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

[DisEnchantment]

Speculate at will

 

[StefanR5R]

Genoa-X [...] was priced only like 10% extra by the way,

FWIW, retail prices in Germany of Milan-X and Genoa-X have always been considerably more expensive than non-X parts, whenever I looked them up.

 

[Joe NYC]

Turin-X scrapped

It disappeared from the 2024 server presentations.

But we have all of 2025 and part of 2026 to go through before Zen 6. I think the odds are that AMD will introduce Turin-X within a year.

Just FYI, there was a cloud vendor that recently standardized all their servers on Genoa-X. My guess is they would not have done it if Turin-X was indeed scrapped.

 

[Win2012R2]

FWIW, retail prices in Germany of Milan-X and Genoa-X have always been considerably more expensive than non-X parts, whenever I looked them up.

In the UK I've seen 9k (inc VAT) for 96 cores normal and 10k-ish for X version (so around 10%), it always struck me as a very small difference, perhaps hyperscalers get far bigger discounts on normal versions than X, and form that point of view X is way too expensive.

if Turin-X was indeed scrapped

Dr Ian asked AMD about Turin-X and they told him it wasn't planned this time around, so maybe not scrapped, but officially it's not appearing, this is annoying given that X version would have kept higher clocks too.

 

[StefanR5R]

Mindfactory publishes sales numbers of Ryzen 7 9800X3D now.
(They withheld these sales figures before.)
https://www.computerbase.de/news/pr...et-das-geheimnis-um-die-verkaufszahlen.90825/ (in German)

9800X3D ....... >8.650 units sold
9700X ............ >2.510 units sold (total of boxed and tray units)
9600X ............... >895 units sold (total of boxed and tray units)
9900X ............... >800 units sold
9950X ............... >780 units sold

or in short,
9800X3D ................................... >8.650 units sold
all other 9000 series parts ....... >4.985 units sold

when Computerbase took their screenshot.

Edit:
Caveat, while 9800X3D could and can be had in Germany almost only at Mindfactory all the time since launch (other retailers evidently received far fewer 9800X3Ds or none at all), the other SKUs are of course available at several retailers. Therefore the above sales ratio of 9800X3D vs. all others is not representative for the entire German or European Ryzen 9000 series sales.

 

[Joe NYC]

Dr Ian asked AMD about Turin-X and they told him it wasn't planned this time around, so maybe not scrapped, but officially it's not appearing, this is annoying given that X version would have kept higher clocks too.

I missed that.

But Mi300c was also "cancelled" until it appeared in Microsoft Azure.

 

[igor_kavinski]

I wonder if the bonding process is reversible so they can take out the V-cache dies from unsold CPUs and divert the V-cache dies to CPUs in demand.

 

[Win2012R2]

But Mi300c was also "cancelled" until it appeared in Microsoft Azure.

Yeah, and it's far easier to conjure Turin-X since 3D chiplets are available, perhaps Threadripper will get it, would be nice

I wonder if the bonding process is reversible so they can take out the V-cache dies from unsold CPUs and divert the V-cache dies to CPUs in demand.

Much easier to do near Just In Time for assembling CPUs, keep low unsold inventory to avoid it being on the books, AMD seems to be very tight in this respect - guess they got burnt way too many times

 

[Joe NYC]

I wonder if the bonding process is reversible so they can take out the V-cache dies from unsold CPUs and divert the V-cache dies to CPUs in demand.

I don't think there is such a thing as unsold V-cache CPU.

And no, I am nearly certain the bonding process is not reversible. There would be no point. If too many are made, just make less for next quarter and sell off the excess.

 

[StefanR5R]

Caveat, while 9800X3D could and can be had in Germany almost only at Mindfactory all the time since launch (other retailers evidently received far fewer 9800X3Ds or none at all), the other SKUs are of course available at several retailers. Therefore the above sales ratio of 9800X3D vs. all others is not representative for the entire German or European Ryzen 9000 series sales.

PS, but there is a caveat to this caveat: All 9000Xs (with the exception of 9950X) have been "in stock" at Mindfactory all the time, while every 9800X3D batch which arrived at Mindfactory was sold out immediately when or rather before it arrived. Nevermind that the sales figures are from ~2 months for 9800X3D vs. ~7 months for the others.

 

[LightningZ71]

Just thinking out loud here...

If having a processor with a single v-cache CCD and a normal CCD is the best of all possible worlds, why hasn't AMD made a Threadripper with a single VCache CCD with the rest being normal CCDs? The scheduling problem would not seem to be an issue. It's a premium product and would give it's users the best possible experience. Why haven't they done it?

 

[StefanR5R]

Just thinking out loud here...

If having a processor with a single v-cache CCD and a normal CCD is the best of all possible worlds,

Well, *if*. But it isn't. It's a terrible kludge whose saving grace is that it costs less.

Threadripper

This is a product for workloads which access large datasets, with a mix of low and high parallelism. For the program portions with low parallelism, very high CPU core clock speed and high memory bandwidth are desired, and for the program portions with good parallelism, good clock speed and very high memory bandwidth are desired. (Generally speaking. Obviously the weighting depends on the particular application and dataset.) *Some* of such applications would benefit from 96 instead of 32 MB L3$/CCX... if all CCXs had it.

The scheduling problem would not seem to be an issue.

But it is. The scheduler can't predict how asymmetric cache can be utilized best.

 

[gdansk]

If having a processor with a single v-cache CCD and a normal CCD is the best of all possible worlds

It's the best for user workloads. Not whatever it is TR/Epyc people do. If you need that many cores you're already prioritizing so far to MT there's little benefit to asymmetry.

Also the opposite would probably be better. 1 normal CCD as 0 and the rest as vcache. The idea of having 1 lower latency, higher frequency CCD is that many user applications (and even some games) perform better on it and aren't heavily threaded.

 

[Joe NYC]

deleted

 

[Philste]

Haven't looked in here for like 2-3 months, so sorry if all of the following has been said or discussed:

Looking at the leaks/rumors about Krackan/Kraken, future roadmaps and judging by Strix' performance it seems to me that 2025 (and maybe 2026) AMD mobile CPU lineup looks pretty dull, or am I wrong?

Everyone is kinda hyped about Strix Halo, but the rest? Strix is decent, but in my eyes it's as underwhelming as Vanilla ZEN5 was. Yeah, CPU ST is +15%, MT +20-25% and iGPU +15%. I just expected more with 50% more threads and 33% more CUs. It's a decent uplift overall but not for a price that is around double than that of it's predecessors.

Now people say Strix is sitting a Tier above the old APUs (Phoenix, Hawk Point) and Krackan is the true successor. But if Strix has ~20% Perf uplift, what does Krackan have? If nothing magical happens it will probably get completely trashed by reviewers and, after that, customers.

4+4 (or 5+3) CPU and 8CU iGPU suggest that it will be at Phoenix Performance at best for CPU MT and GPU. CPU ST will see the standard 15-16% ZEN5 uplift of course. So it's most likely same performance as the 2 year old product with that big NPU. Good thing that MS messed up all the Copilot Stuff and even the uninformed people now know that it's unnecessary stuff.

With all this, Krackan seems to be AMDs worst Mainstream APU in years. AMD had the best APU since at least Renoir. Of course it's nearly impossible to break Intels OEM Dominance but at least Enthusiasts knew: AMD=better in most Cases.

Now with Arrow Lake Intel it looks like Intel has the most competitive Product in Years. Of course it failed miserably in Desktop, but the data looks promising for mobile CPUs. Efficiency in lower clock/power Regions seems to be pretty good, iGPU is also set to be more than decent. And MT Performance, which is LNLs main weakness isn't an issue as well. ARL 24T is kinda competitive to ZEN5 32T in MT so 16T ARL-H should be kinda competitive with 24T Strix.

ARL-H main weakness as "MTL 2.0" will be battery Life, that is where Strix will have a lead. But all in all Strix and Krackan look pretty dull compared to the lead that AMD had with Renoir to Phoenix.

And the worst part is that ZEN6 is more and more shaping up to be a H2 26 product with APUs possibly at CES 2027? Leaks are saying Strix/Krackan Refreshes for 2026. And AMDs refreshes over the last years have been one of the lamest things ever, even Intels Refreshes were great compared to this. It will be like 200MHz at best for CPU and GPU.

Meanwhile Intel should at least launch PTL and even if it's late it should still have at least a half year lead over everything ZEN6. From pure specs I would say PTL with 4+8+4 and 12Xe³ should beat Strix in most Things, while potentially getting close to LNL Battery life.

Doesn't that look concerning overall?

 

[gdansk]

Doesn't that look concerning overall?

And also concerning is the die size of KRK. It's bigger than Phoenix and all it has to show for it is 10% better 1T. Which is better than nothing but not compelling if it costs more. The NPU is ~0 value.

 

[Philste]

And also concerning is the die size of KRK. It's bigger than Phoenix and all it has to show for it is 10% better 1T.

Yeah, I mean the NPU is kinda big, but all the cutdowns in GPU and CPU and it's still a bigger Die? If the ZEN5c cores are capped at 3.3GHz like in Strix, it will probably start to fall behind Phoenix starting at 18-20W and above.

But still, I think Microsoft messing up their Copilot Stuff is what probably breaks KRKs neck. Not because it's actually useful but because they could've at least PRETENDED that it is useful. Now it's basically over, nobody will believe that stuff.

 

[Joe NYC]

Haven't looked in here for like 2-3 months, so sorry if all of the following has been said or discussed:

Looking at the leaks/rumors about Krackan/Kraken, future roadmaps and judging by Strix' performance it seems to me that 2025 (and maybe 2026) AMD mobile CPU lineup looks pretty dull, or am I wrong?

Everyone is kinda hyped about Strix Halo, but the rest? Strix is decent, but in my eyes it's as underwhelming as Vanilla ZEN5 was. Yeah, CPU ST is +15%, MT +20-25% and iGPU +15%. I just expected more with 50% more threads and 33% more CUs. It's a decent uplift overall but not for a price that is around double than that of it's predecessors.

Now people say Strix is sitting a Tier above the old APUs (Phoenix, Hawk Point) and Krackan is the true successor. But if Strix has ~20% Perf uplift, what does Krackan have? If nothing magical happens it will probably get completely trashed by reviewers and, after that, customers.

4+4 (or 5+3) CPU and 8CU iGPU suggest that it will be at Phoenix Performance at best for CPU MT and GPU. CPU ST will see the standard 15-16% ZEN5 uplift of course. So it's most likely same performance as the 2 year old product with that big NPU. Good thing that MS messed up all the Copilot Stuff and even the uninformed people now know that it's unnecessary stuff.

With all this, Krackan seems to be AMDs worst Mainstream APU in years. AMD had the best APU since at least Renoir. Of course it's nearly impossible to break Intels OEM Dominance but at least Enthusiasts knew: AMD=better in most Cases.

Now with Arrow Lake Intel it looks like Intel has the most competitive Product in Years. Of course it failed miserably in Desktop, but the data looks promising for mobile CPUs. Efficiency in lower clock/power Regions seems to be pretty good, iGPU is also set to be more than decent. And MT Performance, which is LNLs main weakness isn't an issue as well. ARL 24T is kinda competitive to ZEN5 32T in MT so 16T ARL-H should be kinda competitive with 24T Strix.

ARL-H main weakness as "MTL 2.0" will be battery Life, that is where Strix will have a lead. But all in all Strix and Krackan look pretty dull compared to the lead that AMD had with Renoir to Phoenix.

And the worst part is that ZEN6 is more and more shaping up to be a H2 26 product with APUs possibly at CES 2027? Leaks are saying Strix/Krackan Refreshes for 2026. And AMDs refreshes over the last years have been one of the lamest things ever, even Intels Refreshes were great compared to this. It will be like 200MHz at best for CPU and GPU.

Meanwhile Intel should at least launch PTL and even if it's late it should still have at least a half year lead over everything ZEN6. From pure specs I would say PTL with 4+8+4 and 12Xe³ should beat Strix in most Things, while potentially getting close to LNL Battery life.

Doesn't that look concerning overall?

If Kracken has the same clock speed, it will have the same ST performance as Strix Point. MT performance above 8 cores is a "don't care" for mainstream market.

If all cores are on the same ring bus, then in low power, the Zen 5c cores will have access to double L3 and will more seamlessly switch from low power 5c cores to full Zen 5 cores running at full speed, while keeping the L3 cache content. So, it may perform better on mix apps outside of trivially parallel workloads.

OTOH, if Kraken has 2 ring buses, 2 cluster of cores with separate L3s, or no L3 at all for dense cores than it is kind of f***d.

Not pretending to be a "premium" chip, hopefully, it will find its way to laptops with no dGPU and is priced attractively. The "premium" status of Strix Point kind of killed it as laptop vendors paired it with nVidia GPUs and price became astronomical.

The 2 new laptop chips, Kracken and Strix Halo, in laptops without dGPU can offer very attractive value and performance proposition.

 

[Mopetar]

There is one answer to this, and as a former scalper myself (sorry, PS3 launch), I can confidently declare--- NEVER pay above MSRP for tech. F***k 'em, dont do it. A little self control will put a world of hurt on scalpers. They can keep their stock until AMD restocks, and restocks again until I get MSRP or better. This is the way.

This doesn't work as evidenced by reality. The best way to deal with this is for the company that makes the product to sell the initial release in an auction format where the people willing to pay the highest prices receive the product first.

People absolutely hate this and accuse companies of being greedy, but the reality is that those prices will ultimately be paid due to scalpers. What you need to ask yourself is who should get any excess money to be paid above MSRP, the company making the product or the scalpers who will invariably exploit that difference?

People who want to only pay MSRP will eventually be able to once everyone willing to pay more has been satisfied. There are some people who will trade time for money and stand in line for release sales (I did this in my younger years) but unless your time is worth nothing to you, you're still paying a cost.

I'd much rather see a company rewarded for a successful product and the additional money paid by early adopters can be invested into increased production to get out more product so that the market price becomes MSRP more quickly. This also eliminates scalpers as they can only exist due to the difference in price between MSRP and what the market will actually pay.

 

[Philste]

Not pretending to be a "premium" chip, hopefully, it will find its was to laptops with no dGPU and is priced attractively.

But even with single CCX it won't be much faster than Phoenix, especially not in the GPU departement. Phoenix and Hawk devices have been priced attractively for 2 years now. In my eyes KRK is pretty much doomed to Fail.

 

[Josh128]

This doesn't work as evidenced by reality. The best way to deal with this is for the company that makes the product to sell the initial release in an auction format where the people willing to pay the highest prices receive the product first.

People absolutely hate this and accuse companies of being greedy, but the reality is that those prices will ultimately be paid due to scalpers. What you need to ask yourself is who should get any excess money to be paid above MSRP, the company making the product or the scalpers who will invariably exploit that difference?

People who want to only pay MSRP will eventually be able to once everyone willing to pay more has been satisfied. There are some people who will trade time for money and stand in line for release sales (I did this in my younger years) but unless your time is worth nothing to you, you're still paying a cost.

I'd much rather see a company rewarded for a successful product and the additional money paid by early adopters can be invested into increased production to get out more product so that the market price becomes MSRP more quickly. This also eliminates scalpers as they can only exist due to the difference in price between MSRP and what the market will actually pay.

Oh it definitely can work. Look at the Zen 5 launch. Prices sold at MSRP at launch even though greedy scalpers attempted to sell them at a premium on Ebay. Then prices plummeted on all Zen 5 SKUs and any scalper who bought at MSRP and didnt offload it before prices dropped lost money. Now if a product is not produced in a sufficient amount, has high demand, and is sold at a decent price, shortages will occur and scalpers will scalp. AMD could have, and IMO should have, recognized that and priced the chip much higher than they did, but that also comes with repercussions and bad PR. Intels 8 core 5960X sold at $1000 around 10 years ago. AMD's single core Athlon FX-57 sold for $1000 20 years ago. Both were halo chips with zero serious challengers. Its AMDs own fault they didnt price 9800X3D higher. Their marketing dept failed to properly guage demand. They could have MSRP'ed it at $599 easily and still sold their first few runs instantly. Opportunities to go unchallenged are usually few and far between for them. I guess they have gone soft.

Its OK though, as I suspect Huang has not gone soft and is about to prove that to everyone with the RTX 5000 series. He knows whats up. Best strap yourself in.

 

[JustViewing]

Dr Ian asked AMD about Turin-X and they told him it wasn't planned this time around, so maybe not scrapped, but officially it's not appearing, this is annoying given that X version would have kept higher clocks too.

Maybe not enough interest for previous X models. And for those target markets, once upgraded they won't upgrade for another 4/5 years.

 

[Win2012R2]

Maybe not enough interest for previous X models. And for those target markets, once upgraded they won't upgrade for another 4/5 years.

That's exactly what they said... BS in my opinion

 

[Joe NYC]

But even with single CCX it won't be much faster than Phoenix, especially not in the GPU departement. Phoenix and Hawk devices have been priced attractively for 2 years now. In my eyes KRK is pretty much doomed to Fail.

Well, it will be incrementally faster than all of its competition (Phoenix, Hawk, Meteor, Raptor), will be the only one of these with Co-Pilot sticker, roughly on par with Lunar Lake, cheaper to produce than MTL, LNL.

 

[poke01]

The 2 new laptop chips, Kracken and Strix Halo, in laptops without dGPU can offer very attractive value and performance proposition.

Halo won't be cheap.

 

[igor_kavinski]

Its AMDs own fault they didnt price 9800X3D higher. Their marketing dept failed to properly guage demand. They could have MSRP'ed it at $599 easily and still sold their first few runs instantly. Opportunities to go unchallenged are usually few and far between for them. I guess they have gone soft.

And give people more excuses to whine about AMD products? I have to admit that I look to AMD for bang for buck. That's always been their allure to me. Take that away and I'll be left looking towards the budget offerings of both Intel and AMD and choose what seems to be the best fit for my needs without blowing a hole in my bank account. I prefer their current strategy. Give consumers stuff at reasonable prices while charging a premium from their enterprise customers. I think that kind of balance helps them stay the "good guys" of this industry. As for 9800X3D, if anyone's paying more than MSRP for it, I truly feel sorry for them. It means they have no patience and they probably have a lot more problems to deal with in life than overpriced CPUs.