Question Zen 6 Speculation Thread

[jpiniero]

Wut. What does this have to do with Zen 6?

 

[AMDK11]

Sandy over Nehalem was childs play my man. Conroe over Netburst was like +85% IPC gain. By far the biggest jump in the last 25+ years.

Yonah(Core(1)M) - Conroe(Core 2) IPC +15-25%.
Tom’s Hardware: SYSmark +~17.7%.

Netbrust(Pentium 4) - Pentium III(Tualtin) IPC+~15-40%
Netbrust(Pentium 4) - Banias(Pentium M) IPC +~20-50%
Netbrust(Pentium 4) - Yonah(Core(1)M) IPC +~40-60%

There's no real breakthrough in the Conroe vs. Netbrust comparison, as the latter had a pathetic IPC even compared to the Pentium III.

Conroe, on the other hand, gained 15-25% IPC compared to its predecessor (Yonah and K8 Athlon64). On the TH, they achieved an average increase of 17.7%. This is a typical generational leap. It only seems significant compared to Netbrust.

 

[Josh128]

With 50% more cores and 1.7-1.75x expected better MT it s definitly no, it wont match the Zen/Zen+ to Zen 2 transition.

AMD Ryzen 9 3950X im Test: Benchmarks in Anwendungen und Spielen

AMD Ryzen 9 3950X im Test: Benchmarks in Anwendungen und Spielen / Benchmarks in Anwendungen / Benchmarks in Spielen (Full HD und UHD)

www.computerbase.de

Core for core MT gains, Zen + to Zen 2 was + 28% from 2950X to 3950X, and only 20% from 2700X to 3700X in Cinebench R20, while going Zen 3 to Zen 4 you got much more. But sure, if you want to compare a 2700X to a 3950X you are getting >100%MT there. I dont think we'll get 100% for 9950X to 11950X, but >75% in some highly threaded apps is certainly plausible. Core for Core MT, I think it will be the biggest gains in Zen history, due to the 2 node jump (if thats true).

 

[DrMrLordX]

Wut. What does this have to do with Zen 6?

Nothing, it's the usual conversation creep that comes from people derailing threads during a leaks/news shortage. The mods will clean it up if anyone complains enough.

 

[DavidC1]

Netbrust(Pentium 4) - Pentium III(Tualtin) IPC+~15-40%
Netbrust(Pentium 4) - Banias(Pentium M) IPC +~20-50%
Netbrust(Pentium 4) - Yonah(Core(1)M) IPC +~40-60%

Banias was much more significant than that. It was 30% over Coppermine P3, so it would be probably ~20% against Tualatin. Yonah got barely anything, and it was mostly about making a proper dual core. Core-wise, Yonah is a tick. 5% average at most. You can go look at the uarch. It changed almost nothing in Yonah.

Also note roughly where people think positively about architectures is relative not about it's FP capabilities, not it's multi-core, but how it performed on scalar Integer. Because that's what really matters on a GP architecture.

Core for core MT gains, Zen + to Zen 2 was + 28% from 2950X to 3950X, and only 20% from 2700X to 3700X in Cinebench R20, while going Zen 3 to Zen 4 you got much more. But sure, if you want to compare a 2700X to a 3950X you are getting >100%MT there. I dont think we'll get 100% for 9950X to 11950X, but >75% in some highly threaded apps is certainly plausible. Core for Core MT, I think it will be the biggest gains in Zen history, due to the 2 node jump (if thats true).

You are thinking 2 node jump in the FinFET era. The GAAFET era will bring a lot less gains. Density gains for example plummets from 100% to just 30%.

The gains came much easier during the planer transistor era. Hype goes up with every new transistor, but gains go down and is more expensive too.


Sandy Bridge was great because it did all that efficiently and the Turbo mode was actually beneficial. It really benefitted the U parts because until then we were stuck with measly 1.33GHz processors, and Sandy Bridge boosted that to 1.86GHz. Along with the new uarch, we got a 30% boost in performance. It was what really enabled the Ultrabook era to make sense, cause previously they were expensive Atoms. It makes you realize how even in the peak days Intel's execution paled to what Apple were able to do, because it needed very fancy schmancy Turbo to have any sort of useful performance while ARM parts did it without.

 

[Chicken76]

Conroe(2006) -> ivy bridge (2012) = intels golden era, banger after banger

Then completely fell off the cliff 🤣

We could include Haswell in that golden period. Jim Keller said he liked Haswell.

 

[adroc_thurston]

We could include Haswell in that golden period. Jim Keller said he liked Haswell.

Yeah, Haswell-U was a market-breaker.

 

[Josh128]

You are thinking 2 node jump in the FinFET era. The GAAFET era will bring a lot less gains. Density gains for example plummets from 100% to just 30%.

No, Im thinking 2 node jump which supposedly enables ~6.5GHz max core freq. If thats true, its the same step change as Zen 3 to Zen 4, which also enables much higher all core frequencies. Power may have to increase yet again, but a 30%-45% MT jump from freq increase, +10% from IPC, and another 35% from 50% more cores would get you to the 75%+ MT gains scenario.

 

[fastandfurious6]

We could include Haswell in that golden period. Jim Keller said he liked Haswell.

Absolutely not, haswell was the beginning of the end

1. Barely any improvement https://www.notebookcheck.net/Review-Intel-Haswell-Processors.93189.0.HTML previous gen Ivy bridge is even faster in some benches

2. Mobile market was FLOODED with the awful 15w U models designed for maximum stuttering.

Yeah, Haswell-U was a market-breaker.

Nope. All U models from haswell gen4 to gen10 were total ass 😂 awful to run. Maybe just 1 hour of extra battery life traded with permanent ultra-stutters...

The whole U 15W idea since 2013 was an absolute failure. CPUs at that time were just not fast enough to run the basic stuff - windows 7 - efficient enough. ULW only became efficient in the last few years: Lunar lake is Intel's first FAST ULW chip. all previous were ass incl meteor lake etc

 

[511]

The whole U 15W idea since 2013 was an absolute failure. CPUs at that time were just not fast enough to run the basic stuff - windows 7 - efficient enough. ULW only became efficient in the last few years: Lunar lake is Intel's first FAST ULW chip. all previous were ass incl meteor lake etc

Nice Cope there were some good chips in between them like Kaby Lake U(i5-8250U) i had one the chip was pretty efficient as long as MB Vendor set sane limits challenge impossible and MTL has battery competitive with Phoenix.
LNL is unmatched in x86 Market the dual core Skylake-U was horrible though.

 

[eek2121]

Man, you guys are bringing back memories. I was always an oddball once I had spending money.

I bought the q6600 when it first came out, the 2600k, Threadripper 1950X, and have bought every AMD 16 core processor except X3D and Zen 5 since (getting 9950X3D soon, need a better board)

I also always bought double the memory that everyone else did at any given time.

 

[511]

Man, you guys are bringing back memories. I was always an oddball once I had spending money.

I bought the q6600 when it first came out, the 2600k, Threadripper 1950X, and have bought every AMD 16 core processor except X3D and Zen 5 since (getting 9950X3D soon, need a better board)

I also always bought double the memory that everyone else did at any given time.

💲💲

 

[Doug S]

Man, you guys are bringing back memories. I was always an oddball once I had spending money.

I bought the q6600 when it first came out, the 2600k, Threadripper 1950X, and have bought every AMD 16 core processor except X3D and Zen 5 since (getting 9950X3D soon, need a better board)

I also always bought double the memory that everyone else did at any given time.

I had a Q6600 back in the day. I haven't followed the train of more and more cores though, because I don't really need them. Just built a new PC a few months ago with a 9600X, six cores is plenty for me. Got 48 GB of RAM even though I only planned to get 32 GB, thanks to a timely Amazon deal for $90.

My first PC, an Atari 800, came with 16K but I convinced my dad to get two 16K upgrade boards for it at $129.99. Each! In today's dollars that's nearly $1000 for the 32K upgrade.

 

[Thunder 57]

Man, you guys are bringing back memories. I was always an oddball once I had spending money.

I bought the q6600 when it first came out, the 2600k, Threadripper 1950X, and have bought every AMD 16 core processor except X3D and Zen 5 since (getting 9950X3D soon, need a better board)

I also always bought double the memory that everyone else did at any given time.

I bought 16GB in oh say 2012/2013? More than I needed. I stuck with it (DDR4 this time) through Zen+ and finally upgraded to 32GB last year or so. I was lucky enough to avoid the memory price hike back in the day.

 

[DavidC1]

Nope. All U models from haswell gen4 to gen10 were total ass 😂 awful to run. Maybe just 1 hour of extra battery life traded with permanent ultra-stutters...

The whole U 15W idea since 2013 was an absolute failure. CPUs at that time were just not fast enough to run the basic stuff - windows 7 - efficient enough. ULW only became efficient in the last few years: Lunar lake is Intel's first FAST ULW chip. all previous were ass incl meteor lake etc

Do you even internet?

I completely ignore laptops before Gen 4, because Haswell was what allowed great battery life increases. It was the only saving grace for the otherwise mediocre uarch. The U laptops got 50% battery life improvement at the same capacity.

No, Im thinking 2 node jump which supposedly enables ~6.5GHz max core freq. If thats true, its the same step change as Zen 3 to Zen 4, which also enables much higher all core frequencies. Power may have to increase yet again, but a 30%-45% MT jump from freq increase, +10% from IPC, and another 35% from 50% more cores would get you to the 75%+ MT gains scenario.

"IPC" itself increases power consumption. You are lucky if you get 10% power increase for 10% perf increase. Usually the ratio is worse such as 10% perf and 25% power increase. You'd be lucky if you get 60% improvement in embarassingly parallel code.

The 10-15% perf improvement TSMC claims is for a dinky Cortex A series CPU. At the frequencies desktop chips are operating at, you are running into circuitry, thermal and power density issues. You aren't getting anywhere near what they claim.

 

[Chicken76]

Absolutely not, haswell was the beginning of the end

1. Barely any improvement https://www.notebookcheck.net/Review-Intel-Haswell-Processors.93189.0.HTML previous gen Ivy bridge is even faster in some benches

2. Mobile market was FLOODED with the awful 15w U models designed for maximum stuttering.

Wasn't Haswell the first generation wide enough for Hyperthreading to really stand out?

 

[OneEng2]

The yields would have been fixed by that time if there are any issues.

Yea, like the yields were fixed so easily on 10nm?

I admire your optimism but BSPDN is apparently PITA to deal with.

I think that BSPDN has been around .... in the lab for quite some time. It is exactly this "PITA" you mention that has kept it out of production for so long.

It's even more PITA for cooling 🤣

This is what I get from process engineers. Evidently the routing requirements leave hot spots in the silicon that result in frequency limitations within that area due to localized heat.

Just a thought, one has to speculate about Intel shortening the pipeline of their P-cores as a pretty good thing. Keeping the same high Fmax, while also having the side effects of slight IPC increases, and lowering the need to spend a bunch of resources on the BPU, since the branch mispredict penalty will also be lower.

It's possible IMHO. If Intel is going for the "more cores" approach, then simplifying and shortening (for higher ST IPC) might make sense.

It is my personal belief that removal of SMT and not focusing on DC is a mistake though.

I think you are right. It is quite conceivable that for 8 and 6 core, AMD will continue to sell Zen 5. And limit Zen 6 to 12 and 10 core SKUs. It might be too much waste to sell 12 core CCD as 8 core SKU.

It doesn't make much sense to throw an N2 die away because it only has 8 good cores IMO.

You are thinking 2 node jump in the FinFET era. The GAAFET era will bring a lot less gains. Density gains for example plummets from 100% to just 30%.

The gains came much easier during the planer transistor era. Hype goes up with every new transistor, but gains go down and is more expensive too.

I think this is the gist of the problem. Process technology changes are simply grinding to a halt (or maybe just a trickle). The entire industry has relied on:

1) Huge transistor budget and power capabilities each gen
2) New instruction sets that greatly increase IPC

I think it is difficult today to get much from either of these ..... but I would bet more on 2 than 1 .

I had a Q6600 back in the day. I haven't followed the train of more and more cores though, because I don't really need them. Just built a new PC a few months ago with a 9600X, six cores is plenty for me. Got 48 GB of RAM even though I only planned to get 32 GB, thanks to a timely Amazon deal for $90.

My first PC, an Atari 800, came with 16K but I convinced my dad to get two 16K upgrade boards for it at $129.99. Each! In today's dollars that's nearly $1000 for the 32K upgrade.

First PC was an 8088 with a Hercules monochrome graphics card . It had a 10Mb hard drive.

 

[511]

Yea, like the yields were fixed so easily on 10nm?

10nm is an exception tbh due to its design has to be the worst node ever made economically.

 

[MS_AT]

It is my personal belief that removal of SMT and not focusing on DC is a mistake though.

I would say removing it and not having anything to compensate is a mistake. If they took the classes in magical frontends at school of hardware wizardy (I am afraid the teachers went to work at Apple so the classes are on hold), then getting rid of SMT wouldn't be so bad. Alas, they did not and are behind both Apple and AMD.

 

[Markfw]

I would say removing it and not having anything to compensate is a mistake. If they took the classes in magical frontends at school of hardware wizardy (I am afraid the teachers went to work at Apple so the classes are on hold), then getting rid of SMT wouldn't be so bad. Alas, they did not and are behind both Apple and AMD.

Why are we talking about removing SMT and 10nm, both of which have nothing to do with Zen 6 ??

 

[Joe NYC]

Why are we talking about removing SMT and 10nm, both of which have nothing to do with Zen 6 ??

One possible reason could be that Zen 6 (Venice) has it and its competition, Diamond Rapids doesn't.

 

[StefanR5R]

One possible reason could be that [...]

[...] for some, the web forum equivalent of loving to hear oneself talk takes precedence over etiquette, such as staying on topic.

Zen 6 (Venice) has it and its competition, Diamond Rapids doesn't.

While it is safe to say that Zen 6 (Venice) will have SMT, it is not a given that Diamond Rapids will lack SMT. But Diamond Rapids speculation does not belong here.

 

[adroc_thurston]

it is not a given that Diamond Rapids will lack SMT

It is.
It just doesn't have it.

 

[511]

[...] for some, the web forum equivalent of loving to hear oneself talk takes precedence over etiquette, such as staying on topic.


While it is safe to say that Zen 6 (Venice) will have SMT, it is not a given that Diamond Rapids will lack SMT. But Diamond Rapids speculation does not belong here.

Bruh their CEO announced it what more proof do you need

 

[StefanR5R]

their CEO announced it

No he didn't. However, that's off-topic here.