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

[DisEnchantment]

Speculate at will

 

[poke01]

AMD is more than fine for 2024 and 2025 but in 2026 when Nova Lake releases and AMD will be on Zen 5 that is not good.

 

[Abwx]

I see what you mean now.

Edit: I like to add that was R23. An M2/M3 core is significantly better in CB 2024.

That s right but at the same time Zen 4-5 consume quite less power in CB 2024, so what is lost on one hand is gained on the other .

As for Apple they dont sell CPUs but complete laptops and as a manufacturer they surely want to get rid of the most annoying and unreliable part wich is the fan, hence their fanless and low power strategy.

AMD is more than fine for 2024 and 2025 but in 2026 when Nova Lake releases and AMD will be on Zen 5 that is not good.

I dont see this as a problem as they managed to sort out the main difficulty wich is to get their 6 ALUs core working flawlessly, it will be much easier to improve the design now that the fundamentals are proved to be reliable, guess that Zen 6 is already well advanced at this point.

 

[lucasworais]

When is Turin expected to launch?

 

[Josh128]

I dont know if anyone has mused this yet, but any chance these weird PBO results (good gains with HUGE power increases) may somehow be similar to the 11900K backport debacle? Like, maybe this architecture was actually designed for 3nm, but worked for a 4nm release, but at the cost of a lot of power for good speed?

Rocket Lake seemed like a turd, and supposedly it was just a backport of Golden Cove-- when Golden Cove was released on 10nm it was super impressive performance. Maybe Zen 5 is similar in some way? I didnt believe that it was a backport, but from what we are seeing, it certainly seems like it could be exactly what we are seeing...

 

[yottabit]

Zen ±5%

 

[adroc_thurston]

When is Turin expected to launch?

Soon™.

I dont know if anyone has mused this yet, but any chance these weird PBO results (good gains with HUGE power increases) may somehow be similar to the 11900K backport debacle

It's a foundry targeted core, backporting is a pretty low effort job.

 

[Jan Olšan]

I dont know if anyone has mused this yet, but any chance these weird PBO results (good gains with HUGE power increases) may somehow be similar to the 11900K backport debacle? Like, maybe this architecture was actually designed for 3nm, but worked for a 4nm release, but at the cost of a lot of power for good speed?

Rocket Lake seemed like a turd, and supposedly it was just a backport of Golden Cove-- when Golden Cove was released on 10nm it was super impressive performance. Maybe Zen 5 is similar in some way? I didnt believe that it was a backport, but from what we are seeing, it certainly seems like it could be exactly what we are seeing...

Rocket Lake was port of Sunny Cove (the Ice Lake CPU core) to 14nm, not Golden Cove.

As for getting worse performance/watt when ramping processor up to higher points of voltage-frequency curve, that is expected behaviour. It generally happens with PBO?

 

[gdansk]

Gamers complain and yet it'll still beat Arrow Lake when AMD stacks some cache on it.

 

[Abwx]

I dont know if anyone has mused this yet, but any chance these weird PBO results (good gains with HUGE power increases) may somehow be similar to the 11900K backport debacle? Like, maybe this architecture was actually designed for 3nm, but worked for a 4nm release, but at the cost of a lot of power for good speed?

Rocket Lake seemed like a turd, and supposedly it was just a backport of Golden Cove-- when Golden Cove was released on 10nm it was super impressive performance. Maybe Zen 5 is similar in some way? I didnt believe that it was a backport, but from what we are seeing, it certainly seems like it could be exactly what we are seeing...

That s not comparable, with N3P instead of N4P they would have either 6% better perf at isopower or 15% lower power at isoperf.

So quite better but at the same time not enough to be the main factor in the resulting efficency.

 

[Josh128]

Well, the 16 core CCX sounds potent, but its obviously not possible with 4nm. We are not getting all of the Zen 5 leaked roadmap features on 4nm, they are only possible on 3nm or smaller. This is why I bring up the possibility.

 

[maddie]

So if the below is true, @adroc_thurston was correct.

View attachment 104680

Now reading through so might be repeating.

Will be interesting to see how Strix Halo with it's 256b memory bus work on CPU tasks?

 

[CakeMonster]

Was there a confirmation of 16 big core CCX for Zen6? I certainly missed that, I thought the consensus (of unreliable rumors) was still 8c?

 

[Goop_reformed]

Kepler_L2 said Zen 6 is planned for 2027.
It must be wrong like other things he said. But if it is true and Microsoft heard of the roadmaps it is rational, not obsessive, to redouble their ARM efforts.

If that is the case then yikes. Zen 5 being severely underwhelmed also doesn't help.

 

[gdansk]

Was there a confirmation of 16 big core CCX for Zen6? I certainly missed that, I thought the consensus (of unreliable rumors) was still 8c?

No but the slide which accurately estimated Zen 5 performance also said "NEW - 32 core complex" for Zen 6.

 

[Markfw]

Zen ±5%

I don't know where you get that, without cherypicking the worst review of the day, 10%-25% which is right in line with AMD's 16% number.

 

[Thunder 57]

If that is the case then yikes. Zen 5 being severely underwhelmed also doesn't help.

I was hoping we would cool it with rumors for a bit after what has happened with Zen 5. But apparantly not. Next thing they will be talking about Zen 6 launching on 6/6/2026 with LPDDR6 and PCIe 6.

 

[gdansk]

I was hoping we would cool it with rumors for a bit after what has happened with Zen 5. But apparantly not. Next thing they will be talking about Zen 6 launching on 6/6/2026 with LPDDR6 and PCIe 6.

Way ahead of you

Here I can make one:
6% more IPC
6 GHz
64 threads
LPDDR6
$699
6/6/2026
There. We have our Zen 6 hype train

 

[Doug S]

I was hoping we would cool it with rumors for a bit after what has happened with Zen 5. But apparantly not. Next thing they will be talking about Zen 6 launching on 6/6/2026 with LPDDR6 and PCIe 6.

66% IPC gain!

 

[Keller_TT]

AMD is more than fine for 2024 and 2025 but in 2026 when Nova Lake releases and AMD will be on Zen 5 that is not good.

Not sure about 2025 too, when QCOM, Apple launch again and if the ARM pressure ramps up with Nvidia too. I get the feeling that AMD will launch a Zen 5+, as Zen 5 was originally poised for what's now N3E, but AMD had to readjust due to TSMC's issues. Turin will make use of the node anyway, and they could even go for N3P next year or beginning of 2026 for a 5+ optimization. Another 5-10% and just chip away with more efficiency for both laptops & desktop. Then, Zen6 in '27 is ok if it's worth the wait.

 

[Jan Olšan]

Problem is, if AMD wanted to make 3nm Zen 5+ for desktop, we would likely know about it, it would be about to tape out.

OTOH, a 9990X unit with 8x Zen5 + 16x Zen5c chiplet (that already exists), that sounds doable on shortish notice, likely? Question is how well would such uneven SKU be accepted. The clocks on the dense cores would be nowhere near the clocks on the classic die, so the performance uplift would be limited. And you can't do 32core dense-only, because the same people that whine about Zen 5 being a "flop" in games would eat them alive.

 

[Keller_TT]

Problem is, if AMD wanted to make 3nm Zen 5+ for desktop, we would likely know about it, it would be about to tape out.

OTOH, a 9990X unit with 8x Zen5 + 16x Zen5c chiplet (that already exists), that sounds doable on shortish notice, likely? Question how well would be accepted. The clocks on the dense cores would be nowhere near the clocks on the classic die, so the performance uplift would be limited. And you can't do 32core dense-only, because the same people that whine about Zen 5 being a "flop" in games would eat them alive.

I think there's still good time if it's not going to launch before Oct/Nov 2025 or early 2026. Will only be a refinement and not an all new design, and Zen 5 being a ground-up design taped out late in 2023 for a mid-2024 launch. So, I hope it happens and there's news.
Already not liking this idea of not offering proper 8 Zen 5 cores for mobile and just loading up on e-cores. So, I'm only eager to see how Strix Halo plays out and how it's priced.

 

[Mahboi]

>get STX reviews
>read a few pages of the forums, stop at 693
>return a week later for Granite Ridge first salvo
>page 731

I am disappointed and surprised at how much AMD stopped giving a dang about INT performance and somehow decided to go all in on what is objectively extreme FP/SIMD preference.
Zen 5 is very interesting and a massive improvement in a lot of special, mostly server/enterprise oriented cases, and kind of a big "BUT WHY?" for almost all standard programs and workloads that you'd expect out of a basic consumer CPU.
X86 seems to want to slowly design itself out of desktop and mobile in favour of server, while ARM seems to enjoy responding to actual client needs with chips designed for clients, whether they're great or not. This concerns me somewhat.

See you in a few days in 40 more pages.

 

[QuickyDuck]

Zen5 is more of a epyc arch, not ryzen.

 

[Markfw]

Zen5 is more of a epyc arch, not ryzen.

The same chiplets that are on Ryzen are on all server CPUs. Its been this way since the Naples CPUs. The only difference is binning. Hence the design change for Ryzrn/Turin

 

[Mahboi]

Even the server landscape is changing rapidly.

1. On one hand you have the HPC applications - compute clusters in university labs, supercomputers in government research etc. These demand peak performance, and where the increased FP perf will matter.

Aye

2. Everything else is better served by a product like Sierra Forest - which is most of your cloud applications. Here, peak performance matters less, often a significant portion of the cores will remain idle or very little utilization at best. Per core perf/Watt matters more than total perf/watt when all cores are active, and for security reasons the cloud providers will disable SMT anyway - so the so-called MT advantage of Epyc is moot.
The most important thing is that Category 1 is dwarfed by Category 2.

AMD is slow to catch up to the bifurcation in the DC landscape.

Wrong. Z5 is precisely what you just described, or trying to.
A higher core count, lower power draw, relatively small but powerful enough SIMD crunching machine. The Phoronix reviews were quite clear on it.
Of course you could argue that SF or that a future Skymont based massive SKU for server could be better than Z5. But Z5 is aiming precisely for that market. While leaving consoooooomers behind, which I'm not too pleased with.