Tuesday, September 24th 2019

AMD Could Release Next Generation EPYC CPUs with Four-Way SMT

AMD has completed design phase of its "Zen 3" architecture and rumors are already appearing about its details. This time, Hardwareluxx has reported that AMD could bake a four-way simultaneous multithreading technology in its Zen 3 core to enable more performance and boost parallel processing power of its data center CPUs. Expected to arrive sometime in 2020, Zen 3 server CPUs, codenamed "MILAN", are expected to bring many architectural improvements and make use of TSMC's 7nm+ Extreme Ultra Violet lithography that brings as much as 20% increase in transistor density.

Perhaps the biggest change we could see is the addition of four-way SMT that should allow a CPU to have four virtual threads per core that will improve parallel processing power and enable data center users to run more virtual machines than ever before. Four-way SMT will theoretically boost performance by dividing micro-ops into four smaller groups so that each thread could execute part of the operation, thus making the execution time much shorter. This being only one application of four-way SMT, we can expect AMD to leverage this feature in a way that is most practical and brings the best performance possible.
AMD isn't the first to implement this kind of solution to its processors. IBM has been making CPUs based on POWER ISA for years now that feature four or even eight-way SMT and they are one of the key reasons why POWER CPUs are so powerful. Nonetheless, we can hope to see more details about Zen 3 core design decisions as we approach 2020 and launch of Milan CPUs. Source: Hardwareluxx
Add your own comment

159 Comments on AMD Could Release Next Generation EPYC CPUs with Four-Way SMT

#26
FordGT90Concept
"I go fast!1!11!1!"
This will be terrible for gaming. Server operators will rejoice at this news but...this is very likely to be harmful in the consumer space.
Posted on Reply
#27
cucker tarlson
FordGT90Concept, post: 4121752, member: 60463"
This will be terrible for gaming. Server operators will rejoice at this news but...this is very likely to be harmful in the consumer space.
epyc is not for gaming.
will likely never be introduced to gaming pc.
there's a way to fit 4 8 core cxx's onto a die.why go with 4 way smt ? that's only beneficial when you're maxing out cores on the die.
Posted on Reply
#28
FordGT90Concept
"I go fast!1!11!1!"
I highly doubt AMD is going to divorce the core design between Epyc and Ryzen. If they are, it's fine; if not, AMD is pulling another Bulldozer with this one.
Posted on Reply
#29
cucker tarlson
FordGT90Concept, post: 4121760, member: 60463"
I highly doubt AMD is going to divorce the core design between Epyc and Ryzen.
but ryzne will never have to be epyc.
is my point.
how long before 128t is not enough ?
Posted on Reply
#30
john_
voltage, post: 4121695, member: 173691"
So, they are now just doing what INTEL has been doing all along, except for the die shrinkage issue. More density is what INTEL has had Over AMD up until now.

NOTHING Innovative about that, at all. AND Amd is about to even start using the +++ HA HA!

AMD, just like apple, nothing new, just copy other peoples ideas and claim as if they thought of it themselves.
I feel your pain, I feel your sadness, I feel your desperation.

On the other hand I have to say that your post had huge success. So many quotes. If it is an attempt on trolling, I have to give it an A+++
Posted on Reply
#31
Frick
Fishfaced Nincompoop
FordGT90Concept, post: 4121752, member: 60463"
This will be terrible for gaming. Server operators will rejoice at this news but...this is very likely to be harmful in the consumer space.
Was waiting for this comment tbh. I mean the admittedly huge caveat is obviously "if keeping the clockspeed/single core IPC up" but ... now is a better time for it fo sho. And it's rumour. And the technical implementation is unknown. vOv
Posted on Reply
#32
kapone32
FordGT90Concept, post: 4121752, member: 60463"
This will be terrible for gaming. Server operators will rejoice at this news but...this is very likely to be harmful in the consumer space.
If you had a Gaming boutique with 1 PC and 12 screens you would be over the moon with the power of the VMs these would create, but not everyone does that so your statement is mostly true.
Posted on Reply
#33
Vayra86
FordGT90Concept, post: 4121752, member: 60463"
This will be terrible for gaming. Server operators will rejoice at this news but...this is very likely to be harmful in the consumer space.
Why harmful? Because of a supposed shift in focus? Its not like CPUs magically get faster single thread if you disable SMT.

If its anything like IBM PowerPC it will scale to whatever is needed and provided the workload is suitable. I'm not seeing the issue. ZEN has been a server CPU tech first and everything else second anyway, nothing new here.

The opposite applies I think. Better SMT will benefit the efficiency and thus market share of AMD CPUs and will lead to more improvements to them in the long run as well. There is only win-win here. Nobody with two brain cells ever made a 'gaming CPU'. No not even Intel; its a mere byproduct. Bump the clocks and poof gaming cpu.
Posted on Reply
#34
HD64G
FordGT90Concept, post: 4121760, member: 60463"
I highly doubt AMD is going to divorce the core design between Epyc and Ryzen. If they are, it's fine; if not, AMD is pulling another Bulldozer with this one.
They will be able to send the less than perfect dies-chiplets to the desktop with "just" 2 or 3-way SMT and with better efficiency to up the clocks closer to 5GHz for all-core boost and thus increase desktop performance again. Some IPC will be improved also i guess. Not a bad upgrade if the 16-core gives 48 threads and the 6-core gives 18 threads for when needed.
Posted on Reply
#35
Steevo
Frick, post: 4121765, member: 23907"
Was waiting for this comment tbh. I mean the admittedly huge caveat is obviously "if keeping the clockspeed/single core IPC up" but ... now is a better time for it fo sho. And it's rumour. And the technical implementation is unknown. vOv
It actually does make sense from a clock efficiency standpoint, if you have any core waiting on work that needs more than X number of refreshes to fetch it should increase IPC as long as the cache has work ready from other threads. We keep seeing more cache added meaning the number of in flight VS executing threads really will dictate how well it works in the real world.

Core 1 waiting on fetch has 30 cycles of wait time, thread 2 assigned is waiting on data from another thread to provide stored results to finish, so why not do work out of the cache while waiting as long as the latency doesn't exceed the 30 cycle wait time its an increase of IPC. The logic behind it all is impressive in hardware, I just wonder if its going to add more latency for the management or if they are going to run the management system at higher frequency to mask it.
Posted on Reply
#36
theoneandonlymrk
kapone32, post: 4121682, member: 181865"
I have said it before and I will say it again. AMD is innovating the CPU space like no one before. I fully expect Intel to respond but 4 way SMT is no joke, especially with the 20% increase in transistor density due to EUV.



I find him to a little too AMD biased
Not really relevant, information is information , it's not yet a fact regardless of who spreads it, it's a rumour ATM, with a lot of people behind it.

And if you had you might get why i mentioned it.

He believes AMD will bring zen 3 to servers next year but not necessarily to desktop.

Desktop would get the free top bins of zen 2 allowing higher clocks for a refresher while zen 3 on EUV wouldn't likely yeild great and would end up expensive ,per chip, making using it just for server seam like The logical path plus they're push to the server room should really start then, once all the trial's have run their course.

He also thinks consumers might not get smt4 anyway, but AMD don't usually segregate sku's like that so it's not likely IMHO.
Posted on Reply
#37
FordGT90Concept
"I go fast!1!11!1!"
Vayra86, post: 4121770, member: 152404"
Why harmful? Because of a supposed shift in focus? Its not like CPUs magically get faster single thread if you disable SMT.
Because you have four threads sharing the same underlying execution resources. If one of those is a game, and the others are something like BOINC, only 25%-40% of the execution time is spent on game thread which means fewer frames per second. Server loads, they care about efficiency over response times which is diametrically opposed to what games (consumer in general) needs.
Posted on Reply
#38
Steevo
FordGT90Concept, post: 4121784, member: 60463"
Because you have four threads sharing the same underlying execution resources. If one of those is a game, and the others are something like BOINC, only 25%-40% of the execution time is spent on game thread which means fewer frames per second. Server loads, they care about efficiency over response times which is diametrically opposed to what games (consumer in general) needs.
Many SMT CPUs do get faster single threaded work done if they don't have to share cache resources with competitive threads.

https://www.techpowerup.com/review/amd-ryzen-9-3900x-smt-off-vs-intel-9900k/5.html
Posted on Reply
#39
Xaled
jeffj7, post: 4121745, member: 189108"
Dont feed the troll
Hahaha 8 people already did :laugh:
Posted on Reply
#40
yotano211
voltage, post: 4121695, member: 173691"
So, they are now just doing what INTEL has been doing all along, except for the die shrinkage issue. More density is what INTEL has had Over AMD up until now.

NOTHING Innovative about that, at all. AND Amd is about to even start using the +++ HA HA!

AMD, just like apple, nothing new, just copy other peoples ideas and claim as if they thought of it themselves.
TSMC is not owned by AMD, both companies are different from each other. So I dont know why you are bashing on AMD.
Posted on Reply
#41
Xaled
Steevo, post: 4121788, member: 19251"
Many SMT CPUs do get faster single threaded work done if they don't have to share cache resources with competitive threads.

https://www.techpowerup.com/review/amd-ryzen-9-3900x-smt-off-vs-intel-9900k/5.html
These are so random scores, but yeah most of the time when SMT is off frames are higher.
As core counts is already more than enough for gaming, Why AMD isnt making something similar to what is being done in some mobile phones? Half of cores with high clocks and others with low clocks for other apps/general use and no SMT at al? Is it that hard really?
Posted on Reply
#42
Frick
Fishfaced Nincompoop
Steevo, post: 4121773, member: 19251"
It actually does make sense from a clock efficiency standpoint, if you have any core waiting on work that needs more than X number of refreshes to fetch it should increase IPC as long as the cache has work ready from other threads. We keep seeing more cache added meaning the number of in flight VS executing threads really will dictate how well it works in the real world.

Core 1 waiting on fetch has 30 cycles of wait time, thread 2 assigned is waiting on data from another thread to provide stored results to finish, so why not do work out of the cache while waiting as long as the latency doesn't exceed the 30 cycle wait time its an increase of IPC. The logic behind it all is impressive in hardware, I just wonder if its going to add more latency for the management or if they are going to run the management system at higher frequency to mask it.
Yeah well as Ford is hinting below the other huge caveat is "software support".

FordGT90Concept, post: 4121784, member: 60463"
Because you have four threads sharing the same underlying execution resources. If one of those is a game, and the others are something like BOINC, only 25%-40% of the execution time is spent on game thread which means fewer frames per second. Server loads, they care about efficiency over response times which is diametrically opposed to what games (consumer in general) needs.
Is that last statement true though, today? How do you define "consumer" anyway? Because for nongamers and general office workers PC performance is basically not interesting as long as one has a semi modern system, an SSD and enough RAM. That's kinda what I meant with "it's a better time for it" in regards to this being a new Bulldozer. If the basic performance and power design is there the other stuff basically don't matter in consumer land, from my point of view.

I mean sure if a hypothetical new quad (or dual, whatever) SMT design meant like a 25% drop in frames in game X it would be bad, but given their current status I really, really, REALLY don't think AMD would implement that in consumer/gamer CPUs. Yes their entire design currently is based on chiplets, but if they take over enough of the enterprise market it might make fiscal sense to create this monster for that specific purpose. That is where the money is after all.

Xaled, post: 4121802, member: 158027"
These are so random scores, but yeah most of the time when SMT is off frames are higher.
As core counts is already more than enough for gaming, Why AMD isnt making something similar to what is being done in some mobile phones? Half of cores with high clocks and others with low clocks for other apps/general use and no SMT at al? Is it that hard really?
Given how AMD makes CPUs, yes. Well not hard, but impractical with their current strategy. Also remember phones have vastly different requirements than PCs when it comes to mostly power and th related heat.
Posted on Reply
#43
AnarchoPrimitiv
theoneandonlymrk, post: 4121680, member: 82332"
Doesn't anyone watch Moore's laws deads videos on here, his leak gave him smt4 leaked weeks ago?
I've been a Moore's Law is dead fan before you were..... Lol

kapone32, post: 4121682, member: 181865"
I have said it before and I will say it again. AMD is innovating the CPU space like no one before. I fully expect Intel to respond but 4 way SMT is no joke, especially with the 20% increase in transistor density due to EUV.



I find him to a little too AMD biased
He's not AMD biased at all, and I don't understand under what criteria you're making such a judgment. Intel basically has NOTHING going on right now, at least nothing that enthusiasts care about, so he has nothing to talk about with respect to Intel. And, whatever Intel does have going on, it's nothing exciting. He's not a journalist and does not have a responsibility to offer equal coverage of all entities whether they deserve it or not. You saying he is biased is like saying the news is biased toward natural disasters for heavily covering a major hurricane that just occurred. It's not biased when you're covering the most interesting PC news and all that news just so happens to be coming from one company.

I'm also assuming that you made that incorrect judgment WITHOUT watching much of his content because if you had watched more, you'd know that his personal system is an Intel system. Also, he's a true enthusiast, and the majority of True enthusiasts (not fanboys, fanboys are not enthusiasts) DO NOT, for example, respect Nvidia for the crap they pull and the harm they do to the community as a whole (e.g. Jacking up prices for no other reason than achieving a 100% profit margin and because fanboys will never check their power by withholding their money from their products) and the majority of enthusiasts know that Ray tracing is not polished enough to warrant buying an RTX card and its price premium... however, knowing and expressing such an opinion is not, by default, a form of bias toward and in favor of AMD, and anyone suggesting that being critical of one company automatically makes you biased toward their competition is logically fallacious and is creating a false binary that does not exist. I offer those as examples.

I really tired of those in this community that try and force their false binary reasoning on others. It's just like conservatives who automatically declare that you're a liberal or a Democrat for criticizing Trump, as if in the vast plethora of opinions and philosophies in our reality, an individual is only limited to two in some sort of perverted, "if, then, else" distorted logic and that any criticism of one automatically and by default makes you the other.

FordGT90Concept, post: 4121752, member: 60463"
This will be terrible for gaming. Server operators will rejoice at this news but...this is very likely to be harmful in the consumer space.
There seems to be this shortcut to thinking that so many "gamers" [that claim to be enthusiasts] constantly perform, namely the completely unfounded assumption and belief that "gaming" comprises the majority of use cases for x86 CPUs and that "gaming", and pleasing gamers, is the paramount concern, or should be, of AMD/Intel when that's not the case.

Gaming does NOT comprise the bulk of the T.A.M. and revenue for the x86 industry, enterprise does. An x86 manufacturer would be committing suicide by choosing to make gaming their number one priority, and yet so many gamers think that it should be, and furthermore, so many "gamers" have this ridiculous mentality that the measure of quality and validity in an x86 CPU is assessed solely by the FPS it can achieve...does anybody else know what I'm talking about or has observed this? It basically occurs everyday on WCCFTech, I witnessed so many fools basically making the argument that a 5 fps inferiority for ryzen 3000 when compared to Intel makes it a failure. Are these people serious or are they so clouded by the psychological idiocy of in group/out group mentality and social identity theory (the two major phenomenon behind fanboy behavior) that they're completely blind to reason?
Posted on Reply
#44
zlobby
Only 4? Why stop there? /kappa

AnarchoPrimitiv, post: 4121807, member: 168101"
I witnessed so many fools basically making the argument that a 5 fps inferiority for ryzen 3000 when compared to Intel makes it a failure. Are these people serious or are they so clouded by the psychological idiocy of in group/out group mentality and social identity theory (the two major phenomenon behind fanboy behavior) that they're completely blind to reason?
Nah, they just stupid, plain and simple.
Posted on Reply
#45
kapone32
AnarchoPrimitiv, post: 4121807, member: 168101"
I've been a Moore's Law is dead fan before you were..... Lol



He's not AMD biased at all, and I don't understand under what criteria you're making such a judgment. Intel basically has NOTHING going on right now, at least nothing that enthusiasts care about, so he has nothing to talk about with respect to Intel. And, whatever Intel does have going on, it's nothing exciting. He's not a journalist and does not have a responsibility to offer equal coverage of all entities whether they deserve it or not. You saying he is biased is like saying the news is biased toward natural disasters for heavily covering a major hurricane that just occurred. It's not biased when you're covering the most interesting PC news and all that news just so happens to be coming from one company.

I'm also assuming that you made that incorrect judgment WITHOUT watching much of his content because if you had watched more, you'd know that his personal system is an Intel system. Also, he's a true enthusiast, and the majority of True enthusiasts (not fanboys, fanboys are not enthusiasts) DO NOT, for example, respect Nvidia for the crap they pull and the harm they do to the community as a whole (e.g. Jacking up prices for no other reason than achieving a 100% profit margin and because fanboys will never check their power by withholding their money from their products) and the majority of enthusiasts know that Ray tracing is not polished enough to warrant buying an RTX card and its price premium... however, knowing and expressing such an opinion is not, by default, a form of bias toward and in favor of AMD, and anyone suggesting that being critical of one company automatically makes you biased toward their competition is logically fallacious and is creating a false binary that does not exist. I offer those as examples.

I really tired of those in this community that try and force their false binary reasoning on others. It's just like conservatives who automatically declare that you're a liberal or a Democrat for criticizing Trump, as if in the vast plethora of opinions and philosophies in our reality, an individual is only limited to two in some sort of perverted, "if, then, else" distorted logic and that any criticism of one automatically and by default makes you the other.
I know what you are saying and do not at all disagree with the Intel opinion. I have an all AMD system I do watch his videos, he has a lot of correct assessments like the Vega 64 and his rants about Nvidia's business practices, but we are talking about rumours. Maybe it is the time when I started watching his videos at which point the internet was awash with AMD love. The thing is that Intel has been making moves in the space like hiring tech journalists and taking people from AMD. Jim Keller is the architect of Ryzen and by association Threadripper so then too Epyc as they are all based on the same architecture. It is clearly speculation but do you not think that he will come up with innovations for Intels architecture? They are a company that is a10x the size of AMD and as such cannot respond in the short term but I can guarantee that Intel have been working on something since they hired Keller in 2018. Everything Intel does is internal so we really have no idea what stage they are at in their next architecture. Weve been gettiing the same tick tock from Intel in the short term agreed. As far as tech journalists go there is no doubt in my mind that social media at least in some small part has contributed to the success of AMD currently. I am sure that Intel is aware of that and as such will use that to their own perceived advantage. The thing for me though is that AMD is just hitting it out of the park right now. Who would have thought that the 3900X would be faster than it's Threadripper big brothers? When the 3950X is actually released I am sure there will be much love for that chip too.
Posted on Reply
#46
Vya Domus
Consumer chips can always have two of the threads disabled or power gated, the matter with regards to gaming is a non issue. Plus it wouldn't be that far fetched for AMD to have a separate design just for servers.
four virtual threads per core
Can we stop this already, it's 2019 for Christ sake, computer architecture isn't that cryptic anymore. There is nothing virtual about them, they are as physical as they can get, you could literally put your finger on the corresponding piece of silicon if you had a chip scaled up.
Posted on Reply
#47
techguymaxc
Aldain, post: 4121702, member: 170164"
Are you F stupid?? When did intel implement 4 way smt??
Here you go:
https://en.wikipedia.org/wiki/Xeon_Phi#Knights_Ferry
Larrabee before that:
https://en.wikipedia.org/wiki/Larrabee_(microarchitecture)

Based on ancient P54 (Pentium non-MMX uarch) even. >2-way SMT is rather trivial to implement, and nothing new. The real challenge will be extracting greater multi-thread performance. That and not falling victim to Windows scheduler inefficiencies.
Posted on Reply
#48
Steevo
Frick, post: 4121805, member: 23907"
Given how AMD makes CPUs, yes. Well not hard, but impractical with their current strategy. Also remember phones have vastly different requirements than PCs when it comes to mostly power and th related heat.
This, by nature X86-64 is a CISC architecture, it has to be for the plethora of software expected to run on it with decent performance.

Low power ARM cores are RISC and since its a "walled garden" they can optimize code for the hardware. They suck at general performance computing but excel at what they are designed for specifically.


techguymaxc, post: 4121844, member: 190760"
Here you go:
https://en.wikipedia.org/wiki/Xeon_Phi#Knights_Ferry
Larrabee before that:
https://en.wikipedia.org/wiki/Larrabee_(microarchitecture)

Based on ancient P54 (Pentium non-MMX uarch) even. >2-way SMT is rather trivial to implement, and nothing new. The real challenge will be extracting greater multi-thread performance. That and not falling victim to Windows scheduler inefficiencies.
"An empirical performance and programmability study has been performed by researchers,[101] in which the authors claim that achieving high performance with Xeon Phi still needs help from programmers and that merely relying on compilers with traditional programming models is still far from reality. "

Anything not scientific community, standard server related? No doubt 4 way SMT works better on highly out of order processing loads, if we could make everything in order we would only need a 4 core CPU with no SMT. Practical implementation for use is the foundation of any of this, and Intel hasn't succeeded at that with any Phi products, just like they don't have a graphics card yet, just artist renderings....
Posted on Reply
#49
techguymaxc
Steevo, post: 4121848, member: 19251"
This, by nature X86-64 is a CISC architecture, it has to be for the plethora of software expected to run on it with decent performance.

Low power ARM cores are RISC and since its a "walled garden" they can optimize code for the hardware. They suck at general performance computing but excel at what they are designed for specifically.





"An empirical performance and programmability study has been performed by researchers,[101] in which the authors claim that achieving high performance with Xeon Phi still needs help from programmers and that merely relying on compilers with traditional programming models is still far from reality. "

Anything not scientific community, standard server related? No doubt 4 way SMT works better on highly out of order processing loads, if we could make everything in order we would only need a 4 core CPU with no SMT. Practical implementation for use is the foundation of any of this, and Intel hasn't succeeded at that with any Phi products, just like they don't have a graphics card yet, just artist renderings....
https://www.pugetsystems.com/labs/hpc/Windows-10-with-Xeon-Phi-733/

If you can run an O.S. on it, it's a general purpose microprocessor, and I think that qualifies it for this discussion.
Posted on Reply
#50
R0H1T
FordGT90Concept, post: 4121752, member: 60463"
This will be terrible for gaming. Server operators will rejoice at this news but...this is very likely to be harmful in the consumer space.
They will in all likelihood disable the extra threads on consumer chips, as many rumors have been predicting/speculating over the last few months.
Posted on Reply
Add your own comment