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

[DisEnchantment]

Speculate at will

 

[TESKATLIPOKA]

Fire range is already niche, only used with high end GPUs. you need to compare the cost to Phoenix + dGPU. And Phoenix wins there

But Phoenix is Zen4, this will be Zen5 and let's not forget that Phoenix has only 1/2 of cores.
Strix Point+dGPU would be a better comparison in my opinion and I think we will see such laptops.

 

[SteinFG]

But Phoenix is Zen4, this will be Zen5 and let's not forget that Phoenix has only 1/2 of cores.
Strix Point+dGPU would be a better comparison in my opinion and I think we will see such laptops.

Zen 4 is still great with the level of GPU that halo will be eqipped with

 

[TESKATLIPOKA]

Zen 4 is still great with the level of GPU that halo will be eqipped with

I don't disagree. Honestly, even for a stronger dGPU It wouldn't be a bottleneck. 7490HS + RTX4900
But let's be honest, Zen4 is almost an old tech.

 

[Timmah!]

I don't disagree. Honestly, even for a stronger dGPU It wouldn't be a bottleneck. 7490HS + RTX4900
But let's be honest, Zen4 is almost an old tech.

If i did not have fairly new rig and the pricing was not so brutal, i would still take 7970x zen4 over 16C zen5 next year. Even 7960x i guess.

BTW the pricing of the Threadripper boards annoys me. Or better said, they are OK, what annoys me my x670e motherboard from last year did cost as much as these TR boards. Despite having less lanes, less slots, less everything.

 

[igor_kavinski]

If i did not have fairly new rig and the pricing was not so brutal, i would still take 7970x zen4 over 16C zen5 next year. Even 7960x i guess.

Check Newegg. Epyc 7373X with Tyan mobo is kinda affordable. That huge cache can come in handy in your workloads, I'm sure.

 

[igor_kavinski]

BTW the pricing of the Threadripper boards annoys me. Or better said, they are OK, what annoys me my x670e motherboard from last year did cost as much as these TR boards. Despite having less lanes, less slots, less everything.

I understand your pain. X670E is way, way overpriced for what it offers, especially when there are $150 mobos from Gigabyte/ASROCK that can run a 7950X without much issue.

 

[Joe NYC]

As I said we don't know how It will really perform or cost compared to CPU+dGPU.
I also wouldn't be so sure that Strix Halo is really cheaper to make than Fire Range + 7600. Fire Range will keep the IOD from Dragon range If I am not mistaken.

That's an interesting comparison. Cost may be in the same range (thanks to 7600 being cheap, on N6) but likely will perform worse and have higher power consumption, bigger size.

The SoC of Strix Halo is going to be quite big and it may be sub-optimal to have the most demanding (GPU) and the least demanding (IO) functions on the same N3E die.

I am wondering what the die size will be. Probably > 200 mm2.

Intel has nothing comparable to Strix Halo. It can be only CPU+dGPU combo, so fire Range should be Its competitor.

What I meant is that it is very important to Intel to have the best performing CPU in client, using existing tools. If +50W is going to give the Intel part extra 5% performance, Intel will do it. If a re-spin gains 2.5%, Intel will do it.

If Intel had V-Cache and could add it to its Halo part, it would be a 100% guarantee that Intel would do it, Intel would have an SKU for it. If Intel had V-Cache, there would a version of 7950x3d with one and both dies having V-Cache.

From the 5800x3d fiasco, it is clear that AMD will not go the extra mile to have a top performing client CPU (AMD let Intel walk all over Zen 3 with Alder Lake, while diverting V-Cache to Milan-X, and gaining nothing by doing it).

As far as V-Cache, it is likely more tricky to incorporate it into monolithic design. But adding it to the separate standalone CCD die (which Strix Halo will likely have) is a problem that AMD already solved. So not incorporating it, and leaving perhaps 5% performance on the table is example of AMD not going the extra mile.

It's not like there is no upside. The client market is ~$10 billion and AMD only has about 15% of it. There is more upside for AMD on client PC than in datacenter GPU and datacenter CPU.

If you look at MTL, it looks like Intel may have over-engineered it, but it does not look like Intel left anything on the table from the tool Intel had in its tool chest.

TOP Dragon Range is clocked 300MHz lower than desktop counterparts, so It's not exactly the last 300-350MHz.
And ~5% higher performance at the same power doesn't look worth it If the price would be $100-200 higher.

Whatever the clock you start from, the last ~300 MHz is always the last. If you take out first 300 MHz, then the next 300 MHz is the last (and the least power efficient).

5% higher performance could help AMD make a bigger dent in the $10 billion client market TAM, from the Halo effect. If AMD even called it Strix Halo, someone obviously thought of this. Even 1% of that $10 billion TAM is $100 million.

As far as $100-$200 price increment, mixing prices and costs only confuses things. Cost of V-Cache is likely less than $10. How AMD prices it is a separate question. Intel is spending more than $10 extra to hit certain targets (lowest power for long battery life, decent graphics).

 

[TESKATLIPOKA]

If i did not have fairly new rig and the pricing was not so brutal, i would still take 7970x zen4 over 16C zen5 next year. Even 7960x i guess.

If It offers you more performance than 16C Zen5 then why not, but that price for 32C 7970x is not pretty at all.
I am also wondering If the difference between 16C Zen5 vs 24C 7960X will be big enough. Let's say that 16C Zen5 will really be 30-35% faster than 7950X, then this 7960X would be only 11-15% faster. In this case, I would probably forgo this extra performance and choose Zen5, which will be more powerful on anything below 16 threads and also cheaper.

 

[Joe NYC]

I understand your pain. X670E is way, way overpriced for what it offers, especially when there are $150 mobos from Gigabyte/ASROCK that can run a 7950X without much issue.

Problem with those is paying for something that you can't use right away (PCIe Gen 5 x16) and there is a chance you may never end up using.

But since AM5 is likely going to last 1 or 2 generations of CPUs, another 1 or 2 generations of GPUs, is seemed a little short sighted to skip PCIe Gen 5 for GPU. So it is a dilemma.

I tried to look for B650E, but somehow, each one I looked at was randomly missing one (or another) features I wanted, so I ended up with x670E. At least mine was under $300, unlike the last year intro pricing...

 

[Goop_reformed]

Problem with those is paying for something that you can't use right away (PCIe Gen 5 x16) and there is a chance you may never end up using.

But since AM5 is likely going to last 1 or 2 generations of CPUs, another 1 or 2 generations of GPUs, is seemed a little short sighted to skip PCIe Gen 5 for GPU. So it is a dilemma.

I tried to look for B650E, but somehow, each one I looked at was randomly missing one (or another) features I wanted, so I ended up with x670E. At least mine was under $300, unlike the last year intro pricing...

The only reason AM4 lasts for so long is because it was conceived as a marketing point for AMD to gain market share. No reason for AM5 to last until zen 7 or equivalence.

 

[FlameTail]

A significant processor with 12 CU RDNA GPU is coming soon, and nobody is talking about it?

 

[jpiniero]

A significant processor with 12 CU RDNA GPU is coming soon, and nobody is talking about it?

See the Phoenix thread?

 

[FlameTail]

Not Phoenix.
Take a guess what it is.

It's not like an ordinary AMD APU we talk about all day.

 

[Mopetar]

But since AM5 is likely going to last 1 or 2 generations of CPUs, another 1 or 2 generations of GPUs, is seemed a little short sighted to skip PCIe Gen 5 for GPU. So it is a dilemma.

It might be an issue for a top of the line GPU at the end of the lifespan of the platform, but any past tests have shown that GPUs can't utilize all of the bandwidth for 16 lanes on the latest standard.

Some of the more recent testing done regarding this topic found it's only an actual problem if the card doesn't have enough VRAM. Here's an image from Techspot illustrating this:

 

[Joe NYC]

The only reason AM4 lasts for so long is because it was conceived as a marketing point for AMD to gain market share. No reason for AM5 to last until zen 7 or equivalence.

AMD only needs to move to the new socket when DDR6 and or PCIe Gen 6 arrive. But it is possible AMD may move to the new socket before it happens.

Moving to new AM5 socket seemed to have gone less than smoothly for AMD, holding hack adoption of Zen 4. So fewer of these transitions the better.

We will see what types of markets AMD tries to tackle with what chip, but AM5 socket may just prove to be too small to add more CCDs, and potentially some sort of on package memory.

From pictures, it seems bigger than it really is, once you hold one in your hand...

 

[TESKATLIPOKA]

Not Phoenix.
Take a guess what it is.

It's not like an ordinary AMD APU we talk about all day.

You mean Hawk Point? That should be a rebranded Phoenix.

 

[adroc_thurston]

No reason for AM5 to last until zen 7 or equivalence.

AMD's really lazy in throwing new platforms out so ughhh.
Either way a platform change is inevitable when client moves to OpenSIL.

See the Phoenix thread?

It's the new Exynos ffs.

That should be a rebranded Phoenix.

not quite

 

[Joe NYC]

Not Phoenix.
Take a guess what it is.

It's not like an ordinary AMD APU we talk about all day.

Hawk Point?

Well, it seems to be only about 2 weeks away, and there are no leaks about what it adds over Phoenix, other than the MLID comments.

My theory is that the chip is just bigger with just more area devoted to AI.

Phoenix is apparently N4. N4P would be a nice surprise.

 

[Joe NYC]

AMD's really lazy in throwing new platforms out so ughhh.
Either way a platform change is inevitable when client moves to OpenSIL.

It's the new Exynos ffs.

not quite

With DDR6 not expected for a while (2026?), Zen 6 will likely arrive before DDR6.

In the meantime, LPDDR5x speeds are racing ahead. I wonder what the practical limit is. Or if the high speed DDR5 will make it to desktop.

It seems LPDDR5 will take more and more market share from DIMM versions DDR5, and that may influence the direction of the next socket.

 

[TESKATLIPOKA]

That's an interesting comparison. Cost may be in the same range (thanks to 7600 being cheap, on N6) but likely will perform worse and have higher power consumption, bigger size.

The SoC of Strix Halo is going to be quite big and it may be sub-optimal to have the most demanding (GPU) and the least demanding (IO) functions on the same N3E die.

I am wondering what the die size will be. Probably > 200 mm2.

I wouldn't be surprised, If It was that big.

Whatever the clock you start from, the last ~300 MHz is always the last. If you take out first 300 MHz, then the next 300 MHz is the last (and the least power efficient).

But you still save more power going down from 5.7GHz -> 5.4GHz than from 5.4GHz -> 5.1GHz.
And from what I saw in TPU's review about 7800X3D, this reduction in power draw is not that big.

5% higher performance could help AMD make a bigger dent in the $10 billion client market TAM, from the Halo effect. If AMD even called it Strix Halo, someone obviously thought of this. Even 1% of that $10 billion TAM is $100 million.

Strix Halo is not really a Halo product If you ask me. It's just too unbalanced to be one, and AMD would first need to make enough of them to make any dent in the market.
There is a reason why there are not that many Zen4 based laptops out there.

As far as $100-$200 price increment, mixing prices and costs only confuses things. Cost of V-Cache is likely less than $10. How AMD prices it is a separate question. Intel is spending more than $10 extra to hit certain targets (lowest power for long battery life, decent graphics).

I honestly don't care If this cache cost only $10 to make when I would have to pay $100-200 extra for such an APU in my laptop. If I had to pay $100-200 extra for an APU with 3d cache, I would like to see an equal benefit at least.

 

[FlameTail]

It's the EXYNOS 2400!

Yes, with 12 RDNA CUs to boot.

 

[adroc_thurston]

With DDR6 not expected for a while (2026?), Zen 6 will likely arrive before DDR6.

How is this platform-relevant?

Or if the high speed DDR5 will make it to desktop.

which one?
We're already running like 8k speed AMP kits.

 

[jpiniero]

It's the EXYNOS 2400!

Yes, with 12 RDNA CUs to boot.

Not Zen 5 either.

 

[TESKATLIPOKA]

It's the EXYNOS 2400!

Yes, with 12 RDNA CUs to boot.

And this is a thread about Zen5 !

 

[Joe NYC]

How is this platform-relevant?

which one?
We're already running like 8k speed AMP kits.

Well, DDR6 will almost certainly need a new socket. If we are still in DDR5 era when Zen 6 launches, I am not sure if there is any point of introducing a new DDR5 platform for Zen 6.

It could make sense to have a new platform for Zen 6, that ditches external memory and moves to on package memory... It may be a better solution than MrDimm on client.

As far as the higher than 5600 DDR5 speeds, I don't think any OEM uses it. Probably not worth the risk for them of reduced stability. But LPDDR5 is at 8533, moving on to 9600 with guaranteed stability...