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

[DisEnchantment]

Speculate at will

 

[A///]

Ah fair, I'm one of those pesky power (mostly sound) conscious desktop users though.

The current woes over energy bills may cause some formerly unbothered desktop/enthusiast users to rethink power in their considerations too.

It's okay! I am too because of how delighted my power company is to ream my behind with high peak hours during the summer. I'm going with Zen 4 simply because replacing several Intel systems over time works out to be cheaper if they deadhead at 250 or whatever watts vs what Intel does on their 13900ks. An extra 300-600 watts an hour can be a matter of an extra 200-300 on my bill.

 

[A///]

Oh soresu! I haven't seen you in ages on here.

 

[Joe NYC]

this video is a complete waste of time

he pulled the 15-26% IPC and 2-9% clock increase out of his behind

thats like saying ST performance is 17-37% higher. No shit, Sherlock!

"my sources say" ... yeah, sure, whatever.

I think the important / informative parts of the video are the overall architecture of MCM, process technologies used, chiplets, core counts etc.

It seems the only big change is the new core (details of which AMD is not going to leak), and the rest of the CPU is conservative, incremental, not taking any big risks with process technology either.

Is it a coincidence that MLID comes out with this leak (roadmap) right on the heals of Intel releasing an updated roadmap? I don't believe in coincidences. Probably, someone from inside AMD confirmed some broad points about Zen 5 to MLID.

It seems, from a couple of tidbits, such as "multiple IO dies" for Venice - Zen 6 - that it will be the next generation when AMD switches to Mi300 like architecture for mainstream CPUs. Which will be a big change.

Small change will likely be only minor tweak to Zen 6 core over Zen 5, derived from N3 process node on the performance parts, and also probably aiming for quick time to market.

It seems that AMD may shrink the cadence between major releases to 18 months.

 

[Joe NYC]

Yes, that is a possibility. It's not uncommon for different market segments to use different manufacturing nodes depending on their specific requirements and priorities. It could be that AMD decides to reserve their most advanced node for the markets where performance, power efficiency, and density are the most critical factors, and use a more cost-effective node for markets that are more focused on price-performance. This approach would allow AMD to balance their production costs with the demands of different markets, and deliver the best possible performance for each market segment.

I think it may also be about risk management. Designing a CPU will at some point narrow the design to target a specific process technology node, which may be a year or more before node is in production, and assuming the node is in production (without delays).

So in case of Zen 5, if MLID and AdoreTV rumors are correct, AMD is targeting the main, high performance version of Zen 5 to the "safe" TSMC N4 node, and only relying on more risky N3 node for the high density version of Zen 5 core.

 

[A///]

Their cadence was less than 18 months at one point. Covid threw them off. But why would anyone from AMD talk to mlid? How he got Peddie to come on for an interview puzzles me months later.

Are we one day going to learn mlid was a psyops hire by AMD or Intel? I try to step away from cpu rumor talk every year and have tried to since the p1 days but every time weird discussions like this pull me back in.

 

[Joe NYC]

Their cadence was less than 18 months at one point. Covid threw them off. But why would anyone from AMD talk to mlid? How he got Peddie to come on for an interview puzzles me months later.

Are we one day going to learn mlid was a psyops hire by AMD or Intel? I try to step away from cpu rumor talk every year and have tried to since the p1 days but every time weird discussions like this pull me back in.

Yeah, Zen4 slipped because of Covid, but, if Zen 5 is targeted for Q1 - Q2 2024, it will be Zen 4 + 18 month timing, which is I think what AMD wants.

Also, CDNA2 to CDNA3 is going to be taking much longer, but it is a major change in platform. Hopefully, after Mi300, the cadence will shrink again.

Re: Psyops

Just well timed leaks, I would say.

 

[soresu]

Oh soresu! I haven't seen you in ages on here.

I'm around most days commenting every so often, usually when something big crops up for AMD or ARM announcements or rumors.

I don't actually post a whole lot though compared to when I first started - went a bit OTT for a few months in 2019, about half of my post count was all from that year 😅

I find that the more I post the more aggressive I tend to get about it, so I throttled back for good measure 😄

 

[A///]

Yeah, Zen4 slipped because of Covid, but, if Zen 5 is targeted for Q1 - Q2 2024, it will be Zen 4 + 18 month timing, which is I think what AMD wants.

Also, CDNA2 to CDNA3 is going to be taking much longer, but it is a major change in platform. Hopefully, after Mi300, the cadence will shrink again.

Re: Psyops

Just well timed leaks, I would say.

18 months is about what I came up with the other day in a reply to our resident AMD phile @DrMrLordX. But referring back to something I remember showing him around the time of Zen 3's launch is that one of the higher ups from AMD stated they will have a variable cadence with 18 months being the extreme. This was pre covid that interview. As you pointed out covid did wreck AMD's plans on getting Zen 4 out earlier than their original intention. I believe if covid were not a factor they would have had more time to attempt selling zen 4 before Raptor came along. Intel's releases are often set in stone within a quarter year after year. Zen 4 was on a 22 or 23 month cadence from Zen 3.

 

[turtile]

18 months is about what I came up with the other day in a reply to our resident AMD phile @DrMrLordX. But referring back to something I remember showing him around the time of Zen 3's launch is that one of the higher ups from AMD stated they will have a variable cadence with 18 months being the extreme. This was pre covid that interview. As you pointed out covid did wreck AMD's plans on getting Zen 4 out earlier than their original intention. I believe if covid were not a factor they would have had more time to attempt selling zen 4 before Raptor came along. Intel's releases are often set in stone within a quarter year after year. Zen 4 was on a 22 or 23 month cadence from Zen 3.

Covid didn't delay Zen 4. It was delayed on purpose for CXL. (from Forrest Norrod)

I think Zen 2 took a lot longer than they planned because of the Fab change and new chiplet design.

 

[A///]

Covid didn't delay Zen 4. It was delayed on purpose for CXL. (from Forrest Norrod)

I think Zen 2 took a lot longer than they planned because of the Fab change and new chiplet design.

Oh yes that's right. I even read and posted that interview myself. The original plan was 4-5 months prior but Norrod wanted CXL. Very off topic but on Intel's side there is a lot of questioning to be done. Someone asked me about their interconnects the other day on here and I had no real answer. I've only read the white papers for DMI4.0 and QPI. Even as we praise or joke on Intel there's a lot of mystery revolving around them due to the engineering brain power glut they suffered not long ago under devilish hands.

 

[maddie]

Their cadence was less than 18 months at one point. Covid threw them off. But why would anyone from AMD talk to mlid? How he got Peddie to come on for an interview puzzles me months later.

Are we one day going to learn mlid was a psyops hire by AMD or Intel? I try to step away from cpu rumor talk every year and have tried to since the p1 days but every time weird discussions like this pull me back in.

Usually, when we get a factual outcome that contradicts our beliefs, we should re-evaluate our base assumptions. Unfortunately too few do these days.

 

[Insert_Nickname]

It's okay! I am too because of how delighted my power company is to ream my behind with high peak hours during the summer. I'm going with Zen 4 simply because replacing several Intel systems over time works out to be cheaper if they deadhead at 250 or whatever watts vs what Intel does on their 13900ks. An extra 300-600 watts an hour can be a matter of an extra 200-300 on my bill.

Summer? Try peak hours during winter here. You'll easily pay 2-3-4-5x per KWh then off-peak. Then add that transport is 3x more expensive during peak. That's before taxes and VAT.

They don't even use lube anymore.

 

[Ajay]

Summer? Try peak hours during winter here. You'll easily pay 2-3-4-5x per KWh then off-peak. Then add that transport is 3x more expensive during peak. That's before taxes and VAT.

They don't even use lube anymore.

What country do you live in, if you don’t mind?

 

[Insert_Nickname]

What country do you live in, if you don’t mind?

Denmark.

Link to the official tarif sheet;
https://andelenergi.dk/kundeservice/aftaler-og-priser/nye-tariffer-for-el/

In my case, you need to look under "Cerius". We don't get a choice in our provider. Costs are in "øre", 1/100 of a Danish Krone.

 

[Anhiel]

Is it me or does it look like the Zen4c leaks show no real gain but just proportional clock/power scaling?
Might be the reason why Zen5c is rumored to be on N3 rather than N4 to show them being better...

Zen4's has a 4-wide front-end so Zen5 expanding to 5-wide would fit the leak but that's still less than ADL/RPL 6-wide.
I do wonder if Zen5 will even get parity with Raptor Coves internal latency & bandwidth or remain the same. There's plenty of room left to address for Zen6. Then again too much IPC will starve the cores. Sooner or later on package DRAM or 3D v-cache become inevitable (for Zen6).

I don't expect much of the 1st gen AI/ML accelerators but I do hope they will be good enough across generations and not become worthless after just 2 gens like the change to DLSS3.

 

[turtile]

Is it me or does it look like the Zen4c leaks show no real gain but just proportional clock/power scaling?
Might be the reason why Zen5c is rumored to be on N3 rather than N4 to show them being better...

Zen 4c isn't designed to be better than Zen 4. It's designed to provide more performance and efficiency for cloud workloads that don't require max performance per core.

Zen 5c is on 3nm because the 3nm variant AMD is using won't be ready in time for the Zen 5 release. They also can't pack the number of cores they want on Zen 5c without using the 3nm process.

 

[yuri69]

Zen4's has a 4-wide front-end so Zen5 expanding to 5-wide would fit the leak but that's still less than ADL/RPL 6-wide.
I do wonder if Zen5 will even get parity with Raptor Coves internal latency & bandwidth or remain the same. There's plenty of room left to address for Zen6. Then again too much IPC will starve the cores. Sooner or later on package DRAM or 3D v-cache become inevitable (for Zen6).

Zen 1-4 cores share the same lineage - a traditional 4-wide + uop cache + SMT design. Sure, Zen 4 is way beefier than Skylake/Zen 1 but still.

Let's hope AMD plan to do something more daring and bold than small evolutionary steps with Zen 5/6. Intel seems to iterate faster than AMD so the upcoming cores must be designed with longevity in mind.

 

[Geddagod]

Is it me or does it look like the Zen4c leaks show no real gain but just proportional clock/power scaling?
Might be the reason why Zen5c is rumored to be on N3 rather than N4 to show them being better...

Zen4's has a 4-wide front-end so Zen5 expanding to 5-wide would fit the leak but that's still less than ADL/RPL 6-wide.
I do wonder if Zen5 will even get parity with Raptor Coves internal latency & bandwidth or remain the same. There's plenty of room left to address for Zen6. Then again too much IPC will starve the cores. Sooner or later on package DRAM or 3D v-cache become inevitable (for Zen6).

I don't expect much of the 1st gen AI/ML accelerators but I do hope they will be good enough across generations and not become worthless after just 2 gens like the change to DLSS3.

I hope Zen 5 changes a bit more. "repipelined" front end makes me think that it might not be as clear cut as just going wider, but that might just be hopium.
Honestly though 5 wide wouldn't be too bad, since the uOP cache is still very large, and as long as cache latency is still very low, Zen 5 should be able to deal with smaller structures and still maintain comparable IPC versus Intel.
I think the "wide issue" is referring to instructions from rename to ROB, and IIRC it hasn't been changed since Zen 2 or Zen 1.

 

[moinmoin]

Intel seems to iterate faster than AMD

Intel is releasing new "gens" faster while the actual iterating is bumpy as best. AMD caught up with that yearly cacedence with their mobile chips anyway where it actually manages to both iterate and launch on time (even if availability on the market leaves to desire).

AMD did well with Zen so far using the new family -> expansion rhythm, it likely won't deviate from that. The new Zen c cores may add a twist by moving the effort of moving to new nodes to in-between gens.

By all accounts, being the first Zen family started within a financially stable AMD and all that, Zen 5 should be more daring and may prepare bigger promises for Zen 6 to expand into.

 

[A///]

Summer? Try peak hours during winter here. You'll easily pay 2-3-4-5x per KWh then off-peak. Then add that transport is 3x more expensive during peak. That's before taxes and VAT.

They don't even use lube anymore.

you're in europe, you're getting it either way. the lack of lube will cause friction heat and warm you up.

 

[eek2121]

18 months is about what I came up with the other day in a reply to our resident AMD phile @DrMrLordX. But referring back to something I remember showing him around the time of Zen 3's launch is that one of the higher ups from AMD stated they will have a variable cadence with 18 months being the extreme. This was pre covid that interview. As you pointed out covid did wreck AMD's plans on getting Zen 4 out earlier than their original intention. I believe if covid were not a factor they would have had more time to attempt selling zen 4 before Raptor came along. Intel's releases are often set in stone within a quarter year after year. Zen 4 was on a 22 or 23 month cadence from Zen 3.

Covid didn't delay Zen 4. It was delayed on purpose for CXL. (from Forrest Norrod)

I think Zen 2 took a lot longer than they planned because of the Fab change and new chiplet design.

I will actually disagree, even with him. AMD was selling every chip they made up until this year.My suspicion is that they were making tons of money and did not want to roll out a new platform until they absolutely had to. If you are the market leader and are eating your competitor's lunch, you absolutely do NOT rock that boat until needed.

Intel is releasing new "gens" faster while the actual iterating is bumpy as best. AMD caught up with that yearly cacedence with their mobile chips anyway where it actually manages to both iterate and launch on time (even if availability on the market leaves to desire).

AMD did well with Zen so far using the new family -> expansion rhythm, it likely won't deviate from that. The new Zen c cores may add a twist by moving the effort of moving to new nodes to in-between gens.

By all accounts, being the first Zen family started within a financially stable AMD and all that, Zen 5 should be more daring and may prepare bigger promises for Zen 6 to expand into.

Correct. Intel's biggest performance increase since Skylake was arguably Alder Lake, why? because they added a bunch of cores. why? because AMD...has perf/watt or even perf/watt increased substantially? not really, what little increase there has been came from the new process.

A further note regarding AMD. They've admitted to intentionally (it was in headlines a bit ago, not going to dig up the source right now) not fulfilling 100% of demand (i.e. they could easily push out more 7950X3D and 7900XTX cards if they wanted). They are doing this to drive ASP. A number of market analysts suspect NVIDIA is doing the same for certain GPUs in the segment.

Technically nothing is wrong with that. I bring this up, because AMD can have the best performing stuff in the world, but if you can't buy it, well, it doesn't matter, does it?

Lisa Su has done great stuff for AMD, but focusing on maintaining margins above all else will risk them everything. Just look at Intel. There was a time they worshiped margins.

 

[Markfw]

Lisa Su has done great stuff for AMD, but focusing on maintaining margins above all else will risk them everything. Just look at Intel. There was a time they worshiped margins.

This I have to agree with. While a sane corporate goal and good for investors, they need to start focusing on market share, and that means not making the most money TODAY.

While the server market is pretty well locked up at the moment and for at least a year or 2, they need to start thinking across the board. I don't think they have the OEMS in their pocket yet like Intel.

 

[Mopetar]

Market share is meaningless unless they have the production capacity to achieve it. Having some of the best CPUs on the market enables that, but there's a limit to how much they can expand production and there's an extent to which they can chase market share.

The idea that AMD could reverse positions and become 70% of the market isn't reasonable on any short term time scale. Just be happy that they're a healthy company and making great products.

Everyone should remember that Apple is one of the largest companies on the planet even though they have a global market share that's lower than AMD in a lot of different segments. If market share were a measure of goodness the best burger you could hope to get would be a Big Mac.

 

[Markfw]

Market share is meaningless unless they have the production capacity to achieve it. Having some of the best CPUs on the market enables that, but there's a limit to how much they can expand production and there's an extent to which they can chase market share.

The idea that AMD could reverse positions and become 70% of the market isn't reasonable on any short term time scale. Just be happy that they're a healthy company and making great products.

Everyone should remember that Apple is one of the largest companies on the planet even though they have a global market share that's lower than AMD in a lot of different segments. If market share were a measure of goodness the best burger you could hope to get would be a Big Mac.

OK, just ONE of the things they should think about.... and 70% ? I don't even want that. 40-50% would be just fine. Competition would thrive ! Best for all of us.

 

[Thunder 57]

OK, just ONE of the things they should think about.... and 70% ? I don't even want that. 40-50% would be just fine. Competition would thrive ! Best for all of us.

Yea, I don't think too many people would want 70%, in fact, no one should. Just look at what Intel did or the GPU market now with NVIDIA abusing it. I can't blame them though since a lot of people think NVIDIA is the only option.