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

[DisEnchantment]

Speculate at will

 

[Mahboi]

Frankly, the existence of the 128-bit "LP" model there puts this whole slide under doubt. That one doesn't make any sense in the slightest, that's just simply far too cut down.

That one implies to me that MLID was told something about "LP" with regards to Strix Halo from an OEM or someone else, and assumed it meant there would be a significantly cut down model, and just made up the specs (by just halving everything).

Ignoring the video, I'm pretty sure I saw a Twitter leak earlier of a 32 CU version...can't remember the number of cores, but frankly, 16 cores for every Strix Halo sounds wasteful compared to typical AMD.

 

[coercitiv]

We have confirmation from Anandtech on the new naming scheme leak:

Upcoming AMD Ryzen AI 9 HX 170 Processor Leaked By ASUS?

www.anandtech.com

Asus changed the CPU names but Anandtech found the receipts in Google's cache:

While we don't speculate on rumors, we confirmed this ourselves by digging through Google's cache. Sure enough, as the image above from Google highlights, it lists a newly unannounced model of Ryzen mobile processor. Under the listing via the product compare section for the ASUS Vivobook S 16 OLED (M5606) notebook, it is listed with the AMD Ryzen AI 9 HX 170, which appears to be one of AMD's upcoming Zen 5-based mobile chips codenamed Strix Point.

 

[adroc_thurston]

Is that so?

Yea.

Just like you knew about the naming scheme and weren't proven wrong the very next day.

Consult the ThinkPad T14 pdf please

 

[adroc_thurston]

Is that so?

Yea.

Just like you knew about the naming scheme and weren't proven wrong the very next day.

Consult the ThinkPad T14 pdf please

 

[Aapje]

Consult the ThinkPad T14 pdf please

Twice!

 

[csbin]

The Zen5 improves 10% relative to the Zen4 IPC, and >10% in the CR23 1T test. Also, Win10 enthusiasts, please note that starting with Strix Point, AMD will not be providing Win10 drivers.

https://weibo.com/7517897889/O8cQlaMP3

 

[Hitman928]

The Zen5 improves 10% relative to the Zen4 IPC, and >10% in the CR23 1T test. Also, Win10 enthusiasts, please note that starting with Strix Point, AMD will not be providing Win10 drivers.

https://weibo.com/7517897889/O8cQlaMP3

10% would mean AMD messed up the design really bad. I very much doubt this is true.

 

[Mahboi]

I don't buy the performance leak for a second, but the abandonment of Win10 drivers is credible to a degree. I wonder about it though. Might be serious backlash if true.
Main reason not to believe it is that drivers for Win 10 should be peanuts if drivers for Win 11 are out, meaning that it would be a conscious decision to drop Win 10 users, which are still plentiful. Maybe some MS pressure to let go of Win 10 or some "AI" justification/stepping into Win12?

 

[SteinFG]

Whatever the result will be, be it +5 or +25%, it's gonna be the same for the end user really, because it'll be priced accordingly. Although I'd be a little sad that all those whispers for the last couple months lead to nothing exciting.

 

[csbin]

10% would mean AMD messed up the design really bad. I very much doubt this is true.

The fact that this guy is a product manager at Lenovo has some credibility to me.

 

[Kaffeekenan]

The fact that this guy is a product manager at Lenovo has some credibility to me.

I don't buy it at all. For a wider grounds-up major redesign that would be a f*** up of Bulldozer proportions. Never ever are 10% true...I hope... xD

 

[ToTTenTranz]

Slide apparently comes from older presentation:
View attachment 98568

Even the 20 CU + 128bit LPDDR5X should be a massive performance upgrade over the 16 CU + 128bit LPDDR5X Strix Point, because it has 32MB Infinity Cache that makes a world of difference at 1080p and below.

 

[leoneazzurro]

I don't buy it at all. For a wider grounds-up major redesign that would be a f*** up of Bulldozer proportions. Never ever are 10% true...I hope... xD

I mean... even the lowest preliminary Geekbench score leak ST % increase was higher than that at iso clock. If we would speak about MT scores at same core number in power limited situation, I will believe for sure. We are basically at the same process node.

 

[DavidC1]

I don't buy it at all. For a wider grounds-up major redesign that would be a f*** up of Bulldozer proportions. Never ever are 10% true...I hope... xD

Lion Cove is "supposed" to be 33% wider decode with more than 2x uop capability and also an expansion for both FP, ALU, and L/S units for a typical 20% increase but with a widely rumored clock decrease.

Maybe both AMD/Intel will deliver massive 40%+ gains but maybe they are both ~10%.

Without knowing the details(meaning insider information) high level descriptions can only provide us with very limited information on what it is actually capable of. Both Prescott and Bulldozer looked promising, until it wasnt.

Competitors tend to be also closer to easier other than expected since the leader relaxes a bit to maximize financial gains while the lagger does better due to the natural tendency of people to work harder in times of distress.

Users of both camps expect massive, sweeping wins that will completely obfuscate the other but it happens few, and far between.

Fanboy and any derivative is not permitted in the tech forums. I changed it to an acceptable term. Thanks for your understanding and cooperation.

Mod DAPUNISHER

 

[Hitman928]

Lion Cove is "supposed" to be 33% wider decode with more than 2x uop capability and also an expansion for both FP, ALU, and L/S units for a typical 20% increase but with a widely rumored clock decrease.

Maybe both AMD/Intel will deliver massive 40%+ gains but maybe they are both ~10%.

Without knowing the details(meaning insider information) high level descriptions can only provide us with very limited information on what it is actually capable of. Both Prescott and Bulldozer looked promising, until it wasnt.

Competitors tend to be also closer to easier other than expected since the leader relaxes a bit to maximize financial gains while the lagger does better due to the natural tendency of people to work harder in times of distress.

Users of both camps expect massive, sweeping wins that will completely obfuscate the other but it happens few, and far between.

The quote is that Zen 5 IPC is only a 10% increase. . .

 

[DavidC1]

The quote is that Zen 5 IPC is only a 10% increase. . .

Intel has higher clocks, so them achieving 20% uarch gain but with lower clocks resulting in overall 10% gains and AMD having 10% gains is really the same thing.

Remember when they got 25% overall gains confused with ~10-15% uarch gain and rest in clocks for Zen 4? I don't really see anyone complaining.

 

[Hitman928]

Intel has higher clocks, so them achieving 20% uarch gain but with lower clocks resulting in overall 10% gains and AMD having 10% gains is really the same thing.

Remember when they got 25% overall gains confused with ~10-15% uarch gain and rest in clocks for Zen 4? I don't really see anyone complaining.

Right, but you could still get a clock regression on the AMD side. I thought you were saying that clock regressions could explain the lower than expected overall performance which wouldn’t apply to the said quote. If you are not saying that, I don’t know what to say other than reiterate that a 10% IPC bump for Zen 5 would mean a massive screw up by AMD just based on the things we already know they’ve changed (edit to add) which is hard to believe given AMD’s track record on the CPU side since Zen arrived and also the leaked GB scores which show significantly higher IPC than 10%.

 

[del42sa]

https://www.anandtech.com/show/21380/amd-zen-5based-epyc-turin-is-sampling-silicon-looking-great

"Looking ahead, we are very excited about our next-gen Turin family of EPYC processors featuring our Zen 5 core," said Lisa Su, chief executive officer of AMD, at the conference call with analysts and investors (via SeekingAlpha). "We are widely sampling Turin, and the silicon is looking great. In the cloud, the significant performance and efficiency increases of Turin position us well to capture an even larger share of both first and third-party workloads."

how modern journalism works : https://wccftech.com/amd-zen-5-cpus...ase-more-in-cinebench-r23-single-thread-test/

 

[marees]

It actually gets worse because TOPs requirements is going up soon. Nutella is single-handledly ruining multiple generations of SoCs.

Anyone has a semiaccurate subscription ?

https://www.semiaccurate.com/2024/05/08/microsoft-absolutely-screws-intel-and-amd-over-ai-pcs/

 

[StefanR5R]

The fact that this guy is a product manager at Lenovo has some credibility to me.

Is he?
And if he was, what do you think, did he post fantasized numbers, or did he break an NDA?
And if he was and did one of the two, how much longer will his contract with Lenovo last?

 

[Joe NYC]

Frankly, the existence of the 128-bit "LP" model there puts this whole slide under doubt. That one doesn't make any sense in the slightest, that's just simply far too cut down.

That one implies to me that MLID was told something about "LP" with regards to Strix Halo from an OEM or someone else, and assumed it meant there would be a significantly cut down model, and just made up the specs (by just halving everything).

Agreed (as I already said in my original post). Cutting down the configuration so much, given there will be only 1 piece of (quite expensive) silicon of SoC. Number of partially recovered dies, from defects, would be too small (IMO) to launch the lowest end model. It would take millions of Strix Halo dies to recover enough.

LP, as it is on the slide, also doesn't make a lot of sense if there are LP Zen 5 cores on SoC.

 

[Fjodor2001]

Anyone has a semiaccurate subscription ?

https://www.semiaccurate.com/2024/05/08/microsoft-absolutely-screws-intel-and-amd-over-ai-pcs/

From that article:

"It looks like Microsoft is turning on Intel and AMD over AI PCs. [...] Microsoft is pushing ‘AI’ not to benefit the users but to offload some very expensive costs from their datacenters to the user."

Could this mean less sales for nVidia of AI related HW to datacenters going forward, since (part of) the work will be done on client PCs instead?

Looking at the nVidia projections, their arrow has so far been pointing upwards for a long time with no end in sight.

 

[DisEnchantment]

"Microsoft is pushing ‘AI’ not to benefit the users but to offload some very expensive costs from their datacenters to the user."

Could this mean less sales for nVidia of AI related HW to datacenters going forward, since (part of) the work will be done on client PCs instead?

No, it means Charlie is deranged.

 

[APU_Fusion]

We are back to Zen 5 having 10% IPC uplift against Zen 4? 😂😂

Zen 5 is now slower than Zen 4 with a 10% IPC gain and 15% clock regression while being 46% faster in cinvench multi-core test while being slower and faster than raptor lake and arrowlake depending on baseline or performance option set for intel cpus.

For the record, I don’t buy for a second a major reworked zen 5 that is wider only gets 10% uplift. If so, they better bulldoze their profits and work on 10% faster clock Zen 4 😜

 

[Joe NYC]

The Zen5 improves 10% relative to the Zen4 IPC, and >10% in the CR23 1T test. Also, Win10 enthusiasts, please note that starting with Strix Point, AMD will not be providing Win10 drivers.

https://weibo.com/7517897889/O8cQlaMP3

We will see about doubling of Zen 5% IPC improvement.

But, regarding Windows 10, it is good to remember that scheduler that can deal with different types of cores has not been added to Windows 10, that it is Windows 11 only feature.