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

[DisEnchantment]

Speculate at will

 

[eek2121]

Just a PSA anyone that claims to know what Zen 5 pricing will be does not know what they are talking about. AMD does not decide on pricing until just before launch. Pricing is set based on existing competitor pricing, channel availability, etc.

I have no idea what the final price for the 16 core (non-v-cache part) part will be either, but I bet it will be < $999…

Hold on, you're saying that Zen 5 will outperform ARL by a wider margin than Zen 3 outperformed RKL?

Just from the memory bandwidth advantages alone it should be more competitive than that. ARL will run DDR5-6400 JEDEC and should implement DDR5 CKD (client clock driver).

Rocket Lake and Zen 3 were very close in IPC.

If Intel could have gotten Tiger Lake clocks/availability up (Rocket Lake was a backport) we would have seen something very different happen.

 

[TESKATLIPOKA]

No, he actually meant just 1 core. Did you not read his response?

So no reasoning of the type you mentioned. If he actually means MT performance is important too after all, he should say that.

He didn't mean It as you think. I will try to explain It to you.

How can a single core offer for example 30% higher performance but the rest of the cores stay the same compared to Zen4?
The only possibility is If the turboboost for a single core increased from 5.7GHz to 7.4GHz.
This is already very unrealistic, because that clock is just too high.
The other illogical thing would be to expect one core to clock that high, but using two or more would result in the same clocks as Zen4 has.

BTW @adroc_thurston already said on this forum that Zen5 won't increase clocks compared to Zen4.
Then how can a single core become faster? Only by higher IPC.

If IPC of one core increased by 30% then what do you think will happen to the rest of the cores and the performance of the whole chip?
Of course the MT performance will increase by about the same amount unless power consumption will hit the power limit resulting in lower performance because of lowered clockspeed.

And if MT performance is important, then core count matters as well. A lot. And then ARL-S will give Zen5 some real competition if Zen5 is only 16C, while ARL-S is 8P+32E. But with no official sources confirming any of this, it's all speculation at this point.

Core count is also important for MT, but It's limited by the number of threads used by soft.
I will give you a simple example, there are 2 CPUs.
CPU A: 16C32T 5.6GHz turbo, 5GHz all-core clock, + 50% IPC compared to Zen4
CPU B: 24C48T 5.6GHz turbo, 5GHz all-core clock, the same IPC as Zen4
What is the performance?
At 48 Threads they would perform the same, because CPU A can execute only 32 threads at once, but at fewer threads CPU A will be faster and at 1-16T the difference will be 50% faster.

P.S. I don't know If what @adroc_thurston said is true or not. We will see after Zen5 is out.

 

[H433x0n]

Just a PSA anyone that claims to know what Zen 5 pricing will be does not know what they are talking about. AMD does not decide on pricing until just before launch. Pricing is set based on existing competitor pricing, channel availability, etc.

I have no idea what the final price for the 16 core (non-v-cache part) part will be either, but I bet it will be < $999…

I was the one that asked that question initially and I specifically said I wouldn't hold him to it. He didn't just throw out that number out of the blue or try to claim it was an authoritative leak.

 

[Timmah!]

To be fair, it felt to me like the 999 pricing was more of an opinion or guess of @adroc_thurston, unlike the other claims. Because Zen5 is apparently so good and because Intel will be supposedly crap, etc... whatever. I think its wrong, no way will AMD bump the price so hard.
That said, i am willing to believe those performance related claims.
We shall see.

 

[itsmydamnation]

Yeah remember back in zen2 days Mike was gushing over zen5. Didn't even care about zen4.

 

[Markfw]

My OPINION on this thread as of late. I do think there is an extremely good chance that Zen 5 will be a big jump over Zen 4, based on what I know of Genoa, and have read about Bergamo, and what my 7950x's are doing. I won't even guess on pricing, but at this point I am just waiting to see the benchmarks, and what happens to pricing.

As for how much they beat the current Intel chips (whatever they may be) all I can say is that I do believe that in the next 2 years Intel will gain some ground, but I don't think they will contend in server in that timeframe.

Bottom line at this point ? Wait and see.

 

[DrMrLordX]

Market conditions will need to improve for AMD to be able to jack up prices on their consumer parts. Unless they just don't care about volume and want to move the peg.

 

[adroc_thurston]

Market conditions will need to improve

They already did, client rev is up QoQ for both x86 vendors.

Unless they just don't care about volume and want to move the peg.

The perf is good enough to be a volume driver in itself.
Only mobile can be a bit of an issue since Intel is very keen about flooding the market with cheapo RPL 282 parts.

 

[Joe NYC]

They already did, client rev is up QoQ for both x86 vendors.

The perf is good enough to be a volume driver in itself.
Only mobile can be a bit of an issue since Intel is very keen about flooding the market with cheapo RPL 282 parts.

Can RPL still be used with DDR4 in notebook space?

The DDR5 transition has caused a setback for AMD in the notebook space. I wonder if the stars are getting aligned better for Phoenix now Srix next year...

 

[adroc_thurston]

Can RPL still be used with DDR4 in notebook space?

Yes!
And 282 is both cheap to make and cheap to sell.

 

[JustViewing]

I feel expecting >30% IPC (per clock) is over optimistic. It may get that in selected micro benchmarks, or games, but I doubt it will get pure IPC increase of >30% compared to Zen4. Getting that much of IPC from existing code is very difficult. We may get higher IPC with new changes proposed by Intel to increase the register count. But that will be in the future and application has to be re-compiled for that.

End of the day same story will repeat, Bad AMD, they promised over 30% IPC and what we got is 15-20%. AMD cheats and lies... Not worth the sand...

Also having 4 generations of 16 core max is setting themselves for disaster regardless of single core performance.

 

[H433x0n]

I feel expecting >30% IPC (per clock) is over optimistic. It may get that in selected micro benchmarks, or games, but I doubt it will get pure IPC increase of >30% compared to Zen4. Getting that much of IPC from existing code is very difficult. We may get higher IPC with new changes proposed by Intel to increase the register count. But that will be in the future and application has to be re-compiled for that.

End of the day same story will repeat, Bad AMD, they promised over 30% IPC and what we got is 15-20%. AMD cheats and lies... Not worth the sand...

Also having 4 generations of 16 core max is setting themselves for disaster regardless of single core performance.

AMD never actually made a specific claim about IPC improvement. It’s just the usual leakers making these statements.

I imagine most of it originates from a Specint 2017 benchmark of Turin. I don’t think this lone Turin benchmark is enough to extrapolate a 30%+ IPC uplift and AMD world domination but that’s the usual AMD hype train. Genoa benchmarks also showed massive gains in Specint (20%+) and the IPC gains ended up at 13%. Milan Specint (1-copy) had 30% gains and the client products had a 19% IPC uplift.

Here’s Milan (Epyc 75F3) showing a 30% improvement over Rome (Epyc 7742).

 

[JustViewing]

AMD never actually made a specific claim about IPC improvement. It’s just the usual leakers making these statements.

That is exactly my point, it is always someone makes outlandish claims about AMD performance then come back and complain AMD lies and cheat.

 

[Joe NYC]

I feel expecting >30% IPC (per clock) is over optimistic. It may get that in selected micro benchmarks, or games, but I doubt it will get pure IPC increase of >30% compared to Zen4. Getting that much of IPC from existing code is very difficult. We may get higher IPC with new changes proposed by Intel to increase the register count. But that will be in the future and application has to be re-compiled for that.

AMD64 / x86-64 already increased the number of registers. So that is going to start reaching diminishing returns.

End of the day same story will repeat, Bad AMD, they promised over 30% IPC and what we got is 15-20%. AMD cheats and lies... Not worth the sand...

Also having 4 generations of 16 core max is setting themselves for disaster regardless of single core performance.

I think most people are learning that for 95% plus desktop users, including gamers, going beyond 8 (!) cores has diminishing returns. Going beyond 16 cores, you are down to perhaps 1% that will derive any benefit from it.

 

[Saylick]

AMD never actually made a specific claim about IPC improvement. It’s just the usual leakers making these statements.

I imagine most of it originates from a Specint 2017 benchmark of Turin. I don’t think this lone Turin benchmark is enough to extrapolate a 30%+ IPC uplift and AMD world domination but that’s the usual AMD hype train. Genoa benchmarks also showed massive gains in Specint (20%+) and the IPC gains ended up at 13%. Milan Specint (1-copy) had 30% gains and the client products had a 19% IPC uplift.

Here’s Milan (Epyc 75F3) showing a 30% improvement over Rome (Epyc 7742).

View attachment 84946

Fwiw, this is what Ian got for his DT Zen 3 review:

About a 19% gain, which is inline with the advertised IPC uplift.

Will we see Zen 5's SPEC scores match up with what AMD advertises? We'll just have to wait until the announcement to find out.

 

[H433x0n]

Fwiw, this is what Ian got for his DT Zen 3 review:
View attachment 84947
About a 19% gain, which is inline with the advertised IPC uplift.

Will we see Zen 5's SPEC scores match up with what AMD advertises? We'll just have to wait until the announcement to find out.

This particular chart appears to take into account clock speed, which is why it comes out to ~15%. Zen 3 also had a slight frequency bump over Zen 2 which got it to that overall 19% uplift.

 

[JustViewing]

When I mentioned IPC, I specifically mentioned Instructions Per Clock and not performance increase. The above graphs are achieved through clock speed increase as well. The claim here is there is no significant clock speed increase with Zen 5 compared with Zen 4.

AMD64 / x86-64 already increased the number of registers. So that is going to start reaching diminishing returns.

16 general purpose registers are not enough. Atleast I personally found it limiting. The issue here is not whether 16 regs are enough or not. But to do wider CPU architecture you need more registers so CPU can run more instructions in parallel. Otherwise, wider architecture will go unused most of the time. Register renaming have only limited use when you have to constantly save and reload values to register. Basically 16 registers is sort of bottleneck for doing wider design.

 

[Saylick]

This particular chart appears to take into account clock speed, which is why it comes out to ~15%. Zen 3 also had a slight frequency bump over Zen 2 which got it to that overall 19% uplift.

15% for SPECfp, 19% for SPECint.

 

[Gideon]

Honestly, I'm not sure why everyone is giving so much attention to this forum leaker. Besides the fact that the CPU forum has had something like a 0% success rate on previous forum leakers, the $1000 claim for 16 core Z5 is totally a red flag. Firstly because of how unlikely it sounds, and secondly because of the sheer amount of confidence he's projecting with this claim. Pricing is the most flexible, most fluid, most likely to be changed before launch. A real industry insider would know this and wouldn't harp on something that's so likely to leave egg on his face. At most he might say to expect a large price increase.

My thought exactly. Price is the only part of the product that can easily be changed willy-nilly up to the launch. It's the most fluid part of the product.

See the slides AMD handed out to the press regarding RX 7800 as an example, where the price is still in the flux, despite being sent, what 24h before the presentation. There were (rather substantiated) rumors that Jensen pretty much decided the prices of some GPUs right before the presentation.

This can only be a guess at best not some kind of 100% set in stone known fact.

 

[BorisTheBlade82]

Interesting, in that case maybe AMD has a 192-core version waiting in the wings. But (hate to be this guy) do you have a source on the 8 ports?

Well, there are annotated die shots around for the sIOD showing 12 ports .That is what Genoa uses in the Top End SKUs.
That Bergamo uses 16c CCDs is common knowledge - so they use 8 ports for 8 CCDs = 128c.
Why did they not come up with a 12 x 16c SKU with 192c in total?
Well, maybe package size is a constraint, maybe TDP, maybe routing. I honestly don't know. Maybe they wanted to save that up for Turin Dense?

 

[eek2121]

My thought exactly. Price is the only part of the product that can easily be changed willy-nilly up to the launch. It's the most fluid part of the product.

Clocks are also decided somewhat late in the game, but earlier than pricing.

As I alluded to in another thread, model numbers can also be tweaked, so for example, if Intel drops a chip that is much faster than a hypothetical 8800X, AMD has the option of changing that chip to an 8700X or 8600X. This also has to be done earlier on otherwise you end up with an NVIDIA type situation.

I am curious as to whether AMD will adopt their new naming scheme for Zen 5.

 

[DrMrLordX]

They already did, client rev is up QoQ for both x86 vendors.

Not to the level it was in 2020 (overall). Plus AMD Osbourned themselves with Zen4 . . . it'll be funny if they do it again.

 

[Joe NYC]

Not to the level it was in 2020 (overall). Plus AMD Osbourned themselves with Zen4 . . . it'll be funny if they do it again.

How did they Osborne themselves with Zen 4?

 

[CakeMonster]

That is exactly my point, it is always someone makes outlandish claims about AMD performance then come back and complain AMD lies and cheat.

Lets place the blame where it belongs though, its not randoms on forums (or even twitter) that cause this, its the people who have gotten low level fame (undeserved) from making stuff up before, make money off it, and are unfortunately referenced by the tech media.

 

[CakeMonster]

The claims about Z5 and Intel 15th gen are interesting, but what is going to be much more decisive is the timing. Will Intel stick to their regular schedule and launch next fall? Will AMD launch earlier than the recent ~24m schedule as claimed? In that case Z5 would really make an impact. Its way too early to say for sure though, even if all the rumors are true we are still some way from release and there could be snags with nodes, production, clock speed, etc. Sounds like AMD is the bigger wildcard though, I would guess Intel is more likely to stick to the schedule for desktop.