16T Zen SKU to be priced under $300. It's not a large chip. It will surely be cheaper to produce than polaris 10, and look where that is being priced! AMD could sell 8C/16T Zen chips for $150 and actually make significantly more money per unit sold compared to polaris 10. If Zen actually comes in under 800mS on Kraken 1.1, then maybe the price might go to $400.
 

superstition

Platinum Member
Feb 2, 2008
2,219
216
101
US labeling laws are extremely strict.
Not really.

You'll find all sorts of food products that say "distributed by" — which is meaningless in terms of knowing whether or not it comes from China's tainted land and water, for instance. That's not strict at all. It's "anything goes".

Some artificial sweeteners like neotame don't have to be listed on a label. Aluminum, which is apparently a neurotoxin, is not required to be labeled on processed cheese (by far the biggest source of aluminum contamination in the body) — nor is such cheese banned. Toxins like caramel colorant and coal tar dyes are still being used in vast numbers of products, despite there being healthy and affordable alternatives. The FDA doesn't test protein drinks and powders despite CR and others finding high levels of heavy metals. The FDA allowed six or seven arsenic powders to be added to livestock feed (now just 2 since they public found out and they did a dog and pony show to convince people the arsenic poisoning is over).

Companies like Coca-Cola have been able to fast track products with highly questionable safety status (like stevia extract) by using the "GRAS" status. The FDA didn't even require lifetime testing in both rats and mice for rebaudioside-A, which is the most basic requirement for nutritional research. Steviol is a mutagen. Acesulfame-K has multiple cancer experts warning about carcinogenic potential and the poor quality of the existing research — and has had that for a very long time. I can go on and on. The last thing I'll say is that it's not that surprising that Romney's chief financial advisor made his billions pushing patent medicine (I meant to say supplements...) — a mostly (intentionally) unregulated business.

Speaking of labeling and Zen. I suggested that AMD have a tiered label, something along the lines of 80 Plus and the Nintendo quality seal — for motherboards, so consumers won't be duped by board makers into buying boards that don't have adequate power delivery for enthusiast-level parts. The Stilt said he made the same suggestion years back. Despite being a faulty standard, 80 Plus has had positive effects in the power supply market. AMD could do quite a bit for Zen by giving customers more reassurance after the AM3+ mess of low-grade VRM boards that blow up and take the chip with it.




The counterpoint is that Zen isn't going to be targeting the power draw of something like a 9590. That's fine but some of the FX boards were apparently so shoddily made that they had problems handling a stock 8350. AMD needs to keep the board makers from giving its CPUs a bad name and also prevent customers from having to overpay (for unnecessarily robust power delivery) if they specifically want a low-power CPU. The lack of clarity in the board market is a problem.
 
Last edited:

lixlax

Member
Nov 6, 2014
143
83
101
Lets have some betting here. Can some setup a combined freq and ipc test? Perhaps add smt efficiency of some sort and max consumer price?

I guess today for 8c part:
Hw ipc class
2.5 GHz boost 3.2
Smt sucks efficiency class
450 euro tops
At these clocks it would be hardly faster than FX8350/70, maybe 10-20%.
Although these are believed to have 95W TDP, while not too much higher clocked Intel's 6-10 core CPUs are rated at 140W. I think anything less than 3GHz for base clock will be a complete failure (at least for the top models).
 

Dresdenboy

Golden Member
Jul 28, 2003
1,730
554
136
citavia.blog.de
Lets have some betting here. Can some setup a combined freq and ipc test? Perhaps add smt efficiency of some sort and max consumer price?

I guess today for 8c part:
Hw ipc class
2.5 GHz boost 3.2
Smt sucks efficiency class
450 euro tops
Is it poll time again? My take (what's at least somewhat likely):
  • HSW IPC for integer, Ivy IPC for FP
  • 3.0 GHz / 4.0 GHz :sneaky:
  • SMT is at HSW level (they're working on it internally since Thunderbird according to my sources!)
  • $999 for top model
 

inf64

Diamond Member
Mar 11, 2011
3,061
1,772
136
Is it poll time again? My take (what's at least somewhat likely):
  • HSW IPC for integer, Ivy IPC for FP
  • 3.0 GHz / 4.0 GHz :sneaky:
  • SMT is at HSW level (they're working on it internally since Thunderbird according to my sources!)
  • $999 for top model
My take:

- IPC between IB and Haswell.
- 3.2Ghz base / 3.7Ghz TB for low thread workloads (on top 8C/16T part)
- SMT at HSW level or better
- intel BDW hex core pricing (while performing a bit better)
 

ASK THE COMMUNITY