Discussion Zen 7 speculation thread

[gdansk]

LNC wasn't a bad tock either. That's just how fruit are picked.

LNC is 7nm to 3nm (tick tick tock?), with a small improvement bundled with a frequency regression.
But yes, in a parallel universe where LNC was on the same major node as its predecessor it would be expected to have smaller gains as returns diminish.

 

[Geddagod]

it seems the balance mantra is still king and they aren't in a rush to take all the sugar many other cores have ( ROB/reg file size ).

They massively increased that with Zen 5.

Capacity increases vs Last Gen:	Zen 5	Zen 3
ROB capacity	+40%	+14%
uOP cache	-11% *	+100%
Rename Width	+33%	+0%
L2 BTB	+14%	-7%
L1 BTB	+967%	+100%
Store Buffer capacity	+63%	+33%
Load Buffer capacity	+130%	+0%
Total scheduler entries	+63%	+25%
INT reg file	+7%	+7%
FP reg file	+100%	+0%

In comparison to other cores:

---	LNC vs Zen 5	RWC vs Zen 4	SNC vs Zen 3
ROB capacity	+29%	+60%	+38%
uOP cache	-15%	-40%	-44%
Rename Width	+0%	+0%	-20%
L2 BTB	+50%	+61%	-23%
L1 BTB	-98%	-92%	-75%
Store Buffer capacity	+15%	+78%	+13%
Load Buffer capacity	-6%	+118%	+78%
Total scheduler entries	+22%	+28%	+0%
INT reg file	+17%	+25%	+46%
FP reg file	-1%	+73%	+40%

It's not a good look for Zen 5 tbh. Even acknowledging diminishing returns, in many areas the Zen 5's capacity increases are massively, as in like 2x or more, larger than Zen 3's capacity increases, for an IPC uplift that's like half as good.

im guessing they would have targeted higher clocks on 4N and some larger internal structures and then Zen5 would have been more inline with typical.

The frequency thing is actually interesting. So Zen 5 has more metal layers, a slightly longer pipeline, and a better node... for no Fmax improvement.

here is hoping for Zen6 to spend 2 gens of sugar at once + good clock boost to boot.

Agreed lol.

But yes, in a parallel universe where LNC was on the same major node as its predecessor it would be expected to have smaller gains as returns diminish.

Both LNC and Zen 5 had organizational changes that should have helped scaling.
LNC at least has the excuse of not blowing up structure sizes like Intel used to do with tocks, to explain its mediocre perf uplift. Idk what Zen 5 has.
And neither AMD not Intel has the best performing cores, or at least not the best performing by any meaningful margin that allows people to handwave away the massive power gap between them and Apple. Clearly there is room to grow for both perf and power improvements.

 

[Win2012R2]

or at least not the best performing by any meaningful margin that allows people to handwave away the massive power gap between them and Apple

This is literally comparing Oranges and Apples.

Does Apple sell x86-64 compatible 64+ core server chips?
Does either of AMD or Intel sell into mobile phone market?

Only real ARM competitor to AMD/Intel are the hyperscalers who use licensed ARM server chip designs, which got nothing to do with Apple and hyperscalers would do it anyway to ensure they secure discounts on x86 stuff.

 

[Geddagod]

This is literally comparing Oranges and Apples.

Does Apple sell x86-64 compatible 64+ core server chips?
Does either of AMD or Intel sell into mobile phone market?

Only real ARM competitor to AMD/Intel are the hyperscalers who use licensed ARM server chip designs, which got nothing to do with Apple and hyperscalers would do it anyway to ensure they secure discounts on x86 stuff.

What does this have to do with what I was talking about?

 

[511]

every x86 is going clustered decode i am wondering what would be the final decoder scheme for Z6/Z7/PNC/ARW/GE

 

[Win2012R2]

What does this have to do with what I was talking about?

And neither AMD not Intel has the best performing cores, or at least not the best performing by any meaningful margin that allows people to handwave away the massive power gap between them and Apple

This bit - who are those people, and I mean actual buyers of those chips? I've got Apple phone/tables, but I won't even consider for a server, even if they had that offering, neither will I get their laptop.

These are different markets, especially given Apple's vertical integration and insistence on charging ridiculous margins

 

[511]

not to mention they are missing all the server features that Intel and AMD have

 

[Win2012R2]

not to mention they are missing all the server features that Intel and AMD have

And what will Dell, HP, Lenovo, SuperMicro do without AMD/Intel chips? There is whole big ecosystem, only reason ARM got anywhere in server space is because of hyperscalers.

 

[itsmydamnation]

They massively increased that with Zen 5.

Capacity increases vs Last Gen:	Zen 5	Zen 3
ROB capacity	+40%	+14%
uOP cache	-11% *	+100%
Rename Width	+33%	+0%
L2 BTB	+14%	-7%
L1 BTB	+967%	+100%
Store Buffer capacity	+63%	+33%
Load Buffer capacity	+130%	+0%
Total scheduler entries	+63%	+25%
INT reg file	+7%	+7%
FP reg file	+100%	+0%

In comparison to other cores:

---	LNC vs Zen 5	RWC vs Zen 4	SNC vs Zen 3
ROB capacity	+29%	+60%	+38%
uOP cache	-15%	-40%	-44%
Rename Width	+0%	+0%	-20%
L2 BTB	+50%	+61%	-23%
L1 BTB	-98%	-92%	-75%
Store Buffer capacity	+15%	+78%	+13%
Load Buffer capacity	-6%	+118%	+78%
Total scheduler entries	+22%	+28%	+0%
INT reg file	+17%	+25%	+46%
FP reg file	-1%	+73%	+40%

It's not a good look for Zen 5 tbh. Even acknowledging diminishing returns, in many areas the Zen 5's capacity increases are massively, as in like 2x or more, larger than Zen 3's capacity increases, for an IPC uplift that's like half as good.

The frequency thing is actually interesting. So Zen 5 has more metal layers, a slightly longer pipeline, and a better node... for no Fmax improvement.

Agreed lol.

Both LNC and Zen 5 had organizational changes that should have helped scaling.
LNC at least has the excuse of not blowing up structure sizes like Intel used to do with tocks, to explain its mediocre perf uplift. Idk what Zen 5 has.
And neither AMD not Intel has the best performing cores, or at least not the best performing by any meaningful margin that allows people to handwave away the massive power gap between them and Apple. Clearly there is room to grow for both perf and power improvements.

Sigh no offence but you missed the point and two you had to resort to percentages because when you look at absolutes it doesn't fit your logic.

The irpf is only 240 entries
The rob is only 448 entries

Both of these are way lower then everyone else. For example even golden cove has more of both and less execution width while also sharing the role of ports. Let's not even bother go to apple or Qualcomm.

I don't have time to go blow by blow and I'm on a phone , but try to understand a CPU critical path for latency/clocks and you consistently find amd has been very sinky in those regards. Thus my comment of big bones but not big muscles.

 

[Geddagod]

This bit - who are those people, and I mean actual buyers of those chips? I've got Apple phone/tables, but I won't even consider for a server, even if they had that offering, neither will I get their laptop.

These are different markets, especially given Apple's vertical integration and insistence on charging ridiculous margins

I'm talking about the people looking at the different architectures of those cores lol. What I was doing for the past couple messages.

Sigh no offence but you missed the point and two you had to resort to percentages because when you look at absolutes it doesn't fit your logic.

The irpf is only 240 entries
The rob is only 448 entries

Both of these are way lower then everyone else. For example even golden cove has more of both and less execution width while also sharing the role of ports. Let's not even bother go to apple or Qualcomm.

That's the entire point of the percentages compared to Intel's cores.
Look at LNC vs Zen 5, and then look at RWC vs Zen 4 and SNC vs Zen 3. The percentages in almost every category has Intel having the lowest or the 2nd lowest lead with LNC vs Zen 5 than previous architectures.
Meaning that AMD actually did comparatively widen the core and core capacities, not just compared to previous Zen architectures, but also compared to the competition (Intel). And ofc Intel also being much better about gaining IPC without increasing core structure capacities dramatically also played a role.

, but try to understand a CPU critical path for latency/clocks and you consistently find amd has been very sinky in those regards

Sinky?

 

[inquiss]

I'm talking about the people looking at the different architectures of those cores lol. What I was doing for the past couple messages.

That's the entire point of the percentages compared to Intel's cores.
Look at LNC vs Zen 5, and then look at RWC vs Zen 4 and SNC vs Zen 3. The percentages in almost every category has Intel having the lowest or the 2nd lowest lead with LNC vs Zen 5 than previous architectures.
Meaning that AMD actually did comparatively widen the core and core capacities, not just compared to previous Zen architectures, but also compared to the competition (Intel). And ofc Intel also being much better about gaining IPC without increasing core structure capacities dramatically also played a role.

Sinky?

I expect "stingy"

 

[adroc_thurston]

but also compared to the competition (Intel)

AMD's competition (aa64 galley) ships far, fat fatter cores.
Zen5 is the size of Skymont outside of SIMD bits.

 

[Geddagod]

I expect "stingy"

Ah lol. I should have made that connection. My bad.
In that case

but try to understand a CPU critical path for latency/clocks and you consistently find amd has been very sinky in those regards.

AMD might still be stingy in those regards, but clearly they aren't nearly as stingy as before.

AMD's competition (aa64 galley) ships far, fat fatter cores.

They also have much better IPC.

Zen5 is the size of Skymont outside of SIMD bits.

It has a much, much bigger BPU and wider in a couple places as well.
Interestingly enough though, with what low quality die shots we have, Zen 5C without the FPU, without the CPL/clock, and the L2 cache blocks... is still smaller edit: larger than Skymont in area. Hope for Unified Core.

 

[adroc_thurston]

They also have much better IPC.

And far lower clk targets. Tradeoffs!

It has a much, much bigger BPU

That's the AMD specialty. They just love extra chungus BTBs and can't help themselves.

 

[Geddagod]

That's the AMD specialty. They just love extra chungus BTBs and can't help themselves.

I hope Huang follows up on what the deal with Zen 5's L0 BTB is.

 

[poke01]

I hope Huang follows up on what the deal with Zen 5's L0 BTB is.
View attachment 127866

Probably won’t till zen 6 review

 

[fastandfurious6]

summarily the consensus is that zen5 is very beefy but has big bottlenecks right? which can be fixed at next iteration

 

[511]

and than Z6 will have another bottle necks designs are tradeoffs depending on their targets.

 

[poke01]

and than Z6 will have another bottle necks designs are tradeoffs depending on their targets.

every design has trade offs but some tradeoffs suck a lot more than others

 

[511]

every design has trade offs but some tradeoffs suck a lot more than others

LNC SMT Removal 🤣. It would have been
like +10% MT over current ARL.

 

[Geddagod]

summarily the consensus is that zen5 is very beefy but has big bottlenecks right? which can be fixed at next iteration

It's still not as beefy as other designs, I agree with the general sentiment there. But the large gap that AMD used to have vs Intel in achieving similar IPC with much smaller core structures seem to have shrunk.
TBF, I think AMD missed what they wanted to hit with Zen 5. The area and perf/ipc topic has been well discussed by this point, but I wonder if there is anyone testing Zen 3 vs Zen 2 core power curves.

 

[fastandfurious6]

TBF, I think AMD missed what they wanted to hit with Zen 5

ZEN 50% ..... RIP 😭

 

[poke01]

This bit - who are those people, and I mean actual buyers of those chips? I've got Apple phone/tables, but I won't even consider for a server, even if they had that offering, neither will I get their laptop.

I mean Apple's M chips don't even have much PCIe lanes for a full length x16 PCIe. You need the Ultra line to even get one, these are so tailor made for Apple laptops. So forget servers lol

That being even if you don't see Apple as competition in the laptop space they along with Qualcomm are competition for AMD and AMD does see these two as competition in the premium laptop space.

 

[Win2012R2]

I'm talking about the people looking at the different architectures of those cores lol.

But chips are designed for different purposes - you are not comparing like for like, say, why not compare power efficiency of M4 vs some random microcontroller?

 

[poke01]

But chips are designed for different purposes - you are not comparing like for like, say, why not compare power efficiency of M4 vs some random microcontroller?

Best course of action would be to make a separate thread for comparisons across different ISAs, I think we have one those already though.