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

[DisEnchantment]

Speculate at will

 

[Thunder 57]

I think they need to improve fetch to make sure if they are having to use the decoders, the instruction code is in L1i not in memory For throughput I guess it would be enough if they could make 2 clusters work together in 1T mode.

Maybe they go back to a 64KB L1i cache when they have more room to work with? If the two decoders could work together 2x4 ought to be plenty. There shouldn't be any reason why they cant, right? Pretty sure Skymont does it. Mike Clark said Zen 5 could but C&C's test dispute that claim.

 

[StefanR5R]

On the CCDs in Strix Halo, we now have statements
– from Mahesh Subramony, AMD Senior Fellow, interviewed by Chips and Cheese at CES,
– from Ben Conrad, Director of Product Management for Premium Mobile Client, interviewed by Notebookcheck at CES.

From the latter interview, when asked about the interconnect between CCD and iGPU:
"The interconnect, we internally we call it DDR SSP. I have to figure out if that internal branding is different from what is used in the desktop chip, because we optimize that interconnect for power. When you hold up the Strix Halo die, you find that the CCDs are really close to the I/O die. And because of this we were able to save multiple Watts of power, which is the whole point of building Strix Halo as being really low power for high performance. So, it is a different interconnect, and it is not the identical CCD silicon as our desktop chips."

To recap, Mahesh Subramony said:
"So these are binned, the same architecture, the same set of pipes, the data parts are the same. The differences are in how we bin the part and how we connect the two dies."

Subramony's statement is not entirely conclusive whether the same or a different chip is used. But if the Strix Halo CCDs are a different line of chips, where do the bins go which don't make it for Strix Halo? (Are its chip design and manufacturing shared with Turin?)

I'd also venture to say that the statement from the Director of Product Management is not entirely conclusive whether the silicon isn't identical due to chip design differences or merely due to different binning [edit: and the SERDES sections fused off, of course]. Although he surely made it sound more like the former.

 

[fastandfurious6]

> Strix Halo's RDNA 3.5 offers equivalent memory bandwidth to an RTX 4070 together with 32 MB Infinity Cache. Conscious decision to not opt for on-package memory.

 

[fastandfurious6]

basically they're saying

9955hx = 9950x

halo = 9950x with a few changes for better power/performance ratio + more mem bandwidth

 

[moinmoin]

A detail that is not hasardous :

There are too many laptops combining good iGPUs with bad dGPUs. Strix Halo got the biggest iGPU outside of Apple, so the danger of a good iGPU being combined with a bad dGPU is even higher here. So in that regard I consider that excellent news.

 

[Thunder 57]

There are too many laptops combining good iGPUs with bad dGPUs. Strix Halo got the biggest iGPU outside of Apple, so the danger of a good iGPU being combined with a bad dGPU is even higher here. So in that regard I consider that excellent news.

Yes, I don't understand the desire for a dGPU on Strix Halo.

 

[igor_kavinski]

Yes, I don't understand the desire for a dGPU on Strix Halo.

To have the best of both worlds. RTX GPU for fake frames up the wazoo.

 

[Thunder 57]

To have the best of both worlds. RTX GPU for fake frames up the wazoo.

That just sounds stupid IMHO. Why shell out for a massive iGPU just to not use it? And who cares about fake frames on laptops that likely won't have super high refresh rates?

 

[igor_kavinski]

That just sounds stupid IMHO.

Maybe a good thing. An RTX GPU would probably drive the value way up because then the sheep who look at RTX like something made out of solid gold will buy it whether they really need it or not.

 

[MS_AT]

And who cares about fake frames on laptops that likely won't have super high refresh rates?

Not sure about fake frames, but working software stack for anything else than gaming is something that might appeal to some people

 

[igor_kavinski]

Not sure about fake frames, but working software stack for anything else than gaming is something that might appeal to some people

But then it should only be a workstation class card. Need to keep these lovely laptops from falling into the hands of immature gamer bros.

 

[Abwx]

There are too many laptops combining good iGPUs with bad dGPUs. Strix Halo got the biggest iGPU outside of Apple, so the danger of a good iGPU being combined with a bad dGPU is even higher here. So in that regard I consider that excellent news.

That s indeed a valuable "feature" since it spare some silicon all while neutering the risk of OEMs coupling it with a useless dGPU, beside i m not sure that OEMs are that hungry to curb Nvidia shady practices that mandate them to use their GPUs exclusively or risk being supply limited, this way there s no possible pressure.

Guess that with the next iterations, and as much CUs as a 9070XT, laptops dGPUs will be relegated to heavy DTRs the same way as SFFs are iGPUs counterparts of the regularly dGPU equipped ATX/uATX PCs.

 

[fastandfurious6]

medusa zen6 with rdna4 may render laptop dGPUs obsolete (except 5090 class)

 

[poke01]

Yes, I don't understand the desire for a dGPU on Strix Halo.

with next gen halo I agree, especially with LPDDR6 and RDNA4/5.

 

[tsamolotoff]

For once, AMD should just release a product without any rehearsed marketing speak.

Looks like the N4x GPUs are going this way, weird things abound On a more serious note, typically the discoverers (tm) (C) G.W.Bush are not doing flashy videos a-la Linus Sebastian and if some David Kantor or Chips & Cheese discovers something interesting, it would have magnitudes less impact than the tech-tubers saying "it's bull-output" after they've failed to setup their systems properly (naturally thanks to MS other bull-products like Gamebar)

 

[Geddagod]

Yes, I don't understand the desire for a dGPU on Strix Halo.

If the Strix Halo LP island on the IO die rumors are true, there's a decent chance that we can get all the performance of the "HX" parts of past AMD and Intel processors while having much, much better battery life.
I do see where people are coming from when they say they want such a product.

 

[MS_AT]

If the Strix Halo LP island on the IO die rumors are true, there's a decent chance that we can get all the performance of the "HX" parts of past AMD and Intel processors while having much, much better battery life.
I do see where people are coming from when they say they want such a product.

But the part of better battery life is sticking to iGPU. Won't dGPU nullify the advantage?

 

[igor_kavinski]

But the part of better battery life is sticking to iGPU. Won't dGPU nullify the advantage?

Use iGPU on battery. Use dGPU when plugged in. More flexibility.

 

[GTracing]

If the Strix Halo LP island on the IO die rumors are true, there's a decent chance that we can get all the performance of the "HX" parts of past AMD and Intel processors while having much, much better battery life.
I do see where people are coming from when they say they want such a product.

How many customers are there that need a) 10 hours of battery life, b) better than a 4070 mobile tier GPU, and c) more CPU performance than Strix Point can offer.

If a customer can compromise on any one of those things, they can save hundreds of dollars. The market for such a laptop would be so small that I wonder if any laptop manufacturer would even be willing to build one.

 

[StefanR5R]

If the Strix Halo LP island on the IO die rumors are true,

Given that AMD did not mention any, either these cores do not exist or they are disabled for this round. I don't expect AMD to suddenly tell us "oh, and one more thing" about Strix Halo until devices get onto the market.

Use iGPU on battery. Use dGPU when plugged in. More flexibility.

But why provide a large iGPU then? With all the implications for extra effort and compromises with the SoC's memory interface (double width, memory-side cache, on-chip fabric...), the corresponding RAM layout and BOM, and the larger cooling apparatus.

 

[yottabit]

How many customers are there that need a) 10 hours of battery life, b) better than a 4070 mobile tier GPU, and c) more CPU performance than Strix Point can offer.

If a customer can compromise on any one of those things, they can save hundreds of dollars. The market for such a laptop would be so small that I wonder if any laptop manufacturer would even be willing to build one.

The thing about battery life is for people actually doing work. You’re not going to get 10-20 hours while loading the APU fully. That’s the reason people want “20+ hour” battery life and the reason Macbook Pro is so popular. It’s not so I can browse ATOT for 20 hours, it’s so I can be running Unreal Engine or performing engineering simulations on-the-go for at least a few hours at a time without stressing about a charger, and no loss of performance. It’s really pretty liberating and I’d love to see x86 deliver on the same premise

I really hope more Strix Halo systems get announced soon. I’m personally hoping for a thicker Thinkpad or other system with a giant battery. I’m a bit disappointed that everything is a thin and light so far.

 

[yottabit]

Honestly I have very little trust in OEMs that they'll produce compelling Strix Halo products

I have the feeling it's going to be thin and lights and then giant beefy gaming laptops with extra dGPU nobody asked for.

AMD themselves doesn't help these things. This is really where having some standard reference hardware for the OEMs would help

Since Strix Halo will likely be fastest single thread and multithread CPU on the market OEMs will want to use it for discrete gaming laptops for those people that want the absolute most performance regardless of cost, then it will get flamed in reviews for being overpriced for what it is.

If they do a good enough job at fusing off the dGPU when running on battery I might still consider a laptop like that, but I'm not too hopeful. Maybe too much pessimism but we'll see

 

[GTracing]

Honestly I have very little trust in OEMs that they'll produce compelling Strix Halo products

I have the feeling it's going to be thin and lights and then giant beefy gaming laptops with extra dGPU nobody asked for.

AMD themselves doesn't help these things. This is really where having some standard reference hardware for the OEMs would help

Since Strix Halo will likely be fastest single thread and multithread CPU on the market OEMs will want to use it for discrete gaming laptops for those people that want the absolute most performance regardless of cost, then it will get flamed in reviews for being overpriced for what it is.

If they do a good enough job at fusing off the dGPU when running on battery I might still consider a laptop like that, but I'm not too hopeful. Maybe too much pessimism but we'll see

Strix Halo doesn't support dGPUs.

Page 926 - Discussion - Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

Page 926 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[igor_kavinski]

But why provide a large iGPU then? With all the implications for extra effort and compromises with the SoC's memory interface (double width, memory-side cache, on-chip fabric...), the corresponding RAM layout and BOM, and the larger cooling apparatus.

ASUS is selling a thin and light model. Assuming Dell, HP and Lenovo release their workstation models with 15.6 and 17 inch displays, there's going to be a lot of space available. They can't put more than a 99 Wh battery in it so I assume there will be space available for a dGPU and its heatsink/cooling pipes etc. And it would let them price the dGPU models higher.

 

[LightningZ71]

Fire Range will likely have the highest ST available in mobile and will be the one with the dGPUs at the top end of the market. Strix Halo doesn't have the PCIe lanes for a dGPU with competitive available lanes left for storage, etc.