New Zen microarchitecture details

Page 132 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Nothingness

Platinum Member
Jul 3, 2013
2,371
713
136
It's part of this article (with paywall). Edit2: Full article
You should link that on RWT, the poor David Kanter couldn't access Wilco reference :)

BTW that article you cite is wrong on libquantum speedup: even if they are right that most of the time is spent in a single loop, they make it sound as if optimizing that loop alone led to the 18x speedup. In fact it's the automagic (pun intended) Array of Structs -> Structs of Array conversion that enables that speedup (cf slides 25 ff http://llvm.org/devmtg/2015-10/slides/Gerolf-PerformanceImprovementsAndHeadroom.pdf).
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
So...GlobalFoundries was taken over by the crazy IBM foundry people. Which in part has effectively lead to a switch, while the management says "dual roadmap." There is really only the FDX roadmap being supported from now on till Mubadala cracks down. In which, it has been planned ahead to persuade them to only support the FDX roadmap. 14nm FinFET is apparently being decided to be dropped as the Samsung & GlobalFoundries relationship sours. There is also rumors of a fast port of 22FDX from Immersion Litho(2H 2017) to EUV Litho(1H 2018). Lower process steps, fits in single exposure, all those 2D metal layers, can be used with NXE:3300B 2nd-phase, etc.

What this means for AMD?
- Return of Bulldozer-likeness with Partitionable Clustered Multithreading(pCMT).
- 2 GenPurpose Cores & 2 FP-Int cores, Both FP cores fuse to do AVX512.
- Some of Zen's LSU IP will be ported.
- Most likely will launch in 2017? as -L series "Great Performance at Unbelievable Value" market segment. (@Cannonlake)

Don't mind me just being a Seronx.
 
Last edited:

superstition

Platinum Member
Feb 2, 2008
2,219
221
101
So...GlobalFoundries was taken over by the crazy IBM foundry people. Which in part has effectively lead to a switch, while the management says "dual roadmap." There is really only the FDX roadmap being supported from now on till Mubadala cracks down. In which, it has been planned ahead to persuade them to only support the FDX roadmap. 14nm FinFET is apparently being decided to be dropped as the Samsung & GlobalFoundries relationship sours. There is also rumors of a fast port of 22FDX from Immersion Litho(2H 2017) to EUV Litho(1H 2018). Lower process steps, fits in single exposure, all those 2D metal layers, can be used with NXE:3300B 2nd-phase, etc.

What this means for AMD?
- Return of Bulldozer-likeness with Partitionable Clustered Multithreading(pCMT).
- 2 GenPurpose Cores & 2 FP-Int cores, Both FP cores fuse to do AVX512.
- Some of Zen's LSU IP will be ported.
- Most likely will launch in 2017? as -L series "Great Performance at Unbelievable Value" market segment. (@Cannonlake)

Don't mind me just being a Seronx.
How about this new CMT CPU having an embedded AI processor, taking advantage of some of that IBM AI research?

It seems like that would be more useful than a banal iGPU.

Even better would be a partial indium arsenide construction of that construction core. : ) Are the nanowires ready yet?
 

KTE

Senior member
May 26, 2016
478
130
76
Wilco at RWT forum posted an interesting link to a book covering some of the FO4 stuff - and provides a FO4 per cycle chart of lots of CPUs till 2011 (p. 105):
https://books.google.co.uk/books?id=PiWOAwAAQBAJ&pg=PA104&dq=Pentium+FO4+inverter+delays+per+pipeline+stage&hl=en&sa=X&ved=0ahUKEwitzKjppt_PAhUB1xoKHXJpBJkQ6AEILDAC#v=onepage&q=Pentium FO4 inverter delays per pipeline stage&f=false

The book is titled "Top-Down Digital VLSI Design: From Architectures to Gate-Level Circuits and FPGAs" by Hubert Kaeslin.

If you don't get the linked pages, a fresh search might bring you there. The whole book is rather interesting.

EDIT: The data used in the book was based on the CPU DB, which has been used by other authors, too. Here is a not so colorful diagram without distinction of the different companies as in the book:
f14.jpg

It's part of this article (with paywall). Edit2: Full article
I will get to this, thanks for providing it... Was skimming it but then get rushed away :)

Just super busy with work/fam.



Sent from HTC 10
(Opinions are own)
 

AMD Polaris

Junior Member
Jul 17, 2016
5
9
36
Hi all! I went into the forest recently and birds have tweeted something again. They spoke about 2 new SKUs that have just hit the post-boxes of the mainboard manufacturers. They are both A0 revisions and Engineering Samples, so there is no improvement since my last post.

The first one is an 8-core design with AMD's HT implementation and it's got a 3150 MHz base clock, it's all-core turbo is 3300 MHz and the max turbo for 1 core is 3600 Mhz. Yes, here are some improvements regarding the previous 8-core SKU under the same TDP envelope.

The second SKU is a 4-core one with AMD's HT. It's got a 65W TDP and the base clock is still 2900 MHz. All-core turbo is 3100 MHz, max turbo is 3400 MHz. I don't know if it's only an SKU for testing mainboards or something is not okay with the clock-wattage correlation. I mean on higher clocks the 4-core SKU steps into the 95W TDP envelope, AMD can't keep the wattage low. Maybe GloFo's 14nm process needs some maturing... Frankly I don't have a clue what's in the background.

Retail AM4 mainboards are under production. The whole platform will be ready to have a paper-launch at the end of the year with a real availability in February of 2017. Performance wise the Zen uarch will be around Haswell and Broadwell (except for FMA), it seems it won't catch Skylake clock for clock. It's not a big deal, but if the clocks can't go higher until the start it won't fulfill the expectations. And we all know that expectations in this case are very high...
 
Last edited:

dark zero

Platinum Member
Jun 2, 2015
2,655
138
106
So...GlobalFoundries was taken over by the crazy IBM foundry people. Which in part has effectively lead to a switch, while the management says "dual roadmap." There is really only the FDX roadmap being supported from now on till Mubadala cracks down. In which, it has been planned ahead to persuade them to only support the FDX roadmap. 14nm FinFET is apparently being decided to be dropped as the Samsung & GlobalFoundries relationship sours. There is also rumors of a fast port of 22FDX from Immersion Litho(2H 2017) to EUV Litho(1H 2018). Lower process steps, fits in single exposure, all those 2D metal layers, can be used with NXE:3300B 2nd-phase, etc.

What this means for AMD?
- Return of Bulldozer-likeness with Partitionable Clustered Multithreading(pCMT).
- 2 GenPurpose Cores & 2 FP-Int cores, Both FP cores fuse to do AVX512.
- Some of Zen's LSU IP will be ported.
- Most likely will launch in 2017? as -L series "Great Performance at Unbelievable Value" market segment. (@Cannonlake)

Don't mind me just being a Seronx.
But isn't Zen a fusion of Bulldozer with Intel's SMT?
Also, why maintain Bulldozer unless is going to ARM?
Or you might meaning that Bulldozer will be on K12 now?

And expecting not to see Bulldozer dead on the lower segment at least. Is the only tier that they can trash Celeron and Pentium without problem.
 

deasd

Senior member
Dec 31, 2013
512
724
136
Hi all! I went into the forest recently and birds have tweeted something again. They spoke about 2 new SKUs that have just hit the post-boxes of the mainboard manufacturers. They are both A0 revisions and Engineering Samples, so there is no improvement since my last post.

The first one is an 8-core design with AMD's HT implementation and it's got a 3150 MHz base clock, it's all-core turbo is 3300 MHz and the max turbo for 1 core is 3600 Mhz. Yes, here are some improvements regarding the previous 8-core SKU under the same TDP envelope.

The second SKU is a 4-core one with AMD's HT. It's got a 65W TDP and the base clock is still 2900 MHz. All-core turbo is 3100 MHz, max turbo is 3400 MHz. I don't know if it's only an SKU for testing mainboards or something is not okay with the clock-wattage correlation. I mean on higher clocks the 4-core SKU steps into the 95W TDP envelope, AMD can't keep the wattage low. Maybe GloFo's 14nm process needs some maturing... Frankly I don't have a clue what's in the background.

Retail AM4 mainboards are under production. The whole platform will be ready to have a paper-launch at the end of the year with a real availability in February of 2017. Performance wise the Zen uarch will be around Haswell and Broadwell (except for FMA), it seems it won't catch Skylake clock for clock. It's not a big deal, but if the clocks can't go higher until the start it won't fulfill the expectations. And we all know that expectations in this case are very high...

Thx for the info. This is the most useful info since.....when? I couldn't even remember, maybe several months ago. 3150MHZ based fits my expectation.
 
  • Like
Reactions: Dresdenboy

Glo.

Diamond Member
Apr 25, 2015
5,657
4,409
136
More importantly, clock for clock Haswell/Broadwell performance for 8 cores in 95W thermal envelope. That can be possibly best news right now.
 

Thunder 57

Platinum Member
Aug 19, 2007
2,647
3,706
136
Thx for the info. This is the most useful info since.....when? I couldn't even remember, maybe several months ago. 3150MHZ based fits my expectation.

He/She was pretty spot on before, so I would think the information is reliable.

The octa-core sounds like it could be a real winner seeing AMD's current position. If it really is in the Haswell/Broadwell IPC range, 8 cores at those frequencies and 95W TDP is not bad at all. If they could get some hex cores in the mid 3GHz range, turbo up to 4Ghz if possible, that would likely be a real winner. The mainstream market has been at quad core for some time now. If AMD follows up well as the process matures and get Zen APU's and Zen+ out on time they may finally gain some traction.
 

lolfail9001

Golden Member
Sep 9, 2016
1,056
353
96
Dunno guys. With what i hear it means to me this thing is less power efficient than Skylake.

So, it's an okay HEDT chip, but something tells me AMD are not going to recapture any visible portion of server market.
 
  • Like
Reactions: Sweepr

PhonakV30

Senior member
Oct 26, 2009
987
378
136
Dunno guys. With what i hear it means to me this thing is less power efficient than Skylake.

So, it's an okay HEDT chip, but something tells me AMD are not going to recapture any visible portion of server market.

Because this , out of AMD's hand
The second SKU is a 4-core one with AMD's HT. It's got a 65W TDP and the base clock is still 2900 MHz. All-core turbo is 3100 MHz, max turbo is 3400 MHz. I don't know if it's only an SKU for testing mainboards or something is not okay with the clock-wattage correlation. I mean on higher clocks the 4-core SKU steps into the 95W TDP envelope, AMD can't keep the wattage low. Maybe GloFo's 14nm process needs some maturing... Frankly I don't have a clue what's in the background.

Also if GF gets matured 14nm Process , then there is a hope that we will see Polaris's new revision with low Wattage for mobile market And CPU too.
 

Vesku

Diamond Member
Aug 25, 2005
3,743
28
86
8 core 16 thread with Haswell-ish IPC would be "good enough" for me to start reincluding AMD CPUs in my builds given reasonable pricing. Will be interesting to see OC headroom going above the 95W TDP.
 

krumme

Diamond Member
Oct 9, 2009
5,952
1,585
136
8 core 16 thread with Haswell-ish IPC would be "good enough" for me to start reincluding AMD CPUs in my builds given reasonable pricing. Will be interesting to see OC headroom going above the 95W TDP.
Lets get back on earth.
A broadwell e 6950 8 core is 3 at base all core and 3.5 all core turbo single 4 turbo.
I dont know what kind of ibm cpu you boot at home since its "good enough" but 3.15 base at hsw like ipc is extremely high in my book since its more or less like the top end bwe for perfectly threadded loads. In a 95w tdp but sans heavy 256b fp loads.
I will see that before i beliewe it.
 

lolfail9001

Golden Member
Sep 9, 2016
1,056
353
96
What is that based on? These ES surely are not meant to be some "E" edition. So if true, this would be 3.3 8C boost @95W vs. 4.2 4C boost @91W.
Look: plain 6700 has 4Ghz single core boost, 3.6Ghz all core boost and 3.4Ghz base clock and iGPU @ 65W. Compared to 3.1Ghz all core turbo, 3.4Ghz single core turbo without iGPU in 65W. Uncore is fairly similar as well.
 

Glo.

Diamond Member
Apr 25, 2015
5,657
4,409
136
Look: plain 6700 has 4Ghz single core boost, 3.6Ghz all core boost and 3.4Ghz base clock and iGPU @ 65W. Compared to 3.1Ghz all core turbo, 3.4Ghz single core turbo without iGPU in 65W. Uncore is fairly similar as well.
It is not finished product. A0 still.
 

Azix

Golden Member
Apr 18, 2014
1,438
67
91
Hi all! I went into the forest recently and birds have tweeted something again. They spoke about 2 new SKUs that have just hit the post-boxes of the mainboard manufacturers. They are both A0 revisions and Engineering Samples, so there is no improvement since my last post.

The first one is an 8-core design with AMD's HT implementation and it's got a 3150 MHz base clock, it's all-core turbo is 3300 MHz and the max turbo for 1 core is 3600 Mhz. Yes, here are some improvements regarding the previous 8-core SKU under the same TDP envelope.

The second SKU is a 4-core one with AMD's HT. It's got a 65W TDP and the base clock is still 2900 MHz. All-core turbo is 3100 MHz, max turbo is 3400 MHz. I don't know if it's only an SKU for testing mainboards or something is not okay with the clock-wattage correlation. I mean on higher clocks the 4-core SKU steps into the 95W TDP envelope, AMD can't keep the wattage low. Maybe GloFo's 14nm process needs some maturing... Frankly I don't have a clue what's in the background.

Retail AM4 mainboards are under production. The whole platform will be ready to have a paper-launch at the end of the year with a real availability in February of 2017. Performance wise the Zen uarch will be around Haswell and Broadwell (except for FMA), it seems it won't catch Skylake clock for clock. It's not a big deal, but if the clocks can't go higher until the start it won't fulfill the expectations. And we all know that expectations in this case are very high...

where does skylake stand with the processor they compared zen with at the same clock speed recently? That one has a lower per clock perf compared to skylake?

Lets get back on earth.
A broadwell e 6950 8 core is 3 at base all core and 3.5 all core turbo single 4 turbo.
I dont know what kind of ibm cpu you boot at home since its "good enough" but 3.15 base at hsw like ipc is extremely high in my book since its more or less like the top end bwe for perfectly threadded loads. In a 95w tdp but sans heavy 256b fp loads.
I will see that before i beliewe it.

it is interesting huh. compared to clock speeds of regular quad i7s you might think meh, but then compared to intels $1000 chips, it looks pretty good. A good price would probably kill the market for those high core count intel processors.
 
Last edited:

itsmydamnation

Platinum Member
Feb 6, 2011
2,743
3,074
136
Dunno guys. With what i hear it means to me this thing is less power efficient than Skylake.

So, it's an okay HEDT chip, but something tells me AMD are not going to recapture any visible portion of server market.
Based on what exactly?
More cores per socket, more memory per socket, ball park power consumption, ball park ipc, ball park clocks.

Just to given an example current client I work at is running 2x8 core Xeon's with 384gb memory for VM farm SKU and we are still running out of memory.
Zen will be able to pack in more memory without going to larger dim's which can be a significant cost reduction in hardware ( especially the 32-64gb dimm transition).
Also remember that ZEN has a full NB/SB integrated with multiple 10gb NIC's so Zen has some extra base load as well in that TDP.

Most people fail to understand what matters in the server market. AMD's hardest market to compete in will be against high clock 4 cores in the consumer market. HEDT AMD looks not to bad and unless intel want to massively reduce prices AMD can find sweet spot, remember Zen looks to be smaller then Polaris (anywhere from 160mmsq to 220 mm sq is a possibility right now) and in laptop space Zen APU should do much better then CON core and with the superior GPU hopefully will be able to get good gains in market share.

Hopefully between now and release AMD can get the base clock upto 3.5ghz. that would be a 10% clock improvement from what AMD Polaris listed hopefully can push boost 1 core boost 10% higher as well as that would be 4ghz...... time will tell. Hopefully the 10gb NIC wont be disabled in the consumer "HEDT" variant as that could be a pretty nice bonus.
 
Last edited:
  • Like
Reactions: .vodka

lolfail9001

Golden Member
Sep 9, 2016
1,056
353
96
It is not finished product. A0 still.
True, but we do hear that it is *process*-related.

Based on what exactly?
More cores per socket, more memory per socket, ball park power consumption, ball park ipc, ball park clocks.
Power efficiency and only that.

Zen will be able to pack in more memory without going to larger dim's which can be a significant cost reduction in hardware ( especially the 32-64gb dimm transition).
What do you mean? It's not that easy to pack in 12 DIMMs per socket as is.

Most people fail to understand what matters in the server market.
Perf/cost of ownership, nothing else, last time i checked.
 

Dresdenboy

Golden Member
Jul 28, 2003
1,730
554
136
citavia.blog.de
Look: plain 6700 has 4Ghz single core boost, 3.6Ghz all core boost and 3.4Ghz base clock and iGPU @ 65W. Compared to 3.1Ghz all core turbo, 3.4Ghz single core turbo without iGPU in 65W. Uncore is fairly similar as well.
I see, you're complaining about the 4C part. If this is not just an ES for fun (bug hunting, sys development, testing, etc.), but will turn into a real product, then it looks to be a harvested product (there are only physical 8C dies), which exists based on its bad characteristics in the first place.