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

[DisEnchantment]

Speculate at will

 

[Timmah!]

Jim from Adore TV appears to have access to Turin Cinebench R23 scores

View attachment 79368


Without any tweaking 2S 9654 Genoa are as fast as 1P 9654 Genoa on Cinebench R23.

So what are you saying i should stay put and not order the second 9654?
:-D

 

[nicalandia]

So what are you saying i should stay put and not order the second 9654?
:-D

Only if you care about Cinebench, most other Apps like Blender do take full advantage of all full cores

 

[xilli_fiberbit]

https://twitter.com/x/status/1646222837361803273

, seems to be coming from linkedin originally ? then the profile has been deleted and source tweet has been deleted but 2nm

 

[Anhiel]

More details, L1 Cache per core is bigger on Zen 5.

A perfect 1.25x increase in line with 4-wide to 5-wide expansion but with same amount each for dispatching.

Anyhow, the given cinebench r23 value is a perfect theoretical match for 64c and +15%IPC. That's suspect. Last time a perfect match was given for ADL and the actual product was 10% off.

But let's assume it's real then with this and given what's said about MTL, and ARL... MTL is no problem but Zen5 won't beat ARL in ST and probably not MT either (with 16c). It's also unlikely Zen5+ can change that either. Zen6 being another refresh will also be in trouble and needs to increase core count to compete in MT. With Intel's follow on products improving further if the leaks are to believed this would be the end for the Zen architecture. AMD might need to rethink about completely new architectures for Zen6 and RDNA5 successors.

 

[Geddagod]

https://twitter.com/x/status/1646222837361803273

, seems to be coming from linkedin originally ? then the profile has been deleted and source tweet has been deleted but 2nm

View attachment 79412

2nm seems a bit ambitious.
TSMC 2nm HVM in 2H 2025 products 2026 makes sense, but wouldn't all that capacity be given to Apple?

 

[Geddagod]

A perfect 1.25x increase in line with 4-wide to 5-wide expansion but with same amount each for dispatching.

Anyhow, the given cinebench r23 value is a perfect theoretical match for 64c and +15%IPC. That's suspect. Last time a perfect match was given for ADL and the actual product was 10% off.

But let's assume it's real then with this and given what's said about MTL, and ARL... MTL is no problem but Zen5 won't beat ARL in ST and probably not MT either (with 16c). It's also unlikely Zen5+ can change that either. Zen6 being another refresh will also be in trouble and needs to increase core count to compete in MT. With Intel's follow on products improving further if the leaks are to believed this would be the end for the Zen architecture. AMD might need to rethink about completely new architectures for Zen6 and RDNA5 successors.

blinks what?
How much better than RWC do you think LNC is?

 

[Tuna-Fish]

A perfect 1.25x increase in line with 4-wide to 5-wide expansion but with same amount each for dispatching.

It's not a 1.25x increase. It's a 50% increase to either the L1i or the L1d.

 

[Anhiel]

blinks what?
How much better than RWC do you think LNC is?

The leaks say

LNC is 45%IPC over Alder Lake; 20%IPC over Meteor Lake
8-wide

I presume these values are inaccurate or missleading.

theory at best 8-wide/6-wide = 1.33x without anything else accounted for
=> ~24%IPC over RWC => ~29% over RPC
exactly matching leak would be 29% IPC
kind of match leak under my assumption and combined with ~12%CLK ~17%power savings with Intel20A

using stats from https://www.anandtech.com/print/187...ks-review-taking-intel-s-raptor-lake-to-6-ghz
for Cinebench R23
Arrow Lake 8c+16c 5.824GHz/5.488GHz/4.144GHz 105W TDP
for +33%IPC
ST:2958.0
MT:51759.27
for 29% IPC
ST:2869
MT:50202.6

Zen5 16c N4 +15%IPC 8%CLK 6.1GHz/5.8GHz 170W TDP
ST:2543.6
MT:48370.4

Lunar Lake 6c+8c Intel18A 6%IPC 5.8GHz/5.4GHz/4.3GHz 55W base Turbo 157W
ST:3135.5
MT:33958.6

 

[Kepler_L2]

The leaks say


I presume these values are inaccurate or missleading.

theory at best 8-wide/6-wide = 1.33x without anything else accounted for
=> ~24%IPC over RWC => ~29% over RPC
exactly matching leak would be 29% IPC
kind of match leak under my assumption and combined with ~12%CLK ~17%power savings with Intel20A

using stats from https://www.anandtech.com/print/187...ks-review-taking-intel-s-raptor-lake-to-6-ghz
for Cinebench R23
Arrow Lake 8c+16c 5.824GHz/5.488GHz/4.144GHz 105W TDP
for +33%IPC
ST:2958.0
MT:51759.27
for 29% IPC
ST:2869
MT:50202.6

Zen5 16c N4 +15%IPC 8%CLK 6.1GHz/5.8GHz 170W TDP
ST:2543.6
MT:48370.4

Lunar Lake 6c+8c Intel18A 6%IPC 5.8GHz/5.4GHz/4.3GHz 55W base Turbo 157W
ST:3135.5
MT:33958.6

RWC is not a 20% IPC increase and LNL is the same core as ARL.

 

[Anhiel]

RWC is not a 20% IPC increase and LNL is the same core as ARL.

That's what I wrote with "=> ~24%IPC over RWC => ~29% over RPC" : RWC is only ~4%-5% over RPL
RPC is only ~3% over GC

In case there's a miss communication due to arrows used:
"=>" this is a conclusion arrow or next step arrow
"->" this is a connecting arrow or lineage arrow

 

[H433x0n]

blinks what?
How much better than RWC do you think LNC is?

The guy you’re responding to seems to be jumping the gun a bit since we don’t really know what to expect from Zen 5.

However, going by historical trends I don’t think it’s a stretch to say it's going to be a big task for Zen 5 to keep up with Arrow Lake. Whenever Intel jumps a full node (22nm -> 14nm and 14nm -> 10nm), it's always been a solid >20% IPC jump. With Lion Cove being 2 full iterations ahead of Raptor Cove coupled with a healthy node improvement from the current Intel 7 process I would be absolutely shocked if Lion Cove doesn't end up at 30% IPC ahead of where we are at now.

I wouldn't bet anything that I care to lose that they actually hit their timeline but for this theoretical conversation, if Zen 5 were competing directly against Arrow Lake they would likely have to bring >20% IPC gain to be at parity.

 

[igor_kavinski]

Even if Zen 5 fails to overtake its future competition, it would still be power efficient coz that is AMD's philosophy in CPU design now. Intel can keep their "too hot to touch" CPU crown. Zen 5 laptops will still deliver better battery life.

 

[H433x0n]

Even if Zen 5 fails to overtake its future competition, it would still be power efficient coz that is AMD's philosophy in CPU design now. Intel can keep their "too hot to touch" CPU crown. Zen 5 laptops will still deliver better battery life.

The efficiency is in large part due to the node. That’s why my inner geek *really* wanted to see both Zen 5 and ARL-S on TSMC N3. I would love to see that comparison but it seems we’ve been robbed of that opportunity.

 

[A///]

blinks what?
How much better than RWC do you think LNC is?

I've used some drugs in the past when I was in my youth but that was eye opening.

 

[yuri69]

Intel can keep their "too hot to touch" CPU crown. Zen 5 laptops will still deliver better battery life.

Yet AMD laptops are nowhere to be found. OEMs would push Intel no matter what.

 

[A///]

I see plenty in stores. Online isn't the be all end all for everything.

 

[soresu]

Lol it looks like I was right back in 2019 😂

AMD Zen 5 "Nirvana" and Zen 6 "Morpheus" Core Codenames Leaked, Confirm Foundry Nodes

An AMD engineer inadvertently leaked the core codenames of the company's upcoming "Zen 5" and "Zen 6" microarchitectures. It's important to understand here what has been leaked. "Zen 5" and "Zen 6" are microarchitecture names, just like the current "Zen 4" and past "Zen 3" or older. AMD uses...

www.techpowerup.com

The CCD based on the upcoming "Zen 5" microarchitecture is codenamed "Eldora," and the "Zen 5" CPU core itself is codenamed "Nirvana." There's no codename for the CCD based on "Zen 6," but its CPU cores are codenamed "Morpheus."

I'd say think K8->K10, rather than Excavator->Zen in terms of change - I think if they meant to make such a dramatic change, they would give it a more interesting/less incremental name than Zen3 (hint hint AMD, Nirvana is perfect!).

 

[A///]

that would depend on whether the Mike Clark interview was out at that time. nirvana in buddhism means to be released from the cycle of birth and death. To start anew would be a stretch to the word. Either way it's the perfect code name for AMD to use. Clark pushed his statements of Zen 5 being a huge game changer.

We now know thanks to the linkedin prolapse of an employee's work those code names and some info about zen 6, specfically for epyc but you take what you can get!still a good guess on your part mr @soreu!!!!

i had more to say but I've forgotten what it was after I got distracted by your reply, it was probably about the same codenames.tpu users seem to think palladiums are still used.

 

[Ajay]

What? Not the alt-rock band. So disappoint.

 

[BorisTheBlade82]

What? Not the alt-rock band. So disappoint.

What would be the interpretation of that?
RYZEN to the stars like Icarus, fallen to the ground (when Cobain joined the 27 club), but gave birth to another EPYC (Foo Fighters)?

 

[soresu]

What? Not the alt-rock band. So disappoint.

I made a vague guess based on Zen itself, and the Zen based µArch (AMD/Chn JV) called Dhyana, which is also a Zen like elevated mental state related to the practice of Yoga.

But ye, I chose Nirvana from the list of related word cos I lek mah Lithium 🤘🏼

 

[DisEnchantment]

Mike Clark randomly repeating himself that IPC isn't everything. Zen 4 IPC is quite conservative, but out of all the new generations in the Zen family, it brings arguably the most amount of perf uplift in a single generation.

 

[soresu]

Mike Clark randomly repeating himself that IPC isn't everything. Zen 4 IPC is quite conservative, but out of all the new generations in the Zen family, it brings arguably the most amount of perf uplift in a single generation.

Depends on the workload.

For SIMD heavy taks that is almost certainly Zen2.

Hoping that Zen5 brings another similar boost along with the scalar IPC improvements.

 

[Exist50]

I think people are giving far too much weight to names. A codename is ultimately chosen because someone sufficiently high up thinks it sounds cool, and possibly with some input from legal. There need not be any deeper reason, fun though it may be to speculate.

 

[Saylick]

I think people are giving far too much weight to names. A codename is ultimately chosen because someone sufficiently high up thinks it sounds cool, and possibly with some input from legal. There need not be any deeper reason, fun though it may be to speculate.

So you're saying if AMD ever had an internal codename called "Intel Killer" or "Da Monster", we should brush it off as some high level engineer's pipedream... Got it.

Joking, of course.