Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[Tigerick]

Kind of boring, I have to say.

The STX1 would seem quite big to be a monolithic die.

It would be surprising if AMD ends up with ~250+ mm2 monolithic die to go against chiplet based Meteor Lake and Arrow Lake. I don't see this as likely...

STX1 is AMD's answer to upcoming M3 Pro which is based on N3B. Come to think of it, STX1 might be supporting 256-bit LPDDR5x 8533 cause with 16 cores and 2048ALU, STX1 need more bandwidth than before..

 

[Kaluan]

The Gigabyte was referring to support of future AM5 processor while launching a Ryzen Server. APU would be irrelevant to this product line.

Are they tho?



The "and NPU" section demarcation heavily implies it's not Rembrandt as well.
Also it's funny they also call TR 5000 "HEDT", since they only had Workstation/Pro variants that gen. I'm guessing TR 7000 changes that.

Kind of boring, I have to say.

The STX1 would seem quite big to be a monolithic die.

It would be surprising if AMD ends up with ~250+ mm2 monolithic die to go against chiplet based Meteor Lake and Arrow Lake. I don't see this as likely...

Phoenix's competition, namely stuff like i7-1370P is already > 250mm2+. Taking this leak at face value, a similar AMD implementation but 2 nodes ahead of Intel and knowing how tiny even AMD "P-cores" are compared to Intel's, you really think that would be the case? I personally can't see it.

That being said, I kinda doubt the monolithic part the most. Also bringing up Meteor Lake shouldn't even be a thing anyway, it's basically vaporware at this point, even for mobile. Arrow Lake is something else, we'll see.

 

[naukkis]

Kind of boring, I have to say.

The STX1 would seem quite big to be a monolithic die.

It would be surprising if AMD ends up with ~250+ mm2 monolithic die to go against chiplet based Meteor Lake and Arrow Lake. I don't see this as likely...

Monolithic is both better and cheaper. Going chiplet is always bad idea if there's possibility to make design monolithic. Intel is in seriously desperate situation so they choose stupid chiplet-design instead going with enough nature manufacturing process making monolithic design possible.

 

[DrMrLordX]

Nothing, he doesn't want to admit he's fussy about AMD's release schedule. the pandemic messed up their scheduling. IIRC Zen 4 was later than expected at 24 months. Zen 5 returns to the 14-18 month schedule they had prior.

Nah my complaint about the XT CPUs was that they re-released the same product at its old price point just to allow them to bump it's successor's price. Turns out I was right. Then the 5900X sold out for 3+ months solid at the higher price anyway, so, it didn't really matter.

The same thing likely won't happen again. Zen3 demand was like lightning in a bottle. We shouldn't expect a run on AMD consumer CPUs like that to happen again anytime soon. Supplies are better while the economy worldwide is driving down demand for a lot of things.

 

[A///]

Nah my complaint about the XT CPUs was that they re-released the same product at its old price point just to allow them to bump it's successor's price. Turns out I was right. Then the 5900X sold out for 3+ months solid at the higher price anyway, so, it didn't really matter.

It sold out because everyone and their nan was buying one. Same with the 7950X3D. I don't think it's a paper launch like some are claiming. The amanager at the local best buy claims they've sold nearly a thousand units since launch. I don't know why amd didn't lead with the 3d processors from the beginning. no one wanted the original release products and stores had to combo deal them to move them.

 

[A///]

Monolithic is both better and cheaper. Going chiplet is always bad idea if there's possibility to make design monolithic.

No it isn't. Chiplet is cheaper with caveat. You the consumer will never benefit from the lower production cost. The company will pocket the money and still charge you rates that a monolithic chip will cost. Monolithic becomes increasingly harder when it comes to yield rates as the area of possible defect rates rise, thus increasing binning rates to lower end processors. Suddenly your high return product devolves into a f tier product you sell because it had so many defects. Waste of time and money.

Furthermore this isn't Intel's first go at chiplets. AMD weren't the first either. Neither them or the preceding companies were on smack when they decided to go with a chiplet direction to reduce their cost and increase yield rates.

 

[naukkis]

No it isn't. Chiplet is cheaper with caveat. You the consumer will never benefit from the lower production cost. The company will pocket the money and still charge you rates that a monolithic chip will cost. Monolithic becomes increasingly harder when it comes to yield rates as the area of possible defect rates rise, thus increasing binning rates to lower end processors. Suddenly your high return product devolves into a f tier product you sell because it had so many defects. Waste of time and money.

Furthermore this isn't Intel's first go at chiplets. AMD weren't the first either. Neither them or the preceding companies were on smack when they decided to go with a chiplet direction to reduce their cost and increase yield rates.

Chiplet is cheaper option when chip size is near reticle limit - for 250mm2 chips chiplet isn't cheaper option if yields are somehow decent. Intel Meteor Lake specially is damn costly design for such a low grade consumer chip - huge silicon interposer and state of art manufactured chiplets for both cpu and gpu. For doing such a design for quite affordable consumer chip is nothing but desperate move and sure isn't increasing product margins.

 

[A///]

Chiplet is cheaper option when chip size is near reticle limit - for 250mm2 chips chiplet isn't cheaper option if yields are somehow decent. Intel Meteor Lake specially is damn costly design for such a low grade consumer chip - huge silicon interposer and state of art manufactured chiplets for both cpu and gpu. For doing such a design for quite affordable consumer chip is nothing but desperate move and sure isn't increasing product margins.

You're presuming perfect yields here on large monolithic dies. That has never been the case. No one outside of Intel knows how much their wafers cost but it certainly can't be cheap with a decent defect rate and needing to bin down on large monolithic dies. Intel nor TSMC could produce something like the Sierra Forest die without having to break it down into chiplets. The defect rate per wafer is high enough where a lot of those expensive Xeon based dies, for example, would have to be binned down for one reason or another, or completely scrapped.

Let's assume Intel's manufacturing is worse than TSMCs, which it is for the most part. It would be safe to assume Intel's bleeding edge is still seeing anywhere from 15-25% scrap rate per wafer. Cost becomes a blurred issue because no one outside of Intel knows the true cost. Using TSMC as a variable isn't accurate because their wafer price is to their customers. They still need to make a net profit on each wafer.

Intel despite lowering their prices vs AMD sells enough in volume to make a net gain in profit. If Intel were throwing money into the wind they would be in worse financial spirits.

 

[naukkis]

You're presuming perfect yields here on large monolithic dies. That has never been the case. No one outside of Intel knows how much their wafers cost but it certainly can't be cheap with a decent defect rate and needing to bin down on large monolithic dies. Intel nor TSMC could produce something like the Sierra Forest die without having to break it down into chiplets. The defect rate per wafer is high enough where a lot of those expensive Xeon based dies, for example, would have to be binned down for one reason or another, or completely scrapped.

I wasn't commenting on large monolithic dies. I was commenting supposedly ~250mm2 AMD chip against Intel Meteor Lake - and AMD have zero reason to go costlier and worse chiplet designs for their mainstream APU as their manufacturer can produce 250mm2 dies with decent yields.

 

[Doug S]

No it isn't. Chiplet is cheaper with caveat. You the consumer will never benefit from the lower production cost. The company will pocket the money and still charge you rates that a monolithic chip will cost. Monolithic becomes increasingly harder when it comes to yield rates as the area of possible defect rates rise, thus increasing binning rates to lower end processors. Suddenly your high return product devolves into a f tier product you sell because it had so many defects. Waste of time and money.

Furthermore this isn't Intel's first go at chiplets. AMD weren't the first either. Neither them or the preceding companies were on smack when they decided to go with a chiplet direction to reduce their cost and increase yield rates.

It depends on your product portfolio. Today if you want stuff that ranges from laptop to server you need a bunch of different dies, which means separate efforts for floor planning, timing closure, mask sets, etc. for fairly large chips. If you design a few smaller chiplets that you can mix n match to cover a wide range of products you may come out ahead - and some of the chiplets may be usable across multiple CPU/GPU core generations (i.e. memory & PCIe controllers) for further savings.

Sure chiplets are ALSO good if your yields are a problem, but that's mostly taken care of by doing stuff like Apple does where they sell Macs with parts that have bad CPU or GPU cores as a somewhat lower spec version - disabling additional cores so even if there's one only bad CPU core you might sell it as a part missing 4 CPU and 4 GPU cores. That's valuable too - if AMD is in the early stages of design for Zen 7 right now they may be designing it for TSMC's N2 but today they don't know what those yields are going to look like.

OK sure AMD is not designing for the bleeding edge like Apple - at least not today - they are taking TSMC's processes once they are fully mature so they've kind of been assuming "we won't move to this new process until the yields are good" but that's a risk. If the yields don't get good with N2 then they can't sell Zen 7, instead they'd have to port it to N2e or whatever fixes N2's problems, meaning the launch is delayed by many months and Intel benefits. If Intel gets their act together and starts executing on their processes like they claim they will, AMD is going to need to go more bleeding edge with TSMC to stay ahead or possibility merely to keep up.

 

[Geddagod]

Zen 5 leek
uOP cache removed, L1i x2 the size with relatively little latency increase (magik)
Dual 6 wide decode clusters , like gracemont but with auto prediction of critical branches
8 wide rename/dispatch
~480 entry ROB, Reg files all see increases
load/store system substantially improved
L2 goes to 1.5 MB per core (private) with marginally higher latency
L3 goes to 3 MB per core and no longer a victim cache
+15-25% IPC gains
posting the YT video rn guys







...in all seriousness though where are all the Zen 5 leaks? MTL is supposed to (with good volume lol) release Q1 2024, and that's around when leakers are claiming Zen 5 releases as well

 

[Geddagod]

Chiplet is cheaper option when chip size is near reticle limit - for 250mm2 chips chiplet isn't cheaper option if yields are somehow decent. Intel Meteor Lake specially is damn costly design for such a low grade consumer chip - huge silicon interposer and state of art manufactured chiplets for both cpu and gpu. For doing such a design for quite affordable consumer chip is nothing but desperate move and sure isn't increasing product margins.

I believe Intel would still be selling RPL chips as the 'cost effective' solution, and keep MTL focused in the high end. And isn't mobile a large industry with decently high margins for Intel too?
Also the iGPU isn't exactly state of the art. TSMC 5nm by 2024 would be extremely mature.
The CPU tile pretty much has to be on Intel 4 though. Intel has to release something for their targets and it would also be great showcase of the node's performance considering that it should be not too far off the design of GLC.
I agree though, in the sense that I think mobile chips should be kept strictly monolithic. I hope the ultra low power LNL chips are monolithic, but based on Intel slides, it doesn't look like it. The low-power SOC island might be something special though, if you can save a bunch of power from virtually turning off the CPU tile.

 

[A///]

It depends on your product portfolio. Today if you want stuff that ranges from laptop to server you need a bunch of different dies, which means separate efforts for floor planning, timing closure, mask sets, etc. for fairly large chips. If you design a few smaller chiplets that you can mix n match to cover a wide range of products you may come out ahead - and some of the chiplets may be usable across multiple CPU/GPU core generations (i.e. memory & PCIe controllers) for further savings.

Sure chiplets are ALSO good if your yields are a problem, but that's mostly taken care of by doing stuff like Apple does where they sell Macs with parts that have bad CPU or GPU cores as a somewhat lower spec version - disabling additional cores so even if there's one only bad CPU core you might sell it as a part missing 4 CPU and 4 GPU cores. That's valuable too - if AMD is in the early stages of design for Zen 7 right now they may be designing it for TSMC's N2 but today they don't know what those yields are going to look like.

OK sure AMD is not designing for the bleeding edge like Apple - at least not today - they are taking TSMC's processes once they are fully mature so they've kind of been assuming "we won't move to this new process until the yields are good" but that's a risk. If the yields don't get good with N2 then they can't sell Zen 7, instead they'd have to port it to N2e or whatever fixes N2's problems, meaning the launch is delayed by many months and Intel benefits. If Intel gets their act together and starts executing on their processes like they claim they will, AMD is going to need to go more bleeding edge with TSMC to stay ahead or possibility merely to keep up.

Yes, that's understandable. Though I would argue designing monolithic for mobile makes more sense because the chiplet approach to add cores like amd is useless. If you need that kind of mobile power you're really better off getting a slightly larger semi mobile form factor like a nuc or a semi mini pc in a small but larger case.

There's a lot of bleeding edge vs design practice that comes into play. It isn't quite as simple as you put it. Though I'll admit that I'm extrapolating from non x86 design. Personally I don't see a point in either intel or amd chasing after apple. apple is apple and their consumers who are die hards are going to be hard to convince to switch to a windows based platform. they're a lost cause. And those who missed windows on mac due to apple ditching intel have since dropped apple for a windows platform. Could apple's choice every come back to haunt them? sure. Just like how intel's choices came back to haunt them. you won't know until you know. And at that point it's a "oh sh** we fu**** up".

 

[A///]

I wasn't commenting on large monolithic dies. I was commenting supposedly ~250mm2 AMD chip against Intel Meteor Lake - and AMD have zero reason to go costlier and worse chiplet designs for their mainstream APU as their manufacturer can produce 250mm2 dies with decent yields.

i think we're facing a communications mistake here. when I saw your post earlier yesterday i thought you meant arrow lake and beyond on desktop amd, not the apu or mobile platforms. as of now my confidence in pat "gunslinger" gelsinger is high but my confidence in intel shipping anything they promise in volume is very low. I don't have high hopes for meteor lake especially after all the back and forth bs I've seen about it and arrow lake, mobile lunar lake and desktop panther lake may as well be a wish you make skipping rocks at the local lake.

Point is and borrowing from you here, if amd is pushed to doing extremes it will take a financial toll on them if they're caught off guard by a rapid intel comeback and intel is screwed if they have a repeat of the last few years. not the manufacturing problem but ending up matching amd in performance. even if they get their power usage under control amd could possibly come in cheaper but I don't think that variance will be as larger due to tsmc increasing their costs. Intel has more on its plate with getting ifs off the ground.

as a consumer we want both to succeed otherwise intel or amd will power on and then after a few years begin to stagnate. Imo AMD were caught off guard by just how equal rpl was in performance, insane power usage put aside it's a compelling series of processors. Zen 5 needs to come roaring to life to put a gap between amd and intel for what they have now and whatever crap they release on desktop over the next 2 years.

 

[A///]

I believe Intel would still be selling RPL chips as the 'cost effective' solution, and keep MTL focused in the high end. And isn't mobile a large industry with decently high margins for Intel too?

on mobile for the 1st part?

yes it is. amd laptops are hard to come by, even buying them as a business. idk if it's a supply issue or a demand issue. far easier to get an intel laptop through volume sales.

 

[A///]

Zen 5 leek
uOP cache removed, L1i x2 the size with relatively little latency increase (magik)
Dual 6 wide decode clusters , like gracemont but with auto prediction of critical branches
8 wide rename/dispatch
~480 entry ROB, Reg files all see increases
load/store system substantially improved
L2 goes to 1.5 MB per core (private) with marginally higher latency
L3 goes to 3 MB per core and no longer a victim cache
+15-25% IPC gains
posting the YT video rn guys







...in all seriousness though where are all the Zen 5 leaks? MTL is supposed to (with good volume lol) release Q1 2024, and that's around when leakers are claiming Zen 5 releases as well

You can't out nostra the original nostra.

It was 1h 2024, not q1. Zen 4 came out in september last year and if zen 5 comes out in june 2024 it'll be about or a little over 18 months. Right on the cadence mike peppermaster talked about 2-3 years ago.

 

[DrMrLordX]

...in all seriousness though where are all the Zen 5 leaks? MTL is supposed to (with good volume lol) release Q1 2024, and that's around when leakers are claiming Zen 5 releases as well

I would be really, really surprised if Zen5 came out in any form other than EPYC to select customers by Q1 2024. Keep your eyes open for word on when Zen5/Turin starts sampling. Consumer parts could easily be July 2024 or later depending on how much of a pessimist you are.

 

[Abwx]

You can't out nostra the original nostra.

mike peppermaster talked about 2-3 years ago.

We all hope that Mark Papermaster cooked a spicy CPU but for the time any info should be taken with a pile of salt...

 

[A///]

We all hope that Mark Papermaster cooked a spicy CPU but for the time any info should be taken with a pile of salt...

while amd has improved on not saying much bull over the years and hoping i got the person right, the corpo talk was zen 4 set the platform even with its current aib based bugs that zen 5 will bring the performance. whatever that means. intel should be treated with the same healthy skepticism. nvidia too but jh is too busy with his fist around nvidia addicts nads squeezing them for more money.

 

[Geddagod]

on mobile for the 1st part?

yes it is. amd laptops are hard to come by, even buying them as a business. idk if it's a supply issue or a demand issue. far easier to get an intel laptop through volume sales.

I believe RPL will be selling in the majority of systems while MTL focuses on the high end thin and lights, where things get a bit pricier.
I have an Intel laptop rn because there literally was no good selection of AMD based models at my local best buy. Idk why AMD isn't focusing on mobile more, because the margins should still be higher than desktop no?

 

[A///]

I believe RPL will be selling in the majority of systems while MTL focuses on the high end thin and lights, where things get a bit pricier.
I have an Intel laptop rn because there literally was no good selection of AMD based models at my local best buy. Idk why AMD isn't focusing on mobile more, because the margins should still be higher than desktop no?

Supply? I'm taking a stab in the dark but they might not be able to guarantee enough raw materials to their partners. if I'm at a store that sells laptops and have 30 models on display maybe 7-10 will be amd.

 

[Abwx]

while amd has improved on not saying much bull over the years and hoping i got the person right, the corpo talk was zen 4 set the platform even with its current aib based bugs that zen 5 will bring the performance. whatever that means. intel should be treated with the same healthy skepticism. nvidia too but jh is too busy with his fist around nvidia addicts nads squeezing them for more money.

If anything Mike Clark hinted in an interview that Zen 5 was a significantly bigger step than Zen 4, wich he admitted behind the lines that it was an improved Zen 3 and not as impressive as the following design.

 

[Exist50]

Idk why AMD isn't focusing on mobile more, because the margins should still be higher than desktop no?

The flagship mobile platforms tend to be heavily co-designed with the SoC vendor. Intel has very long-established relationships and capabilities there. Seems like AMD is still building up their own. Should gradually improve over time.

 

[A///]

If anything Mike Clarck hinted in an interview that Zen 5 was a significantly bigger step than Zen 4, wich he admitted behind the lines that it was an improved Zen 3 and not as impressive as the following design.

yesssssssssssss that one, not papermaster. that is the interview I was referring to.

AnandTech Interviews Mike Clark, AMD’s Chief Architect of Zen

www.anandtech.com

 

[Geddagod]

Anyway back on topic, while my 'leek' was a complete troll lol, I am genuinely curious about the use of dual decode clusters in the front end.
And I'm certainly no CPU architect, so obviously there is a good chance I'm wrong, but I would love to hear your thoughts...
AMD was at least investigating the topic with these two patents, though I want to make clear I'm not claiming that there's even a good chance or anything these will end up in Zen 5, just that it's an interesting possibility.
I think one of the patents has 2 microOP caches while another has 2 decoders whose instructions are then re-ordered.
I'm very curious about the fate of the uOP cache. On a Zen 4 die shot, it looks like it takes up a marginal amount of die space, and the performance benefits of it in a core designed with it in mind seems to be ~10% best case, and in many cases being less than that, from chips and cheese testing.
Maybe we just don't see a very large increase in uOP cache with Zen 5?
With increased decode width, you don't need as an large uOP cache too.