Question Zen 6 Speculation Thread

[marees]

that's it? so no Zen7 before 28-29?

No roadmaps with zen 7 so far

Florence is a code name for zen 7 sku as per semi accurate

 

[gdansk]

So... functionally nothing is what you are saying...

Yes, officially nothing. If you want to hype big IPC gains again feel free. But it still won't happen.

 

[Joe NYC]

You're super eager to torture poor-poor SOC people.

A lot is riding on them. Client CPU market is 2x the size of Server CPU market.

forget about SoIC anywhere but extra premium client.

When Lisa was asked about cost of SoIC, she said the cost will come down with volume and time. By the time Zen 7 ships, there will be a lot of time and a volume will be multiple of what it was when first 5800x3d and MilanX were introduced.

Intel's Foveros is very expensive for client applications. So it is not like AMD would be pricing itself from the competition.

AMD may have advantage in cost, competing with monolithic vs. Intel chiplet with expensive packaging, but resulting notebooks are just meh, there are no clear wins. AMD needs clear wins in specific segments to move the needle in market share.

All the wins for AMD in datacenter and desktop come from having better products. In segments where AMD is just meh (notebook, client GPU), nothing happens for years.

Having mastered SoIC is an advantage that AMD has now, but it is not going to last forever. It's great that AMD is going to use it to maximum advantage in server, in Zen 7, but client CPU market is 2x the size of server CPU market.

Also, the upside in client is much larger than in server. By the time of Zen 7, AMD may already have 50% of server revenue. The other 50% of 1/2 size market is much smaller opportunity than 80% of 2x sized market.

 

[adroc_thurston]

Client CPU market is 2x the size of Server CPU market.

Who cares, margins are a lot thinner there.

So it is not like AMD would be pricing itself from the competition.

yeah they would in fact lose their cost advantage and stuff.

AMD may have advantage in cost, competing with monolithic vs. Intel chiplet with expensive packaging, but resulting notebooks are just meh, there are no clear wins. AMD needs clear wins in specific segments to move the needle in market share.

All the wins for AMD in datacenter and desktop come from having better products. In segments where AMD is just meh (notebook, client GPU), nothing happens for years.

Having mastered SoIC is an advantage that AMD has now, but it is not going to last forever. It's great that AMD is going to use it to maximum advantage in server, in Zen 7, but client CPU market is 2x the size of server CPU market.

Also, the upside in client is much larger than in server. By the time of Zen 7, AMD may already have 50% of server revenue. The other 50% of 1/2 size market is much smaller opportunity than 80% of 2x sized market.

mucho texto and you do not understand how the laptop market works.

 

[Cheesecake16]

Yes, officially nothing. If you want to hype big IPC gains again feel free. But it still won't happen.

I am not hyping anything... maybe the increase is 30%, maybe the increase is effectively 0%, who knows...

All I am saying is we don't know much of anything about the improvements to Zen 6... the one official number we have is that 1.7X slide we saw at Advancing AI... which tells us very little other than perfromance increase from a 2P 9965 system to a 2P top of SKU stack Venice system... that is all... nothing else...

 

[gdansk]

I am not hyping anything... maybe the increase is 30%, maybe the increase is effectively 0%, who knows...

All I am saying is we don't know much of anything about the improvements to Zen 6... the one official number

Sure. But you know I'm right. It's like past Zens and inline with the slide. Remind me whenever it ships.

 

[adroc_thurston]

Remind me whenever this turd ships.

That's some unhealthy angst.

 

[gdansk]

That's some unhealthy angst.

Nah, I called it that because only frequency can polish it.
Which is one way of doing things but I wager they'll miss.

 

[adroc_thurston]

I called it that because only frequency can polish it.

That's true for every core on the market.
Pegasus opted for N3X overdrive xtors to hit their 1t CAGR.

 

[poke01]

That's true for every core on the market.
Pegasus opted for N3X overdrive xtors to hit their 1t CAGR.

There is no such thing as a free lunch, someone has to pay for it somehow.

 

[adroc_thurston]

There is no such thing as a free lunch, someone has to pay for it somehow.

his point is frequency bad when AMD and frequency good with ARMcels because ARM good and AMD bad.
That's it.

 

[gdansk]

his point is frequency bad when AMD and frequency good with ARMcels because ARM good and AMD bad.
That's it.

Nope, I've said it a dozen times. IPC obsession is wrong. Performance is what matters.

But ARM vendors are increasing frequency faster than AMD. Go plot a graph from 2022 if you need to.

Strix is a whopping 100MHz ahead. Real speed demon you have there AMD...

 

[adroc_thurston]

But ARM vendors are increasing frequency faster than AMD.

They're a node ahead.

Strix is a whopping 100MHz ahead.

The only N4x part shipped is Granite Ridge/Fire Range and they're like 5.5-5.7 aren't they?

 

[gdansk]

They're a node ahead.

So? Even if they ship a 6.1GHz Medusa Halo they're not increasing performance from frequency as much as Glymur does. And its successor will have some N2 too. Basically, frequency is fine if they can actually increase it faster than the competition. But they won't.

 

[adroc_thurston]

Even if they ship a 6.1GHz Medusa Halo they're not increasing performance from frequency as much as Glymur does.

?
It's 19% fmax for both then.

And its successor will have some N2 too. Basically, frequency is fine if they can actually increase it but being stuck for 48 months at the same fmax is a poor showing.

I get it you want to complain but this is just silly.
The fastest Hamoa is 4.2, that's N4 for you.

 

[gdansk]

?
It's 19% fmax for both then.

Sorry I phrased it wrong. They need to ship over 6100MHz to close that gap. That would simply restore the pre-Glymur status quo, by using N2 to compete with an N3 part from over a year earlier.

And do you really think they'll be shipping 6.1GHz laptops? If they were then those Venice numbers should be much bigger than 1.7x. Maybe 2x.

 

[adroc_thurston]

They need to ship over 6100MHz to close that gap

what if they do that would be funny right.

with an N3 part from over a year earlier.

It's not a year earlier, we're half a year away from Glymur shipping.

And do you really think they'll be shipping 6.1GHz laptops?

Why not, that's what Gator Range is for.

If they were then those Venice numbers should be much bigger.

what? why.
it's 70% more skt oomph at 33% more cores with 20% (generously) more power.

 

[gdansk]

what if they do that would be funny right.

It would be very funny building another set of rails off a cliff. But it's less funny if you board the train yourself.

 

[adroc_thurston]

But it's less funny if you board the train yourself.

Jokes on you, I like trainwrecks.

 

[reaperrr3]

Strix is a whopping 100MHz ahead. Real speed demon you have there AMD...

AMD managed to hit the same clocks at same power with Zen5 as with Zen4, despite only slightly better process and a ~30% fatter core (not counting the denser L3).
It was also a full tock redesign that apparently had some hiccups during development and has a bunch of bugs/errata, some of which will certainly get adressed in Zen6.

I wouldn't take Zen5 as too much of a reference point, it's a server core first and was ported back from N3 to N4P for client on top of the redesign.
Sure, it's disappointing in terms of end result, but that also means a lower bar for Zen6 to improve on.
Note that Zen3 didn't look that impressive outside of games and CCD-latency-sensitive workloads, either.

Zen4 was also only a single-digit IPC improvement outside of games, but still a homerun in the end because clocks.

And Zen6 is to Zen5 what Zen4 was to Zen3.
Sure, the percentual clock increase might land a tad lower, but in turn there's a bigger CCD/L3 and other surrounding improvements this time.

Yes, officially nothing.

The bigger monolithic L3 alone will lift effective IPC improvement in most games over the 10% hurdle, in some cases probably substantially.
Then I'd say 10% higher 1T turbo clocks (6.3 ghz) are pretty much the most pessimistic estimate for clocks, with 14% (6.5 ghz) for top SKUs sounding realistic enough to me, and maybe we'll actually see a tad more.
So ~20-25% uplift at worst and 30+% at best for games, and that's the only performance-critical consumer/desktop workload that matters anyway, everything else is either too niche to matter, or way past "good enough" already.

In server market, I think the mix of modestly higher general IPC, some server-related instructions getting an extra boost, higher core counts per $ (at least some SKUs, presumably), bigger L3 per CCD and (from rumors so far) rather impressive clocks should be good enough, too.
There's still enough slices of cake to take away from Intel, and 10% more or less IPC vs. ARM won't decide what any server customer goes for.
No server vendor would switch from x86 to ARM over 10% more or less IPC.
They either stick with x86, or go ARM for strategic reasons (like TCO in the case of some big names).

 

[adroc_thurston]

Zen4 was also only a single-digit IPC improvement outside of games

hey no, low teens pretty consistently.

 

[gdansk]

Then I'd say 10% higher 1T turbo clocks (6.3 ghz) are pretty much the most pessimistic estimate for clocks,

Hmm. It seems some rumor monger managed to re-define expectations. Well, not problem.

 

[DrMrLordX]

hey no, low teens pretty consistently.

Remember that some people are still stuck on the Zen5% meme.

 

[Joe NYC]

Who cares, margins are a lot thinner there.

There is more operating margin dollars Intel earns from this "thin margin" client business than AMD earns as the entire company.

yeah they would in fact lose their cost advantage and stuff.

AMD mobile strategy is not working. Less than "Me Too" product with a cost advantage is not leading anywhere.

Unlike Desktop and Server, where AMD is competing with superior products and winning in Mobile, the unremarkable SKUs are not gaining any new traction:

mucho texto and you do not understand how the laptop market works.

I understand that current AMD strategy in notebook is not working.

While Intel is struggling everywhere else, AMD managed to gain only 0.2% market share in past year.

In desktop, Raptor Lake is on the ropes, and in notebooks, AMD still has not figured out, after 3 years, how to effectively compete with Raptor Lake in mobile, never mind Lunar Lake.

 

[adroc_thurston]

There is more operating margin dollars Intel earns from this "thin margin" client business than AMD earns as the entire company.

?

AMD mobile strategy is not working

Yes it does.

Unlike Desktop and Server, where AMD is competing with superior products and winning in Mobile, the unremarkable SKUs are not gaining any new traction:

just full on stop, you have no idea what goes into Intel maintaining their share.

I understand that current AMD strategy in notebook is not working.

Works really well doe.

In desktop, Raptor Lake is on the ropes, and in notebooks, AMD still has not figured out, after 3 years, how to effectively compete with Raptor Lake in mobile, never mind Lunar Lake.

Yeah they did, that's why LNL turned from a $1200 laptop part into a $700 one.