Question Zen 6 Speculation Thread

[adroc_thurston]

is going to look quite week.

compared to what exactly

 

[Joe NYC]

uhhh no, AMD is not building tablet parts.

Matching MacBook Air and MacBook Pro needs to be the target for AMD Zen 6 and Zen 7 generations. Not IPad.

MTL and PTL are both performance regressions lmao

PTL should be an improvement over LNL and ARL at the same power envelopes and core counts. We will see how it turns out.

 

[poke01]

Considering that the goal is to match and eventually surpass Mac laptops, there is no way AMD can get there with > 2 year or longer cadence (if first Zen 6 Medusa slips to 2027)

lol. There’s a proper MacBook Pro refresh due in 1.5 years. Intel will have do it by then which means nova lake-H/HX.

These last 5 years Intel caught up with what Apple did in 2020.

uhhh no, AMD is not building tablet parts.

I think he means the Pro/Max SKUs, not the iPad parts.

 

[adroc_thurston]

Matching MacBook Air and MacBook Pro needs to be the target for AMD Zen 6 and Zen 7 generations. Not IPad.

tablet IP.

PTL should be an improvement over LNL and ARL at the same power envelopes and core counts

1t is all that matters and PTL 1t is < ARL/LNL 1t.

I think he means the Pro/Max SKUs, not the iPad parts.

meh AMD will continue cobbling those together off dGFX parts.

 

[Fjodor2001]

1t is all that matters and PTL 1t is < ARL/LNL 1t.

No. And especially not for mainstream / general laptops.

 

[poke01]

meh AMD will continue cobbling those together off dGFX parts.

I mean AMD has no competition there. intel and nvidias ip seem a long way off.

 

[Geddagod]

People say this about every release. They have yet to fall behind.

They already are behind tbh

I mean Panther lakes cpu isn’t exactly anything special

competitive CPU, but better iGPU and battery life.

1t is all that matters and PTL 1t is < ARL/LNL 1t.

Maybe better than ARL but it's going to be at worst = vs LNL.

 

[adroc_thurston]

No. And especially not for mainstream / general laptops.

yeah.

I mean AMD has no competition there. intel and nvidias ip seem a long way off.

forget about NV, they always screw it all up.

Maybe better than ARL but it's going to be at worst = vs LNL.

that's not a real achievement.
4+0 PTL is also like 4.8G top bin. pathetic.

 

[jpiniero]

1t is all that matters and PTL 1t is < ARL/LNL 1t.

It's not Ayy Eye?

 

[adroc_thurston]

It's not Ayy Eye?

no, you only need enough of that to get a windows license rebate.

 

[poke01]

No. And especially not for mainstream / general laptops.

I mean 1t plays a part. MacBook Airs weren’t even in the top 5 best selling laptops before M1/2020.

Most of the stuff people do on general laptops is 1t based anyway.

 

[Fjodor2001]

Most of the stuff people do on general laptops is 1t based anyway.

No. Absolutely not. We left the 1 core era about 2 decades ago.

 

[jpiniero]

No. Absolutely not. We left the 1 core era about 2 decades ago.

Like Cinememe?

 

[Fjodor2001]

Like Cinememe?

Yeah, if you think cinememe is everything above 1c. But the actual PC market left that approach behind about 2 decades ago.

 

[poke01]

No. Absolutely not. We left the 1 core era about 2 decades ago.

i don’t mean a single core CPU. I mean a CPU with multiple cores capable of excellent 1t.

Big difference and makes a huge difference in web based stuff.

 

[Bigos]

In client the performance of a single core is still the most important.

It does not mean it is the only factor. A very fast 1-core CPU would be garbage. Still, it is better to have faster threads than to increase the number of them beyond a certain point. Whether that point is 8/12/16 threads depends on the use case.

You might think otherwise, but do not force other people to think the way you do.

Incidentally, I would really make a use of 16+ cores. I still do not think it is a good fit for client workloads.

 

[Fjodor2001]

i don’t mean a single core CPU. I mean a CPU with multiple cores capable of excellent 1t.

Big difference and makes a huge difference in web based stuff.

So then both 1c and nT matters after all.

 

[poke01]

So then both 1c and nT matters after all.

yes but for most people more nT becomes irrelevant after a certain amount of cores whereas 1t is always important.


* in client workloads

 

[Fjodor2001]

RDNA 5 thread starts with that roadmap of AMD planning a end of H2 2026 release.

That was before what happened with RDNA 4 — when AMD ended up waiting for nvidia to go first

If that repeats — all rumours point to Q3 2027 for nv launch — then RDNA 5 is for sure H2 2027 launch

From here:

AMD RDNA 5 "Next-Gen Radeon" GPUs To Launch In Mid 2027, Utilize TSMC N3P Node

AMD is rumored to launch its next-generation "RDNA 5" GPU family for Radeon graphics cards in the mid of 2027.

wccftech.com

”Rumors Point Towards Mid-2027 Launch For RDNA 5 "Next-Gen Radeon" GPUs, Taped Out on N3P Node”

But it’s Wccftech and based on a Kepler_L2 post:

”It looks like 2026 will mostly be a silent year as far as new AMD Radeon GPU launches are concerned, since Kepler_L2 has stated that the next-gen RDNA 5 GPU launch isn't expected until mid-2027.”

So perhaps just a speculation circle.

 

[Fjodor2001]

yes but for most people more nT becomes irrelevant after a certain amount of cores whereas 1t is always important.


* in client workloads

Problem is we’ve hit a (soft) wall w.r.t. frequency and IPC increase.

Which is why everyone is going for increasing core count instead, and SW will adapt to make use of that.

 

[adroc_thurston]

Problem is we’ve hit a (soft) wall w.r.t. frequency and IPC increase.

nope.

 

[Fjodor2001]

nope.

You’re hoping for 32% IPC increase going from Zen5 -> Zen6?

 

[adroc_thurston]

You’re hoping for 32% IPC increase going from Zen5 -> Zen6?

no, that's for Zen7.
Zen6 is a mix of IPC and fmax both.
Ideally we'll get 35% 1t via 15% IPC and 20% fmax.

 

[Fjodor2001]

no, that's for Zen7.
Zen6 is a mix of IPC and fmax both.
Ideally we'll get 35% 1t via 15% IPC and 20% fmax.

Speculation, and yet they are bumping core count by 50%.

 

[adroc_thurston]

and yet they are bumping core count by 50%.

Then lowering it back to 8c for Zen7.
1t rules.