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

[DisEnchantment]

Speculate at will

 

[Thunder 57]

Alder Lake had a huge IPC gain. The 10nm silicon was holding back Intel. It was inferior to the 5nm TSMC silicon that Zen 4 was built on.

What do you think "Intel 7" is? If MTL is any indication, ARL will dissappoint.

 

[H433x0n]

My concern is that Zen 5 will be on TSMC 5nm which they call 4nm because it's more advanced than N5 silicon. It trails TSMC 3nm in efficiency by 15-20%. Zen 5 was supposed to be on the 3nm silicon but TSMC has a shortage of 3nm or AMD elected to wait it out because they are cheap.

That's not the case at all. The perf/watt improvement for N3B over N4P is ~5%, it's basically a wash.

 

[Hans Gruber]

What do you think "Intel 7" is? If MTL is any indication, ARL will dissappoint.

Remember the Athlon 1800+ (equal to or better than Intel 1800mhz). I go by silicon Intel 7 was 10nm. Arrow Lake will be on 5nm silicon. Intel said they will gain silicon superiority with 18A over TSMC 2nm. I am not saying they will or won't. They said 18A will be 10% more efficient than 20A. AMD was supposed to be on N3 for Zen 5. N4 is not bad but 10-15% less efficient than N3. On 3nm, AMD could have a 45w eco mode and reduce TDP. I personally think a low power efficient processor is as important as a top performing processor. People who run servers boast of the energy efficiency and performance of their AMD servers compared to the Xeon servers they had in the past.

If AMD releases Zen 5 early. Then they will take the performance crown in all categories. That is an important factor to consider.

 

[Hans Gruber]

That's not the case at all. The perf/watt improvement for N3B over N4P is ~5%, it's basically a wash.

It's N4 that Zen 5 will be riding the lightning on. N4P is more advanced than N4. I do wish TSMC would better publish the performance differences between their processes.

 

[adroc_thurston]

It's N4 that Zen 5 will be riding the lightning on

nope.
lmao.

Remember the Athlon 1800+ (equal to or better than Intel 1800mhz). I go by silicon Intel 7 was 10nm. Arrow Lake will be on 5nm silicon. Intel said they will gain silicon superiority with 18A over TSMC 2nm. I am not saying they will or won't. They said 18A will be 10% more efficient than 20A. AMD was supposed to be on N3 for Zen 5. N4 is not bad but 10-15% less efficient than N3. On 3nm, AMD could have a 45w eco mode and reduce TDP. I personally think a low power efficient processor is as important as a top performing processor. People who run servers boast of the energy efficiency and performance of their AMD servers compared to the Xeon servers they had in the past.

nodes don't matter.
cores do.

On 3nm, AMD could have a 45w eco mode and reduce TDP

buy a fire range and be happy.

 

[Abwx]

What do you think "Intel 7" is? If MTL is any indication, ARL will dissappoint.

They have a theorical core count advantage with MTL since 6 + 8 amount to 10P, so that could largely offset the process disadvantage against a 8C Hawk Point.

At 25% more cores you could reduce average frequency by 20% andbe at 0.6x the power of their Intel 4 process and still have the same throughput as a 8 P cores, yet they fail even at same perf, so it s not the process that is the main culprit but an uarch with notably lower intrinsical efficency.

It's N4 that Zen 5 will be riding the lightning on. N4P is more advanced than N4. I do wish TSMC would better publish the performance differences between their processes.

They published enough numbers but apparently it s not up to anyone to understand the data, to summarise N4P and N4X have respectively 28% and 39% better perf/watt at isofrequency than their N5, that s a big and sufficent margin to accomodate Zen 5 higher throughput/Hz even if it s less efficient uarchitecturaly wise.

 

[adroc_thurston]

even if it s less efficient uarchitecturaly wise.

but it's more.
AMD does 1:0.8 IPC:Cac ratio.

 

[Hans Gruber]

They have a theorical core count advantage with MTL since 6 + 8 amount to 10P, so that could largely offset the process disadvantage againsta 8C Hawk Point.

At 25% more cores you could reduce average frequency by 20% and 0.6x the power of their Intel 4 process and still have the same throughput as a 8 P cores, yet they fail even at same perf, so it s not the process that is the main culprit but an uarch with notably lower intrinsical efficency.


They published enough numbers but apparently it s not up to anyone to understand the data, to summarise N4P and N4X have respectively 28% and 39% better perf/watt at isofrequency than their N5, that s a big and sufficent margin to accomodate Zen 5 higher throughput/Hz even if it s less efficient uarchitecturaly wise.

We are not comparing N4P to N5. N5 is the base node of 5nm. Zen 5 was supposed to be based on N3. Compare N3 to N4. N4P or whatever Nvidia uses for their 40 series cards is the most advanced 5nm silicon which TSMC calls 4nm. It's safe to assume the 4nm nomenclature is done to signify the superiority of the silicon vs. standard N5.

 

[adroc_thurston]

It's safe to assume the 4nm nomenclature is done to signify the superiority of the silicon vs. standard N5.

it's all very incremental.
nodes are means to an end.

 

[gdansk]

Blackwell is on "4NP" so uhh RIP to the theory that Ada's "4N" is equivalent to N4P.
Not that Nvidia branding matters much.

 

[Markfw]

nevermind

 

[Thunder 57]

Remember the Athlon 1800+ (equal to or better than Intel 1800mhz). I go by silicon Intel 7 was 10nm. Arrow Lake will be on 5nm silicon. Intel said they will gain silicon superiority with 18A over TSMC 2nm. I am not saying they will or won't. They said 18A will be 10% more efficient than 20A. AMD was supposed to be on N3 for Zen 5. N4 is not bad but 10-15% less efficient than N3. On 3nm, AMD could have a 45w eco mode and reduce TDP. I personally think a low power efficient processor is as important as a top performing processor. People who run servers boast of the energy efficiency and performance of their AMD servers compared to the Xeon servers they had in the past.

If AMD releases Zen 5 early. Then they will take the performance crown in all categories. That is an important factor to consider.

Every generation you seem to say AMD should release Zen x early or AMD messed up by launching Zen x too late. It's not like they are going to sit on a bunch of inventory and not launch it until whenever because... reasons. It is ready when it is ready.

Might as well say that Intel should release ARL early, cause they sure could use it.

 

[BTRY B 529th FA BN]

What's this Zen5 Execution Engine leak about? AVX 512 and things of the sort

 

[igor_kavinski]

Intel really has entered the dying company cycle: "if I just push harder with the same thing that worked 10 years ago but absolutely doesn't work anymore, surely I'll get back to N°1"
Pat was supposed to save Intel's soul, not give the last rites.

It's not like this hasn't happened before. Intel's doom was prophesized when they clung onto Pentium 4 harder than a leech inside a donkey's arse and ignored the donkey's blood curdling screams. Athlon 64 should've made AMD the No.1 semicon company in the world. Sadly, all it did instead was make enthusiasts and the gaming crowd happy for a few years who then promptly ditched AMD for Banias, Dothan, Merom and Conroe. AMD then promptly drove themselves into the ground with Bulldozer and Piledriver. I don't know why people bought those despite being priced lower. I guess electricity was cheaper back then and few, if any, were sounding the global warming alarms back then. AMD also fooled them into believing they were selling them more cores and paid for that deception by having to settle a class action lawsuit. Funnily, this is what a commenter wrote back then (https://www.anandtech.com/show/14804/amd-settlement):

By filing a meritless lawsuit unscrupulous lawyers force the hand of companies who could rack up tens of millions in litigation costs and then be subjected to the technical ignorance of a jury with the mentality to pay a woman a million dollars for pouring hot coffee on her crotch in a moving vehicle. This outrageous behavior that only exists in the U.S. judicial system is named: Jackpot Justice where paid liars dupe a jury into believing that it's always someone else's fault and that any corporation with deep pockets should pay for any perceived injustice even when as in the AMD CPU core case there was none. If you can't understand what a CPU core is then you should educate yourself instead of filing frivolous lawsuits that make bottom feeders wealthy. The real injustice in this case is that AMD is raped out of $12 million by devious players manipulating the system for profit. If $35 makes you "whole" in judicial terms then you were never "damaged" in the first place by buying an AMD CPU - which is precisely the case.

I would love to know his thoughts on the billions raped out of AMD by Intel during the P4 era

If it weren't for AMD's measly and pathetic Jaguar CPU core powering consoles, they could've been decimated and relegated to making dirtcheap embedded CPUs for eternity. So no. Intel has shown that they can sell crap over and over again, with absolute frickin' impunity coz they are the IBM of the modern corporate world. And as some others have noted, their biggest threat isn't AMD. It's ARM and RISC-V and anything else that pops out of nowhere and suddenly takes over the world. Heck, Apple could give them a much worse time than AMD if they weren't so obsessed with being a lifestyle company. Intel WILL make a comeback. It's just a matter of when, not if.

 

[Ghostsonplanets]

The whole Zen 5 line-up, from DT to Mobile seems very exciting. But Strix Halo is the one I'm really most keen on seeing it. Will be interesting to see if OEMs will have designs with it. Could be a killer play from AMD as the costs on Laptop system design with an Mega APU should be much lower than the classic CPU + dGPU.

And as some others have noted, their biggest threat isn't AMD. It's ARM and RISC-V and anything else that pops out of nowhere and suddenly takes over the world.

This is a good point. AMD maintaining x86 relevance is good for Intel because it means they're keeping others players at bay and the market is still dominated by x86 designs. If ARM or RISC-V start to dominate and players change their tooling, software, etc to seriously support these ISAs, it will be much harder to recover market share.

Intel WILL make a comeback. It's just a matter of when, not if.

That I really don't know. I think Intel position currently is way more precarious compared to P4 days. They're also not nearly as dominant nor have the insane reach and grip in the market players as they used to.

Intel might come back. It's their aim, specially with things like "Exciting new high performance project" and supposedly Nova Lake. But I don't really see AMD being displaced and relegated like how it happened back then. Only if AMD suddenly stalls, but we have no reason to think so.

 

[Wolverine2349]

It's Intel who need to show ability to produce meaningful IPC gains without going for more and more bloat.
GLC and Zen 4 shouldn't even be close on paper on most metrics.

Are you saying Golden Cove and Zen 4 should not be close on paper in most metrics of IPC in that Golden Cove should be far superior with its resources and being 6 wide to Zen 4 4 wide, but its branch predictor is so much worse that Zen 4 and Golden Cove are almost even because it has more bloat but is even with Zen 4 clock normalized in IPC or barley wins?

 

[Goop_reformed]

Are you saying Golden Cove and Zen 4 should not be close on paper in most metrics of IPC in that Golden Cove should be far superior with its resources and being 6 wide to Zen 4 4 wide, but its branch predictor is so much worse that Zen 4 and Golden Cove are almost even because it has more bloat but is even with Zen 4 clock normalized in IPC or barley wins?

You're basically making a bloated version of what he said

 

[Anhiel]

Intel's uarch has slightly less latency internally and since Intel's nodes traditionally are aimed for more clockspeed while TMCS more efficiency... you can see that in past tests when clockspeed for both are OCed the GLC-type's show less diminishing ISO performance than Zen4. RPL has improved threaddirector & e-core latency. It seems certain LNC will keep that little advantage. AMD will keep infinity fabric & cache as a counter but it has obviously connectivity limits.

Anyhow, I don't like the way how everything is going to demand NPU power. At this rate no amount will be enough to satisfy all the MS OS and apps requirements. For a second I do feel the need for a co-processor-esque solution. Welp, I suppose it'll eventually end up being an extra large tile/chiplet.

 

[adroc_thurston]

Intel's uarch has slightly less latency internally

average mispredict and a whole lot of exeuction latencies are higher on GLC. lmao.

 

[Mahboi]

It's not like this hasn't happened before. Intel's doom was prophesized when they clung onto Pentium 4 harder than a leech inside a donkey's arse and ignored the donkey's blood curdling screams. Athlon 64 should've made AMD the No.1 semicon company in the world. Sadly, all it did instead was make enthusiasts and the gaming crowd happy for a few years who then promptly ditched AMD for Banias, Dothan, Merom and Conroe. AMD then promptly drove themselves into the ground with Bulldozer and Piledriver.

If it weren't for AMD's measly and pathetic Jaguar CPU core powering consoles, they could've been decimated and relegated to making dirtcheap embedded CPUs for eternity. So no. Intel has shown that they can sell crap over and over again, with absolute frickin' impunity coz they are the IBM of the modern corporate world.

My gut feeling in the case of Intel though, is that their master problem is in getting lean solutions fused together into a coherent whole. It's not just good cores anymore, it's cores, advanced packaging, all kinds of cache strategies, uncore performance, etc.
I'm not worried about Intel because their cores are bad. I'm worried because the cores are bloated, the wattage is ridiculous, the uncore is far more costly than AMD's and yet MTL has somehow not whipped PHX's butt. If anything MTL should have trounced through in low power efficiency, or beaten it in GPU since RDNA 3 has its power problem. It did none, and even has the luxury of falling apart in both, with terrible low power control and poor graphics drivers.

Intel seems to lack not in one domain but rather seems to be slipping in every way, like an old man whose business isn't popular anymore and who just tries harder than before. Nothing says they can't come back, but if you take the current projections from adroc, then Zen 5 is going to establish dominance until 2026. After which Zen 6 will. Zen 6 will also be about the uncore more than the core, AMD's dropping the Zen 2 era I/O die and a lot of cheaper, older packaging designs. What happens when the circle will be complete on the Red side, and every part of the Ryzen package is modern? Up to now Intel threw silicon interposers at a competitor that ran traces in the PCB. When everyone will have the same level of costly tooling, how behind will they really be?

 

[FlameTail]

I am talking out of the gate when Zen 4 was released. They added the eco mode later which was good. With power efficiency, Intel cannot touch AMD. That was the big selling point for AMD. If AMD was on N3P out of the gate with Zen 5.

You want AMD to release Zen5 next year(2025)??

N3P mass production will start only next year. Heck, N3E mass production has barely started.

 

[FlameTail]

It's N4 that Zen 5 will be riding the lightning on. N4P is more advanced than N4. I do wish TSMC would better publish the performance differences between their processes.

Take this,

And this:

 

[Tuna-Fish]

We are not comparing N4P to N5. N5 is the base node of 5nm. Zen 5 was supposed to be based on N3. Compare N3 to N4. N4P or whatever Nvidia uses for their 40 series cards is the most advanced 5nm silicon which TSMC calls 4nm. It's safe to assume the 4nm nomenclature is done to signify the superiority of the silicon vs. standard N5.

You seem very confused. nVidia 40 series did not use N4P for the very simple reason that they entered production nearly a year before N4P was available. They used 4N, which was one of the early 5nm processes, and not anywhere near the most advanced 5nm silicon. 4NP was developed later, using knowledge accumulated during 3N development, and is both faster and more economical to manufacture. TSMC is pushing all clients to use N4P, like they earlier pushed everyone to use N6, using favorable prices. Blackwell is using N4P, and so will Zen5, and frankly every other newly developed 5nm-class product.

 

[poke01]

So on the AMD side Zen 7 for 2nm with BSPDN, maybe?

 

[FlameTail]

There will be no Zen 7.

Zen6 is The Last Zenbender.