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

[DisEnchantment]

Speculate at will

 

[TESKATLIPOKA]

Best Buy USA currently listed ASUS TUF A16 with 7600S 8GB @ $799 (drop from $1,099), seems like this is the model will be replaced by Sarlak base model.

Sarlak IGP with only 16CU would need to have ~3.3GHz game frequency to be in theory on par with 7600S.

I never believe in 40CU of RDNA3+; even with 256-bit LPDDR5x-8533, total memory bandwidth is around 272GB/s which is similar to 7600S with 32CU. Remember APU have to share memory bandwidth with CPU and GPU, thus I believe with real dual issue ALU, Sarlak will most likely having 20CU or WGP with 2560 ALU...

Maybe @adroc_thurston can provide some insights?

7600S has only 28CU. 7600M XT or 7700S have 32CU.
20CU IGP to match 7600M XT would need 3.7GHz game frequency.

Can It clock that high within a limited TDP? We will see.

 

[TESKATLIPOKA]

Hmm, the issue is with around $1,500, we have much better choice than large iGPU. I have updated the table to include ROG Strix G17 with Fire Range 12 Zen5 cores and RTX4070 GPU. Based on current pricing, the upcoming G17 with full fat Zen5 cores will offer much better graphics performance at slightly higher price. Would you rather go for RTX4070 or iGPU?

You have less cores (12 vs 16), less memory(32GB vs 48GB) and although RX 4070 is significantly more powerful It's limited to 8GB Vram.
So ROG Strix G17 with 4070 is not such a great deal honestly.

 

[adroc_thurston]

Sarlak IGP with only 16CU would need to have ~3.3GHz game frequency to be in theory on par with 7600S.

Wdym 16CU?
No one's gonna chop 60% of STX-halo iGP tile off.

You have less cores 12 instead of 16, less memory and although RX 4070 is more powerful It's limited to 8GB Vram.

What the hell are you on, none of STX-halo dies chops are anywhere near frozen

 

[TESKATLIPOKA]

Wdym 16CU?
No one's gonna chop 60% of STX-halo iGP tile off.

It's from @Tigerick not from me.
At least he has his answer about how many CUs are there -> 40.

What the hell are you on, none of STX-halo dies chops are anywhere near frozen

I am commenting what @Tigerick wrote about Ryzen 9 8950HX vs Ryzen 9 8855HX+RTX4070.

 

[Timmah!]

Correct!
Think of Zen5 as Conroe and Zen6 as Nehalem.

Zen6 still on AM5? That comparison would suggest it wont be.
Just want to know whether to consider replacing Zen4 with Zen5 or wait for 6.

 

[adroc_thurston]

Zen6 still on AM5?

For now, yes.

Just want to know whether to consider replacing Zen4 with Zen5

High chances you will.

 

[A///]

Nehalem brought a lot of changes for Intel that helped them out. think ad rock thurton is suggesting more cores for consumer ryzen with zen 6 as one example.

 

[adroc_thurston]

suggesting more cores for consumer ryzen with zen 6 as one example.

No, just a system/SoC design approach.

 

[igor_kavinski]

No, just a system/SoC design approach.

with soldered RAM, even in desktop?

 

[adroc_thurston]

with soldered RAM, even in desktop?

What?
No.
Why would you need LPDDR on desktop?

 

[igor_kavinski]

Why would you need LPDDR on desktop?

Faster speeds, for one. Plus, more memory bus width, especially if they want to bring more than 16 cores to desktop without increasing the number of memory slots on mobos and keeping most of them adhering to micro-ATX form factor. I personally don't want soldered RAM but I can see the point of it if it enables the best performance without needing to tinker with memory overclocking. They could even make the whole SoC with soldered RAM in the form of a large slot module that plugs vertically into the mobo. There are pros (easier assembly) and cons (could kill DIY) to it.

 

[adroc_thurston]

Faster speeds, for one

And horrible latency (LPDDR is GDDR levels of it).

Plus, more memory bus width

Don't need that.

especially if they want to bring more than 16 cores to desktop

Don't need that either.

 

[Joe NYC]

And horrible latency (LPDDR is GDDR levels of it).

Is LPDDR5 latency worse than equivalent speed DDR5 DIMM?

 

[adroc_thurston]

Is LPDDR5 latency worse than equivalent speed DDR5 DIMM?

Yea and by a lot.

 

[Timmah!]

For now, yes.

High chances you will.

Thanks.
How about 3D cache? Will we see clock downgrade cause by its addition resolved in Zen 5 or 6?

I went vanilla 7950x with Zen4. cause i did not want to wait for 3D chips later down the road, but ultimately did not like how 7950x3D was designed anyway - with inequal CCDs. If nextgen fixed that and introduced 16C chip with both v-cache dies and same clocks as vanilla ones - i would be tempted, even at somewhat higher price.

If Zen6 comes with some hybrid approach and non-equal dies already regardless of v-cache, then i guess this is a moot point.

 

[igor_kavinski]

And horrible latency (LPDDR is GDDR levels of it).

AMD can hide that easily with V-cache. Imagine a design where both CPU and GPU (over PCIe 5.0) access system memory through a large 256MB or even 512MB V-cache. Gobs of bandwidth, acceptable latency and enough RAM (if they go with minimum 48GB) to make even 8GB VRAM graphics cards perform well.

 

[Joe NYC]

Thanks.
How about 3D cache? Will we see clock downgrade cause by its addition resolved in Zen 5 or 6?

I went vanilla 7950x with Zen4. cause i did not want to wait for 3D chips later down the road, but ultimately did not like how 7950x3D was designed anyway - with inequal CCDs. If nextgen fixed that and introduced 16C chip with both v-cache dies and same clocks as vanilla ones - i would be tempted, even at somewhat higher price.

If Zen6 comes with some hybrid approach and non-equal dies already regardless of v-cache, then i guess this is a moot point.

There is just something unappealing about the asymmetric approach of 7950x3d. I think it was also a waste of time for AMD to introduce it, and delay further the 7800x3d.

Looking at the recent sales from Mindfactory, the 7900x3d and 7950x3d are non entities. So it was yet another own goal by AMD marketing.

https://twitter.com/x/status/1693298485917864362

Hopefully, by the time of 8950x3d, the V-Cache will not be reducing the clock speeds, and a symetrical CPU can be released, with V-Cache on both dies...

 

[adroc_thurston]

Will we see clock downgrade cause by its addition resolved in Zen 5 or 6?

No.

If nextgen fixed that and introduced 16C chip with both v-cache dies and same clocks as vanilla ones

No.

AMD can hide that easily with V-cache

That costs money and not really suitable for mainstream segments like laptop.

and GPU (over PCIe 5.0) access system memory through a large 256MB or even 512MB V-cache.

Oh hell no why would you want to DMA into some cache blob over PCIe.

 

[igor_kavinski]

Oh hell no why would you want to DMA into some cache blob over PCIe.

Why not?

 

[adroc_thurston]

Why not?

you gotta check PCIe bandwidth and latency to understand

 

[A///]

There is just something unappealing about the asymmetric approach of 7950x3d. I think it was also a waste of time for AMD to introduce it, and delay further the 7800x3d.

Looking at the recent sales from Mindfactory, the 7900x3d and 7950x3d are non entities. So it was yet another own goal by AMD marketing.

Only useful for the people who want it all, a fast processor with lots of cores and that cache, except the regular compute die is still a lower clocker than the reg processor. I think amd will eventually figure out a sweet spot and way for everything.

 

[dr1337]

Looking at the recent sales from Mindfactory, the 7900x3d and 7950x3d are non entities. So it was yet another own goal by AMD marketing.

They sold more 7950x3ds than 13900k per your source, 7900x3d matches the Intel chip in sales but both are still ahead of the vanilla 7950x. For halo chips they seem to have a very solid place in the lineup, that's a lot more than just existing for a marketing goal. And your logic makes the same argument for Intel which I would also completely disagree with, since flagships aren't supposed to be huge sellers in a random month well after launches.

There is just something unappealing about the asymmetric approach of 7950x3d.

Well the customers prefer it much more over the non-cache version. AMD definitely released the halos first to mop up impulse sales but the market seems to always agree that budget is more valuable than sheer cores.

 

[itsmydamnation]

People need to watch PC jesus trip to amd campus as to why you wont see dual vcache 16 core parts. The actual answer is to put the vcache on the memory controller/ MALL.

 

[Joe NYC]

People need to watch PC jesus trip to amd campus as to why you wont see dual vcache 16 core parts. The actual answer is to put the vcache on the memory controller/ MALL.

That would allow the MALL cache to be shared between multiple CCDs.

Between 2 CCDs each having its own 64MB of L3 as one alternative and MALL having 128MB shared cache as another alternative, MALL would come out ahead in most scenarios as far as achieving cache hits, but there would be small latency hit vs. a cache hit in CCDs own or stacked L3.

If CCD is connected to IOD+MALL using Hybrid Bond bridges, as could happen with Venice, the extra latency would be reduced.

Looking at AMD deploying MALL to client GPU, datacenter GPU / APU in Mi300, I think MALL will be the answer to CPU, both client and server.

As far as when we could see this, definitely not in Zen 5 client, highly unlikely in Zen 5 Turin server (even though there is a new IOD coming)

But I think highly likely with Zen 6, like 90+ % likely.

 

[adroc_thurston]

The actual answer is to put the vcache on the memory controller/ MALL

MALL doesn't need SoIC-X (MI300 MALL is run-thru a bunch of 2.5D interfaces really).

but there would be small latency hit vs. a cache hit in CCDs own or stacked L3.

A large one, it's a memory side cache.

the extra latency would by

Good lord someone missed his comparch course in college.

I think MALL will be the answer to CPU, both client and server.

Client yes, battery life is nice.
Server ehhhh.