Tuesday, September 26th 2017

AMD Zen 2 Architecture: Socket AM4, 2019, Code-named "Matisse"

AMD's Zen-based Ryzen and Threadripper have been said by the company as representing the "worst case scenario" of performance for their architecture. This is based on the fact that there are clear areas for improvement that AMD's engineers were keenly aware of even at the moment of Zen's tapping-out; inadvertently, some features or improvements were left on the chopping block due to time and budget constraints. As unfortunate as this is - who wouldn't love to have even more performance on their AMD processors - this also means AMD has a clear starting point in terms of improving performance of their Zen micro-architecture.

Spanish website Informatica Cero have gotten their hands on what they say is an exclusive, real piece of information from inside AMD, which shows the company's CPU roadmap until 2019, bringing some new details with it. On the desktop side, there's mention of AMD's "Pinacle Ridge" as succeeding the current Zen-based "Summit Ridge" Ryzen CPUs in 2018. These leverage the same Summit Ridge architecture, but with a performance uplift; this plays well into those reports of 12 nm being used to manufacture the second-generation Ryzen: it's an AMD tick, so to say. As such, the performance uplift likely comes from increased frequencies at the same power envelope, due to 12 nm's denser manufacturing design.
Another interesting tidbit from this slide is the confirmation of AMD's Zen 2 processors being released in 2019. Based on the company's revised Zen 2 cores (which should see those architectural improvements we were talking about in the beginning of this piece), AMD also seems to be doing away with the ridge-like codenames they've been using in recent times, and taking a more artistic approach. Matisse seems to be the code-name for AMD's Zen 2 architecture, and if you know your painters, it looks like AMD is betting on its Zen 2 cores to further define and influence CPU design. There is no mention of increased Zen threads, however, which likely points towards AMD keeping the same 8-core, 16-thread design as with their current Summit Ridge. This likely plays into AMD's plan for keeping the AM4 socket relevant - while good for consumers who might want to see longevity in their platform, it does prevent AMD from making more radical design changes to their architecture.
Jumping towards its notebook/desktop APU designs, AMD is looking to accompany Matisse with their code-named Picasso APUs for 2019. These should still leverage the company's upcoming Raven Ridge line of APUs, which pair Vega 11 NCUs on the graphical side of the equation and up to 8 Zen threads (4-core CPUs with SMT). However, usage of 12 nm processes across both its Zen CPU cores and Vega 11 NCUs apparently give AMD confidence that they can improve Picasso over Raven Ridge not only in the performance fields, but also the power consumption metrics.

A little bit of lore: AMD seems to be changing their code-names towards painters, and the marriage of Matisse in the CPU and Picasso in the APU space isn't a coincidence. Both artists are considered two of the most important players in the arts space of the beginning of the twentieth century, responsible for both defining and developing the arts. AMD's choice of both of them is, therefore, not a fluke; it's likely more of an artistic take on the company's vision of further defining itself as a key innovative player in the CPU space.

Sources: Informatica Cero, Pablo Picasso @ Wikipedia, Henri Matisse @ Wikipedia
Add your own comment

65 Comments on AMD Zen 2 Architecture: Socket AM4, 2019, Code-named "Matisse"

#1
Vayra86
If AMD still wants to capture all those potential buyers of an 8700k, they had better push out some leaks regarding high clock speeds, fast.
Posted on Reply
#2
birdie
Vayra86 said:
If AMD still wants to capture all those potential buyers of an 8700k, they had better push out some leaks regarding high clock speeds, fast.
As much as people hate Intel, they are currently the only company in the semiconductor industry which can churn out high speed (as in high frequency) 14nm chips, soon to be followed by 10nm.
Posted on Reply
#3
Ferrum Master
Where's the cut down Vincent van Gogh version? :laugh:
Posted on Reply
#4
jigar2speed
Looking at the results so far i feel Ryzen at 4.5GHZ should be extremely competitive if not overwhelm Intel's top dogs.
Posted on Reply
#5
R-T-B
birdie said:
As much as people hate Intel, they are currently the only company in the semiconductor industry which can churn out high speed (as in high frequency) 14nm chips, soon to be followed by 10nm.
GloFo actaully supposedly has a non-lpp process that could probably touch 5GHz. No idea what'd do to Zen's energy/heat advantage though.
Posted on Reply
#6
XiGMAKiD
Ferrum Master said:
Where's the cut down Vincent van Gogh version? :laugh:
Good one :roll:

If they can keep this up, it's gonna make PC great again
Posted on Reply
#7
Vya Domus
Vayra86 said:
If AMD still wants to capture all those potential buyers of an 8700k, they had better push out some leaks regarding high clock speeds, fast.
Buyers of the 8700K are unlikely to touch anything from AMD to begin with.

R-T-B said:
GloFo actaully supposedly has a non-lpp process that could probably touch 5GHz. No idea what'd do to Zen's energy/heat advantage though.
Key word , probably.

birdie said:
As much as people hate Intel, they are currently the only company in the semiconductor industry which can churn out high speed (as in high frequency) 14nm chips, soon to be followed by 10nm.
It's not that simple. Every one of these semiconductor companies can produces chips that push the frequency limits of these nodes , they just can't do it on all architectures/products because they prioritize other things.
Posted on Reply
#8
GreiverBlade
Vayra86 said:
If AMD still wants to capture all those potential buyers of an 8700k, they had better push out some leaks regarding high clock speeds, fast.
well specially as it need a new chipset albeit same socket ....

a 8700K is becoming less and less desirable for me, hell even the 8600K is loosing his appeal in favor of a R5 1600/1600X (or R7 if i had the mean to take a 8700K instead of a 8600K) .... even with the actual clock speed and meager OC potential

since the Zen 2 will be also on AM4 ...

but will it need a new chipset? probably not, will it be retro compatible? probably ... we are not talking about Intel in this case ...


the last good point of Intel, is their set of blind follower that will buy my 6600K+Z170G7+16gb 2800 literally near the initial launch price .... (did some search checked some similar offers ... i could buy a full set in AM4 CPU/Mobo + compatible RAM and still have ~200chf if taking a R5 or ~100-145chf if taking a R7 :laugh: )

Vya Domus said:
Buyers of the 8700K are unlikely to touch anything from AMD to begin with.
"close minded blind follower" is the keyword :laugh: (joke)
Posted on Reply
#9
Mark T
First table looks v. dodgy as it suggests '2017 Bristol Ridge' includes a Polaris GPU...
Posted on Reply
#10
theoneandonlymrk
Vya Domus said:
Buyers of the 8700K are unlikely to touch anything from AMD to begin with.



Key word , probably.



It's not that simple. Every one of these semiconductor companies can produces chips that push the frequency limits of these nodes , they just can't do it on all architectures/products because they prioritize other things.
Glofo used SOi on bulldozer to hit 5ghz successfully and recently announced 14nm hp+ on SOi ,regarded as a node for automotive.
Its interesting that automotive is mentioned ,but on 12nmlpp similar but slightly conflicting, SOI on zen would push the price up and allow higher clocks but the cost rises too so im thinking this 12nm is all the goodies in the 14nm hpe minus the SOi.
Anyway good times ahead ,both big x86 players are really in the game.
Posted on Reply
#11
TheinsanegamerN
theoneandonlymrk said:
Glofo used SOi on bulldozer to hit 5ghz successfully and recently announced 14nm hp+ on SOi ,regarded as a node for automotive.
Its interesting that automotive is mentioned ,but on 12nmlpp similar but slightly conflicting, SOI on zen would push the price up and allow higher clocks but the cost rises too so im thinking this 12nm is all the goodies in the 14nm hpe minus the SOi.
Anyway good times ahead ,both big x86 players are really in the game.
Perhaps you missed that, while it could hit 5GHz, it also consumed well north of 200 watts while doing so, giving it worse performance/watt then stock, while still getting its behind whooped by a stock i7?

As the previous poster pointed out, sure GloFo could do higher clocks, but it would eliminate the power advantage ryzen has, not to mention make the mobile market a hopeless endeavor AGAIN. I would ratherAMD continue improving IPC rather then getting into another disastrous clockspeed war.
Posted on Reply
#12
Vya Domus
theoneandonlymrk said:
Glofo used SOi on bulldozer to hit 5ghz successfully and recently announced 14nm hp+ on SOi ,regarded as a node for automotive.
Its interesting that automotive is mentioned ,but on 12nmlpp similar but slightly conflicting, SOI on zen would push the price up and allow higher clocks but the cost rises too so im thinking this 12nm is all the goodies in the 14nm hpe minus the SOi.
Anyway good times ahead ,both big x86 players are really in the game.
Thing is , I don't want them to start a clock speed race again. If Zen+ clocks a little higher that's good but they already established the Ryzen lineup as high core count low price products and that's what they need to maintain. Intel can maintain their clock speed advantage we need diversity.
Posted on Reply
#13
EarthDog
jigar2speed said:
Looking at the results so far i feel Ryzen at 4.5GHZ should be extremely competitive if not overwhelm Intel's top dogs.
Truth...for todays top dogs... what about in 2019 when these release?
Posted on Reply
#14
R0H1T
R-T-B said:
GloFo actaully supposedly has a non-lpp process that could probably touch 5GHz. No idea what'd do to Zen's energy/heat advantage though.
You mean at least 5GHz ;)
Each IBM z14 SC CPU consists of 6.1 billion transistors, runs at 5.2 GHz and contains 10 cores with dedicated 6 MB L2 per core (2MB L2 for instructions, 4MB L2 for data) and 128 MB shared L3.
https://www.anandtech.com/show/11862/globalfoundries-weds-finfet-and-soi-in-14hp-process-tech-for-ibm-z14-cpus
Posted on Reply
#16
R0H1T
EarthDog said:
Those are mainframe cpus...
Yes but that shows GF can do high density, high clock CPUs.
The z14 along with the SPARC M8 are the only high end processors that can do 5GHz 24x7 & that has to count as some sort of achievement, considering the Intel advantage wrt x86 & their own foundries.
Posted on Reply
#17
EarthDog
But SPARC is not the same as x86. Its a proof of concept I guess, but not proof on x86 CPUs.
Posted on Reply
#18
R0H1T
EarthDog said:
But SPARC is not the same as x86. Its a proof of concept I guess, but not proof on x86 CPUs.
What do you mean proof of concept?
With little fanfare, Oracle formally tore the wraps off its SPARC M8 data-center-class processors and servers on Monday.

The House of Larry Ellison unveiled the fifth-generation successor to the fourth-generation SPARC M7, along with five server models, and a commitment to support its Solaris operating system for at least another 17 years. That's all the way to 2034 and potentially beyond.

The M8, first [COLOR=rgb(221, 0, 0)]sighted and described in July
, is a modest uptick from the M7. That's the M7 unveiled in 2015 with some anti-malware mechanisms, plus encryption and SQL query hardware acceleration built in.[/COLOR]

The M8 is clocked to 5GHz, topping the M7's 4.13GHz. The latest generation processor has a 32KB L1 instruction cache, where the M7 has 16KB. The M8 has a 128KB L2 data cache per core, whereas the M7 has 256KB per core pair. Clearly, Oracle has figured its CPU needs a larger fast code cache near the execution engines and a rejig of the room at the back for stuff like program variables. The M8 can also issue up to four instructions at a time, compared to the dual-issue M7. This is about what you'd expect from a server-grade chip in 2017.

Interestingly, you can have up to eight M8 processors per machine, whereas the M7 supported a maximum of 16. This is a pattern for Oracle; the M7's predecessor, the SPARC M6, supported up to 32 sockets per system. In other words, Oracle believes you want fewer chips per box. The M8 can handle up to 8TB of RAM, whereas the M7 can hit 16TB due to the higher socket limit.
Beyond that, the M8 and the M7 are pretty much the same on paper. Both have 32 cores – the M8's are codenamed "SPARC Next" – and eight threads per core. (We've been [COLOR=rgb(221, 0, 0)]waiting
for SPARC Next to arrive since the turn of the year.)[/COLOR]

Both are 20nm parts fabricated by TSMC, and both support up to 16 DDR4 RAM DIMMs. Both have 16KB of L1 data cache, 256KB of L2 instruction cache per four cores, and 64MB of shared L3 cache. Both support 16GB memory pages, and both had the aforementioned security measures and database acceleration hardware on the die.

https://www.theregister.co.uk/2017/09/19/oracle_sparc_m8_solaris/
Posted on Reply
#19
Vya Domus
Can't compare chips like that.

FX CPUs did 5 Ghz easily too , on a node that is what , half a decade old ? It's all down to the node and architecture. How many pipeline stages it features , how wide the instructions are , how much cache there is and how fast it runs. There are so many things that affect clock speed besides the manufacturing process.
Posted on Reply
#20
EarthDog
Vya Domus said:
Can't compare chips like that.

FX CPUs did 5 Ghz easily too , on a node that is what , half a decade old ? It's all down to the node and architecture. How many pipeline stages it features , how wide the instructions are , how much cache there is and how fast it runs. There are so many things that affect clock speed besides the manufacturing process.
this... cant compare is my point.
Posted on Reply
#21
R0H1T
Vya Domus said:
Can't compare chips like that.

FX CPUs did 5 Ghz easily too , on a node that is what , half a decade old ? It's all down to the node and architecture. How many pipeline stages it features , how wide the instructions are , how much cache there is and how fast it runs. There are so many things that affect clock speed besides the manufacturing process.
Oh absolutely, the point is GF (as a foundry) can do that. Now whether Zen+ or Zen2 can achieve those clocks is a totally different thing.
Posted on Reply
#22
EarthDog
R0H1T said:
Oh absolutely, the point is GF (as a foundry) can do that. Now whether Zen+ or Zen2 can achieve those clocks is a totally different thing.
lol
Posted on Reply
#23
R0H1T
EarthDog said:
lol
What happened :confused:
Posted on Reply
#24
Vya Domus
R0H1T said:
Oh absolutely, the point is GF (as a foundry) can do that.
That's why we can't agree , that statement is not entirely relevant. Like I said GF manufactured the FX chips on 32nm and those could hit 5 Ghz as well.

The fact that GF can make 5Ghz chips on their newest nodes doesn't tell us anything with regards to Zen+ or Zen 2. Can those do 5 Ghz on that node ? No way to tell , it's all 100% guesses at this point.

And it's not even entirely about technical capabilites , maybe GF and AMD can make these chips run at 5 ghz+ , but is it cost effective ? Those IBM CPUs must cost a fortune.
Posted on Reply
#25
theoneandonlymrk
TheinsanegamerN said:
Perhaps you missed that, while it could hit 5GHz, it also consumed well north of 200 watts while doing so, giving it worse performance/watt then stock, while still getting its behind whooped by a stock i7?

As the previous poster pointed out, sure GloFo could do higher clocks, but it would eliminate the power advantage ryzen has, not to mention make the mobile market a hopeless endeavor AGAIN. I would ratherAMD continue improving IPC rather then getting into another disastrous clockspeed war.
I didn't miss your pady wtf. I have a 5ghz chip right here ,i know what it uses and i understand nodes.

The contradiction on forums is getting ridiculous, who mentioned mobile, not me im talking Hpc and how do you improve Ipc without arch changes on mass without frequency going up.
It's ok for intel to pr blurb 4.8ghz ocs with 5ghz via delid for its new coffee lake but Amd better not try for higher clocks too because you won't like it.
Do you not read,,,, that 8800k at 5ghz is not going to be power efficient.
Shit check the recent I9 world record , 1000watts plus through an Intel cpu.
Physics are what they are and intel play with the same rulebook.
. and all decrying 14nmhpe irrelevant because its used for mainframes, its not just for mainframes its a foundry node you could easily make x86 on it, they aren't and i started this by saying they could but are not.
But they could be using some of its pdk improvement recipe points on its 12nm lpp as to me that's whats implied by separate news items rumour and a bit of smoke.
Posted on Reply
Add your own comment