Question Zen 6 Speculation Thread

[adroc_thurston]

I guess depending on what clocks SP7 Venice hits that is 100% possible...

You know well enough how it clocks.

just that outside of HPC, 16 channels of MRDIMM with 1.6TB/s of memory bandwidth would be a struggle to use...

Naa, 256c socket in a world of oversubscribed favelas? 1.6TB/s actually feels appropriate, like a wee bit over 6GB/s a core.

 

[igor_kavinski]

1.6TB/s actually feels appropriate, like a wee bit over 6GB/s a core.

Howz the latency?

 

[adroc_thurston]

Howz the latency?

Normal EPYC latency.
But it's a manycore system so it's not like you'll be able to hide it all that effectively anyway.

 

[Tuna-Fish]

One thing that would make it a killer product is if TSMC and AMD mastered Wafer on Wafer packaging, which would cut cost and improve throughput on packaging dramatically.

Cutting costs does not equate to lower prices to us, though.

One thing that's obvious to everyone on the market is that the 9800X3D was mispriced. It would probably not have sold a single unit less so far if it was $499.

To me, the 3D core leak seems like AMD chasing higher ASP by delivering as good a desktop high-end part as they can, costs be damned. 224MB of L3 directly attached to 32 cores lets them sell a no-compromises part that is better than the current favorite gaming chips in everything, and can justify a much higher ASP because of the core count.

 

[Josh128]

I'm interested to see if they will produce an 8 core (recovery bin of 12) that keeps the full 48MB of L3, and if so, how it compares to the 9700x. I suspect that the combination of 50% more cache and better throughout will make a notable improvement in games.

Again, look to historical perf jumps with node advancements. >nT benefits the most. Look at 5800X vs 7700X (+26% R23 nT), then look at 7980X vs 5995WX (+51% R23 nT). Uplift from node advancement increases with # of cores.

Zen 5 to Zen 6 is a two node advancement though, compared to the example above, so its not naive to expect more. 8 core Zen 5 to 8 core Zen 6 could easily hit +30%nT, and that margin will only increase going up the core count stack until you start running into more uncore losses.

 

[511]

Again, look to historical perf jumps with node advancements. >nT benefits the most. Look at 5800X vs 7700X (+26% R23 nT), then look at 7980X vs 5995WX (+51% R23 nT). Uplift from node advancement increases with # of cores.

Zen 5 to Zen 6 is a two node advancement though, compared to the example above, so its not naive to expect more. 8 core Zen 5 to 8 core Zen 6 could easily hit +30%nT, and that margin will only increase going up the core count stack until you start running into more uncore losses.

the biggest improvements are at lower power levels at higher power levels the gains are less.

 

[Abwx]

the biggest improvements are at lower power levels at higher power levels the gains are less.

Still 32% better perfs for 7950X vs 5950X when both are set at 142W PPT, that s considerable because at same throughput a 65W PPT 7950X match a 142W PPT 5950X,
that s 2.18x the perf/watt, and most of it came from the process, the IPC can be counted as 1.35x for the total, so that s 2.28/1.35 = 1.62x just for the process.

 

[MadRat]

If AMD can cut production costs then its good for stockholders.

The consumer price is a reflection of market forces not production costs.

 

[511]

If AMD can cut production costs then its good for stockholders.

The consumer price is a reflection of market forces not production costs.

Yeah forget reducing production cost it's impossible especially with meager node gains and high cost.

Still 32% better perfs for 7950X vs 5950X when both are set at 142W PPT, that s considerable because at same throughput a 65W PPT 7950X match a 142W PPT 5950X,
that s 2.18x the perf/watt, and most of it came from the process, the IPC can be counted as 1.35x for the total, so that s 2.28/1.35 = 1.62x just for the process.

Process has always been a big factor as for IPC wasn't it just 13% for Zen3 to Zen4.

 

[Abwx]

Process has always been a big factor as for IPC wasn't it just 13% for Zen3 to Zen4.

IPC count a lot when we re talking of same throughput, with 13% better IPC you get 35% better perf/watt at same throughput because you can downclock by 1/1.13 at same perfs.

 

[511]

IPC count a lot when we re talking of same throughput, with 13% better IPC you get 35% better perf/watt at same throughput because you can downclock by 1/1.13 at same perfs.

Even without IPC just a shrink provides same performance at 25-30% lower power IPC is just cherry on top of that.

 

[Abwx]

Even without IPC just a shrink provides same performance at 25-30% lower power IPC is just cherry on top of that.

I gave the numbers, 1.35x for IPC and 1.62x for the process, that s respectively 26% lower power due to IPC alone and on top 38% less power for the process alone.

 

[511]

I gave the numbers, 1.35x for IPC and 1.62x for the process, that s respectively 26% lower power due to IPC alone and on top 38% less power for the process alone.

Ipc or rather perf/clock is measured at the same clock lol

 

[Abwx]

Ipc or rather perf/clock is measured at the same clock lol

Not always, AMD published a comparison of a stock 5950X vs a 7950X@65W, guess that it was hardly the same clocks.

 

[Abwx]

the biggest improvements are at lower power levels at higher power levels the gains are less.

That s not true, at 142W the difference is 37% between the 5950X and the 7950X, and it s still 35% at 230W.

AMD Ryzen 7000 im Test: So schnell sind 7950X und 7700X: Leistungsaufnahme und Effizienz

Ryzen 9 7950X & Ryzen 7 7700X im Test: Leistungsaufnahme und Effizienz / TDP und PPT steigen deutlich

www.computerbase.de

 

[MS_AT]

That s not true, at 142W the difference is 37% between the 5950X and the 7950X, and it s still 35% at 230W.

It is literally true, you even admit it. 35% is less than 37%... the difference is not big but it does exist...

 

[511]

That s not true, at 142W the difference is 37% between the 5950X and the 7950X, and it s still 35% at 230W.

it proves my point at
65W -> 75% gains
at 105W -> 37%

 

[Josh128]

it proves my point at
65W -> 75% gains
at 105W -> 37%

Thats because the 7950X MT clocks much higher at 65W power level than the 5950X does. Gains from IPC are going to be ~8-15% depending on the workload. R23 IPC increase was only about 7% from Zen 3 to Zen 4. The rest of the ~40% gains in default TDP nT score were entirely from increased all core clock speeds when comparing the two SKUs.

 

[Abwx]

It is literally true, you even admit it. 35% is less than 37%... the difference is not big but it does exist...

The 2% differerence is negligible, beside the 5950X wasnt designed for power as high as 230W.

 

[Joe NYC]

Yes in Asia.
Anywhere, really, 9800X3D rules the world with an iron fist.

Asia, especially China were Intel's best markets, which means, they will likely be last to move to AMD. But gamers are inquisitive people they will find their way to AMD

No, WoW is faster but not really cheaper and you have other caveats in play.

Cost of constructing carrier wafer (or two) individually placing the die there etc, having to deal with each individual die that has previously been cut, it seems like an order of magnitude more steps to accomplish than WoW.

I think that High Yield is probably right, and AMD / TSMC are still working on the WoW packaging. Because if they had the answer, we would be seeing more proliferation of 3D stacking in AMD roadmap.

Maybe by Zen 7, the companies estimate they will have WoW working, just from the fact that one (or more) of Zen 7 will be a 2 die implementation.

Straight to the wet nerd dreams he goes.
Pipe down.

Mainstream market, served by Kraken (and successor) is area where AMD could use some differentiating features.

AMD is already doing chiplets for premium market, but the lower to mid market, served by a single monolith may not have enough to win over the market could use something that would make them more desirable.

Intel is already stuck with their expensive chiplet overhead, which is probably higher than cost of stacking if WoW stacking was available... So AMD would not be pricing itself out of the competition...

 

[adroc_thurston]

Asia, especially China were Intel's best markets, which means, they will likely be last to move to AMD.

Cracked some time ago.

Cost of constructing carrier wafer (or two) individually placing the die there etc, having to deal with each individual die that has previously been cut, it seems like an order of magnitude more steps to accomplish than WoW.

yeah but with WoW you're doubling the raw node d0 since you can't pre-sort dies before packaging.

I think that High Yield is probably right, and AMD / TSMC are still working on the WoW packaging. Because if they had the answer, we would be seeing more proliferation of 3D stacking in AMD roadmap.

again, scribe lines match. please.

Mainstream market, served by Kraken (and successor) is area where AMD could use some differentiating features.

Mainstream is won on cost, not weird nerd wank materiel.
Pipe down already.

served by a single monolith may not have enough to win over the market could use something that would make them more desirable.

AMD already wins on cost vs ARL/PTL something fierce. They do not need to do anything.

 

[OneEng2]

You could say that the higher clocks coupled with more cache easily outperform 128 cores plus leaving some room to challenge and beat anything Intel may have up their sleeves, instead of making everyone suspect that you didn't get to spend enough time with your dad in your childhood.

LOL. I was literally already considering a very similar remark.

@adroc_thurston,

The lack of any intelligent response from you is a testament to your maturity level. Instead of tossing adolescent personal insults at your betters, go find your binkie and suck on it until the urge to be a baby dissipates.

@ the mods. Before you criticize me for breaking forum rules, note: I wouldn't need to had you dispelled this infantile behavior yourselves.

This forum has been a model of decorum for the most part and I honestly don't wish it to decay into personal insults as it is a great source of intelligent discussion and a resource for lots of people across the web.


Two things wrong here.
1) No member callouts. Period.
2) No moderator callouts either. You do not get to tell us how to run the forum
3) There's only one way to take care of this. Make a Reported Post on the offending post.
If you were up on our rules you would know this.

AnandTech Forum Guidelines

AnandTech Forum Guidelines Introduction In an effort to make our management of the forums as transparent as possible, we have refined our member and posting guidelines and we have created a set of moderator guidelines. It is our hope that we can better satisfy members by being upfront and...

forums.anandtech.com

Rule # 1 and 13.


esquared
Anandtech Forum Director

 

[Joe NYC]

yeah but with WoW you're doubling the raw node d0 since you can't pre-sort dies before packaging.

But mostly SRAM die has enough redundancy that yield is close to 100%. So it would be only function of yield of logic die, while throwing out cheap(er) SRAM die when there is a defect in logic die.

I also think that there has to be more uniformity when using the entire wafers vs. positioning every single die one by one on a carrier wafer, which would lead to increased yield from the packaging, fewer losses to failed hybrid bond connection.

again, scribe lines match. please.

This would be quite a technological accomplishment to get WoW packaging into high volume manufacturing. What's stopping AMD and TSMC from taking this victory lap?

Mainstream is won on cost, not weird nerd wank materiel.
Pipe down already.

AMD already wins on cost vs ARL/PTL something fierce. They do not need to do anything.

I think that's right. Kraken is (likely much) better on cost, but "meh" otherwise, on features and performance.

I think the "meh" part needs some improvement.

 

[Joe NYC]

LOL. I was literally already considering a very similar remark.

@adroc_thurston,

The lack of any intelligent response from you is a testament to your maturity level. Instead of tossing adolescent personal insults at your betters, go find your binkie and suck on it until the urge to be a baby dissipates.

@ the mods. Before you criticize me for breaking forum rules, note: I wouldn't need to had you dispelled this infantile behavior yourselves.

This forum has been a model of decorum for the most part and I honestly don't wish it to decay into personal insults as it is a great source of intelligent discussion and a resource for lots of people across the web.

People on this forum are generally not children, so it does not have to have kindergarten rules.

 

[adroc_thurston]

But mostly SRAM die has enough redundancy that yield is close to 100%.

It ain't a naked SRAM die, it features a full power delivery and signal network.
You indeed have to actually yield those.

This would be quite a technological accomplishment to get WoW packaging into high volume manufacturing.

WoW SoIC shipped years ago with Graphcore BOW.

What's stopping AMD and TSMC from taking this victory lap?

Graphcore did that years before AMD.

I think that's right. Kraken is (likely much) better on cost, but "meh" otherwise, on features and performance.

I think the "meh" part needs some improvement.

the whole point of mainstream is you deliver comp perf while being lower on cost so your platform is more palatable to OEMs.