New Zen microarchitecture details

[Eric1987]

I'd love to upgrade to a good AMD chip. I hope they get good again.

 

[Dresdenboy]

Not for the Integer,

BD Module has 4x Integer ALUs + 4x AGUs

ZEN Core has 4x Integer ALUs + 2x AGUs.

IPC of a single thread fluctuates a lot. So it's good to be able to share those ALUs between two threads depending on the individual thread's needs.

The fixed 2x (2 ALU + 2 AGU) CMT configuration aimed at high clocks (helped by the less complex scheduler etc.) cut off all the cases with 3 or 4 required ALUs per cycle. But during a high IPC phase (e.g. a loop) a steady IPC of 2 is much lower than a possibly achievable IPC of 3 or maybe even 3.5.

 

[itsmydamnation]

Not for the Integer,

BD Module has 4x Integer ALUs + 4x AGUs

ZEN Core has 4x Integer ALUs + 2x AGUs.

AGU's aren't integer execution resources, they are for moving data to and from registers. Its actually interesting and something that internet strong man will never accept, but i see SMT as a way of maximizing value from the cost of AGU's and the L/S system more then anything else (hear the cries of balance).

Executing data is easy and cheap powerwise, moving data and the infrastructure to move data is hard and expensive. When you look at CON cores from this perceptive its another nail in the proverbial coffin, they have a large amount of expensive, power hungry resources and yet are so bottleneck by simple arithmetic operations they had to waste time executing some of these on the expensive AGU rather then leaving them to the ALU's.

Yes.

u'll have to wait until i have time to go searching, im sure it was in video, which is time consuming to listen to, i dont have that time right now.

 

[NostaSeronx]

Well, it would appear GlobalFoundries has ditched FX14/14LPP.

The new big hype FinFETs from GlobalFoundries is labelled GF14HP. What does it mean for AMD with Zen/Polaris!

 

[itsmydamnation]

nothing because you are on crack

 

[Azuma Hazuki]

Rude!

But seriously, Nosta, 2/3 of the stuff you say never materializes. Are those Harvester and Crane dice still in the works, or...?

 

[Arachnotronic]

Rude!

But seriously, Nosta, 2/3 of the stuff you say never materializes. Are those Harvester and Crane dice still in the works, or...?

Azuma, NostaSeronx has been posting on various forums for years with fantastical claims like this. This user acts as though he/she has "special" knowledge but in reality the signal to noise ratio is not good.

NostaSeronx's posts are generally not grounded in reality and should not be relied upon for any serious analysis/decision making. Entertainment purposes only, but at this point it's not even funny, it's just sad.

 

[VirtualLarry]

NostaSeronx's posts are generally not grounded in reality and should not be relied upon for any serious analysis/decision making. Entertainment purposes only, but at this point it's not even funny, it's just sad.

That could be said of certain other forum members too, but I'm not going to single them out.

 

[NostaSeronx]

Nosta, 2/3 of the stuff you say never materializes.

The GF14HP node was apparently not marketed as much as 22FDX.

Blips of GlobalFoundries' 14nm HP FinFET that is no way related to Samsung's 14LPE/14LPP from linkedin;

14HP Design Rule Development Intern,
-Summer Intern-14HP Design Rule Development at GLOBALFOUNDRIES, U.S
-Support of 14nm High Performance Design Rule Development

IBM assignee at GlobalFoundries, working on 14HP BEOL integration
working on 14HP BEOL integration to support 14nm chip production for IBM Server Group

14HP Lead Manufacturing Integrator, Responsible for coordinating manufacturing and volume production readiness activities for 14nm SOI technologies.

CMOS Device Modeling Engineer,
Delivered FEOL, MOL and BEOL passive device compact models in world-class PDK for multiple generations of SOI and bulk CMOS technology nodes: 32SOI, 20LPE, 22SOI, 14SOI/14HP, 10nm and 7nm.

Fellow, Advanced Technology Development, Advanced Technology Development - Integration and Device Design Group at GLOBALFOUNDRIES.
Responsible for 14HP, 10nm and 7nm Technology Development

One of the big things to note is that ALL three 300mm foundries from GlobalFoundries will be doing GF22FDX and GF14HP. Some profiles doing 22FDX EDA-related work are now recently doing 14HP work. The ditch of FX14/14LPP basically means only Samsung foundries will be doing Samsung nodes.

Are those Harvester and Crane dice still in the works, or...?

Maybe. 14XM Bulldozer added a lot of things to 15h.

Floating Point; FP128+FP256 (FN8000_001A_EAX // FP256_V=1 // FP128_V=1), 2 FP256 to 4 FP128, the units are more uniform implying MMX being fused into FMAC units. Which was seen with Steamroller/Excavator w/ P0 and P2 being fused from Bulldozer/Piledriver.
Gen. Execution; AGUs have adders for ADD, SUB, Convert, Shift, etc. // ALUs are uniform; 2 imuls, 2 idivs, 2 branch, 2 jumps
Front-end and cache; L2 is Jag/Leo/Mar derivived replacing the larger, more slow L2 design. That has been seen in every 15h design to date. Much faster usage latency, same read/write latency as Steamroller/Excavator though. L1D is implied to have three ports; Read/Write for port 0, Write for port 1, Load for port 2. 1RW1W1R(4 reads + 4 writes) instead of 2R1W(4 reads + 2 writes). Branch predictor and L1i+Fetch split for more single threaded performance. Saw the less complete split with Steamroller/Excavator.
Misc; AVFS which was added to Excavator, AFS which was added to Steamroller, and IVR which hasn't popped up yet.

 

[AtenRa]

I seems GF14HP is an internal name for Samsungs 14LPP

 

[Arachnotronic]

I seems GF14HP is an internal name for Samsungs 14LPP

14HP is most likely IBM's 14nm high performance node to be used to build POWER9. In fact, it's literally right there in Nosta's quotes.

 

[Silverforce11]

Indeed, very interesting. We're fixated on TSMC, GF/Samsung and forgotten about IBM. Hmm.

 

[CHADBOGA]

1. These "misdeeds" intentionally caused unrecoverable damage to x86.

Where is the proof of this? D:

 

[swilli89]

Where is the proof of this? D:

What? http://www.cnet.com/news/intel-to-pay-amd-1-25-billion-in-antitrust-settlement/

AMD produced better silicon during the Pentium 4 days so Intel used its massive amounts of cash to pay off OEM's to continue to use their inferior hardware.

Please disregard if you were joking.

 

[DrMrLordX]

Indeed, very interesting. We're fixated on TSMC, GF/Samsung and forgotten about IBM. Hmm.

GF now owns all of IBM's old fabs and has many of their former engineers on-staff. So IBM's process R&D is now GF's.

 

[Vortex6700]

Where is the proof of this? D:

What do you need proof of? The "misdeeds" are common knowledge. Heck, they've been found guilty of anticompetition in almost every western nation (and a few easterns).

As for the irreparable damage, how many x86 manufacturers do you see nowadays?

 

[CHADBOGA]

What? http://www.cnet.com/news/intel-to-pay-amd-1-25-billion-in-antitrust-settlement/

AMD produced better silicon during the Pentium 4 days so Intel used its massive amounts of cash to pay off OEM's to continue to use their inferior hardware.

Please disregard if you were joking.

I'm asking about the irreparable harm.

What do you need proof of? The "misdeeds" are common knowledge. Heck, they've been found guilty of anticompetition in almost every western nation (and a few easterns).

As for the irreparable damage, how many x86 manufacturers do you see nowadays?

Wow, what a weak proof for your argument.

Back during the time just before Intel was playing funny games, how many x86 manufacturers were there?

Same as today, Intel, AMD, VIA.

People keep wanting to blame Intel for AMD's incompetence, yet forget that AMD were capacity constrained during their x64 days, so they sold every processor they could make, yet some continue with the fantasy that they could have sold so much more.

 

[coercitiv]

Sigh, even the slightest mention is enough to stir up the hornet nest. My apologies, will not make this mistake again.

 

[AtenRa]

14HP is most likely IBM's 14nm high performance node to be used to build POWER9. In fact, it's literally right there in Nosta's quotes.

Forgot about that.

 

[Xpage]

The GF14HP node was apparently not marketed as much as 22FDX.

Blips of GlobalFoundries' 14nm HP FinFET that is no way related to Samsung's 14LPE/14LPP from linkedin;One of the big things to note is that ALL three 300mm foundries from GlobalFoundries will be doing GF22FDX and GF14HP. Some profiles doing 22FDX EDA-related work are now recently doing 14HP work. The ditch of FX14/14LPP basically means only Samsung foundries will be doing Samsung nodes.Maybe. 14XM Bulldozer added a lot of things to 15h.

Floating Point; FP128+FP256 (FN8000_001A_EAX // FP256_V=1 // FP128_V=1), 2 FP256 to 4 FP128, the units are more uniform implying MMX being fused into FMAC units. Which was seen with Steamroller/Excavator w/ P0 and P2 being fused from Bulldozer/Piledriver.
Gen. Execution; AGUs have adders for ADD, SUB, Convert, Shift, etc. // ALUs are uniform; 2 imuls, 2 idivs, 2 branch, 2 jumps
Front-end and cache; L2 is Jag/Leo/Mar derivived replacing the larger, more slow L2 design. That has been seen in every 15h design to date. Much faster usage latency, same read/write latency as Steamroller/Excavator though. L1D is implied to have three ports; Read/Write for port 0, Write for port 1, Load for port 2. 1RW1W1R(4 reads + 4 writes) instead of 2R1W(4 reads + 2 writes). Branch predictor and L1i+Fetch split for more single threaded performance. Saw the less complete split with Steamroller/Excavator.
Misc; AVFS which was added to Excavator, AFS which was added to Steamroller, and IVR which hasn't popped up yet.

So any juicy 14nm SOI coming?

 

[el etro]

14HP is most likely IBM's 14nm high performance node to be used to build POWER9. In fact, it's literally right there in Nosta's quotes.

This.

 

[Ma_Deuce]

People keep wanting to blame Intel for AMD's incompetence

Can you quote some of those "people" please. All I have seen Intel blamed for was illegal business practices.

 

[exar333]

I'd love to upgrade to a good AMD chip. I hope they get good again.

Same here.

100% Intel in my machines since Core 2 launched a decade ago. AMD hasn't had anything compelling (to me) since AX2 days. Have built a few machines here and there with X3 or X4s for budget builds, a while back, but its been a LONG time.

Would love to mess around with a new (competitive) AMD chip.

 

[myocardia]

Can you quote some of those "people" please. All I have seen Intel blamed for was illegal business practices.

1. These "misdeeds" intentionally caused unrecoverable damage to x86.

2. These "misdeeds" historically happen when a company releases something competitive with intel (386, athlon)

3. These "misdeeds" make the arguement of a competitive product moot.

How exactly does it happen to be Intel's fault that during the entire, multi-year run of the 130nm and 90nm 'rule' of AMD's superior-performing Athlon 64 and Athlon 64 X2s, that AMD was 100% supply-constrained? AMD sold every single CPU they were able to make, and at a very nice price, I might add. As a matter of fact, they did so up until the day that the Intel Conroe uArch launched.

 

[Vortex6700]

I'm asking about the irreparable harm.



Wow, what a weak proof for your argument.

Back during the time just before Intel was playing funny games, how many x86 manufacturers were there?

Same as today, Intel, AMD, VIA.

People keep wanting to blame Intel for AMD's incompetence, yet forget that AMD were capacity constrained during their x64 days, so they sold every processor they could make, yet some continue with the fantasy that they could have sold so much more.

You are out of your depth if you think the 2000's counts as "back during the time before intel was playing funny games."

In fact, you are off by decades, but you're probably too young to remember.

I never said intel was entirely to blame for anything, but they do play their part. I don't really like AMD, so feel free to lash out defensively some more.