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

[DisEnchantment]

Speculate at will

 

[static shock]

What is the expectation for Lunar Lake iGPU. Vs. Meteor Lake and vs. Strix Point?

I searched a little and find that lunar lake igpu is as strong as STX gpu. But LNL is 3nm TSMC, STX is 4nm.

 

[jpiniero]

I searched a little and find that lunar lake igpu is as strong as STX gpu. But LNL is 3nm TSMC, STX is 4nm.

Quality difference between N3B and N4P isn't much. That's a big part of the reason why TSMC had so much trouble getting adoption of the node.

 

[adroc_thurston]

Quality difference between N3B and N4P isn't much

Nah, it's genuinely, measurably better at lower power.

That's a big part of the reason why TSMC had so much trouble getting adoption of the node.

It's just very expensive.
N3e is literally cheaper and better so might as well?

 

[Joe NYC]

I searched a little and find that lunar lake igpu is as strong as STX gpu. But LNL is 3nm TSMC, STX is 4nm.

I also seem to recall reading that LNL iGPU is stronger than MTL.

BTW, it looks like Intel did a pretty good job of packing the iGPU, 4+4 CPU, NPU, memory controllers and SLC all inside ~140-150 mm2 die area. N3 helped a ton, for sure, but still...

 

[adroc_thurston]

I also seem to recall reading that LNL iGPU is stronger than MTL.

It is.

BTW, it looks like Intel did a pretty good job of packing the iGPU, 4+4 CPU, NPU, memory controllers and SLC all inside ~140-150 mm2 die area

I/O is external in a PCH so not quite directly comparable to what AMD ships.

 

[Joe NYC]

It is.

I/O is external in a PCH so not quite directly comparable to what AMD ships.

I think Intel got it right partitioning Lunar Lake.

We will see how it comes out cost-wise vs. Strix Point. I am guessing roughly even, with KRK having some cost advantage over LNL...

 

[adroc_thurston]

I think Intel got it right partitioning Lunar Lake.

It's literally the old CPU + PCH partitioning from Haswell-U onwards, just over 2.5D substrate now.

 

[Joe NYC]

It's literally the old CPU + PCH partitioning from Haswell-U onwards, just over 2.5D substrate now.

Just comparing the partitioning of Strix Halo with Lunar Lake.

Supose there is 70 mm2 worth of analog / IO in Strix Halo SoC. Suppose you partition that out and bring in the CPU, as a unified 16 core, with unified L3, but instead of 2 x 69 mm2 of N4P, you add 1 x 100 of N3E to the SoC.

That seems like a good tradeoff, unless AMD is planning to sell the bulk of Strix Halo SKUs with only a single CCD.

 

[adroc_thurston]

Supose there is 70 mm2 worth of analog / IO

I/O needs the shoreline so to tile it out you need a whole lot more stuff than -halo has.

 

[Abwx]

Just comparing the partitioning of Strix Halo with Lunar Lake.

That s not the same concepts and markets, LNL if 4 + 4 and without SMT will have trouble competing even against KRK, 8C/16T with 25% SMT yield can reduce their frequency by 20% and get the same throughput but at 0.6x the power.

With something that scale like Cinebench then it s the same throughput at 0.5x the power, or when a CPU is fully loaded by different applications, that s equivalent to CB and is a real case scenario.

 

[adroc_thurston]

That s not the same concepts and markets

He's comparing disaggregation schemes.
-halo is SOC + a pair of compute tiles and LNL and CPU + HSIO.

 

[adroc_thurston]

May be so. It's just that the bar value usually means overall TS score unless explicitly stated otherwise.

https://twitter.com/x/status/1803286635188199672

There you go.

 

[PJVol]

There you go.

Yeah. Just watched the review of gpd win max 2 (2023) with 7840U inside that scored 3070-ish in TS graphics at 40W cTDP

 

[adroc_thurston]

Yeah. Just watched the review of gpd win max 2 (2023) with 7840U inside that scored 3070-ish in TS graphics at 40W cTDP

Phoenix is ass, Strix is the new champ, especially in more power limited scenarios, where GPU crawls closer and closer to Vmin.
It's time to Feel the Wang again.
Amen.

 

[branch_suggestion]

Halo/KRK pincer at CES brings joy.
Hopefully someone makes the right laptop with Halo, would be a shame to squander such a cool new part with dumb compromises.
Medusa gets so many great upgrades though, along with a lot more platform maturity that I might keep waiting.

 

[itsmydamnation]

Phoenix is ass, Strix is the new champ, especially in more power limited scenarios, where GPU crawls closer and closer to Vmin.
It's time to Feel the Wang again.
Amen.

Spoiler: semi NSFW

 

[PJVol]

Phoenix is ass, Strix is the new champ, especially in more power limited scenarios

Is the TS graphics score for the 12CU Strix known also?

 

[Kolifloro]

Are the X3D core variants so good for virtual machines as they are for gaming ???

I mean ... 'domestic' VMs ...

And what about for generic user tasks ... do they (X3D cores) worth it ??

Till now I had honestly dismissed X3D cores as them being just mainly useful for gamers ... but I am starting to think that MAYBE for a 'power efficiency' sensitive generic user ... they might be interesting ...

OTOH ... I think X3D cores are now ONLY available for desktops, aren't they ??

Any chance for Zen 6 X3D mobile cores in the future ?????


Thank you !!!

 

[biostud]

OTOH ... I think X3D cores are now ONLY available for desktops, aren't they ??


Thank you !!!

Nope 7945HX3D

AMD Ryzen 9 7945HX3D Processor - Benchmarks and Specs

Benchmarks, information, and specifications for the AMD Ryzen 9 7945HX3D

www.notebookcheck.net

 

[TwistedAndy]

I think X3D cores are now ONLY available for desktops, aren't they ?

I'm aware of only one Asus laptop with AMD Ryzen 7945HX3D.

There are some non-gaming use cases, where X3D is useful: https://www.phoronix.com/review/amd-ryzen-7-7800x3d-linux

Personally, I would like to see more AMD laptops with X3D.

 

[CouncilorIrissa]

https://twitter.com/x/status/1803286635188199672

There you go.

Now I'm even more pissed at MS for making AMD ditch MALL.
We'd have a GOAT handheld SoC on our hands, if not a little overkill on the CPU side.

 

[DisEnchantment]

Now I'm even more pissed at MS for making AMD ditch MALL.

Why would MS make AMD ditch MALL?

MS has a requirement for the NPU TOPs for labelling the Windows PCs with CoPilot+.
It is AMD's choice to play in the CoPilot game, seems like they are tripping all over themselves to get the 45+ TOPs and basically throw everything out of the window just to get to label anything as AI.

 

[DisEnchantment]

AMD as usual counting their mm2, be it in GPUs or CPUs. This company is too conservative.
If it were Jensen, you can bet he will have the fattest VCache, bigliest SLC and the widest CPU core possible.

 

[FlameTail]

AMD as usual counting their mm2, be it in GPUs or CPUs. This company is too conservative.
If it were Jensen, you can bet he will have the fattest VCache, bigliest SLC and the widest CPU core possible.

Nvidia and Apple are probably the only 2 companies who willingly do such things. No wonder, they rake in tens of billions and have a gaint $$$ war chest.

 

[coercitiv]

If it were Jensen, you can bet he will have the fattest VCache, bigliest SLC and the widest CPU core possible.

Technically speaking, Jensen tried his luck with a CPU core it did not go so well. To his credit though, the second attempt was indeed as ballsy as it gets: just buy ARM and be done with it.