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

[DisEnchantment]

Speculate at will

 

[poke01]

It’s here!

Edit: Pictures

It’s 9% Specint increase over Zen 4.


Comparison with M3/M4.

 

[mostwanted002]

It’s here!

Edit: Pictures
View attachment 104146
It’s 9% Specint increase over Zen 4.

View attachment 104148
Comparison with M3/M4.

So it’s pretty much the “intel” moment for AMD right now; might be able to overcome it with future updates.

- cross CCX latency b/w Z5 & Z5c is worse than MTL

- They set the game’s affinity to full Zen5 cores only which saw the performance boost. Hinting towards better scheduling can give performance boost all around

- SteelNomad wasn’t able to demonstrate any difference at all, although there was significant uplift in TimeSpy

- Games which are memory bandwidth dependent don’t perform well (ex: PUBG PC)

 

[poke01]

SteelNomad wasn’t able to demonstrate any difference at all, although there was significant uplift in TimeSpy

Steel Nomad is more demanding and 890M needs faster memory and more cache for that benchmark I suppose.

 

[Geddagod]

 

[DrMrLordX]

Perf/Watt on HX370 is really good. Really really good.

 

[H433x0n]

View attachment 104149
View attachment 104150
View attachment 104151

Interesting that below 9W Zen 4 has better integer performance. I suppose that's a good demonstration to the cost of going wider.

 

[ErichZannMusician]

Interesting that below 9W Zen 4 has better integer performance. I suppose that's a good demonstration to the cost of going wider.

And at ~6W Zen5 has identical int perf to Redwood Cove.

One thing is for sure, there is no way +16% IPC uplift is materializing on the desktop in the most general applications. It'll be lucky if AMD gets +10%.

 

[FlameTail]

AMD should pay someone for better memory controllers. Strix is way behind MediaTek, Intel, Qualcomm SOCs. 1000MT/s less can't be helpful for that GPU.

I heard a rumour that AMD's working in validating 8533 MT/s for Strix Point's memory controller.

 

[FlameTail]

So this means H370x is more than double the performance of the 155H at 15W, probably almost double the 185H too. Good luck to to Lunar Lake to catch up with advertised 50% improvement.

Really funny people here are acting like these aren't the best mobile CPUs outside of Apple.

Well, Intel isn't the only mobile CPU maker outside of Apple. Now we also have Qualcomm.

I'd love to see power curves of Ryzen AI 9 HX 370 vs Snapdragon X Elite.

 

[adroc_thurston]

AMD should pay someone for better memory controllers. Strix is way behind MediaTek, Intel, Qualcomm SOCs. 1000MT/s less can't be helpful for that GPU.

It's a platform thing.
-halo runs faster since it's relevant there.

 

[mikegg]

So Apple Silicon roughly 50-59% faster while using 50% - 60% of the power.

 

[igor_kavinski]

So Apple Silicon roughly 50-59% faster while using 50% - 60% of the power.

Try running your precious silicon at 5.5 GHz.

 

[yuri69]

Try running your precious silicon at 5.5 GHz.

It doesn't need to. They reach their perf goal with less frequency.

 

[RnR_au]

So Apple Silicon roughly 50-59% faster while using 50% - 60% of the power.

On a node better than AMD, with ram in the soc, with an arguably simpler instruction set, and can add instructions as needed since it owns the entire platform.

Oranges and apples mate 🙂

 

[okoroezenwa]

On a node better than AMD, with ram in the soc, with an arguably simpler instruction set, and can add instructions as needed since it owns the entire platform.

Yet both on the market now.

 

[tsamolotoff]

https://twitter.com/x/status/1818151400658141412

Interesting data, I wonder how this comparison would change with proper full-width cores.

 

[MarkizSchnitzel]

It's still an extremely impressive look for apple here.
I think it's obvious that there is a large market for someone to do Mx for Windows, but for some reason it can't be done. Qualcomm seems most likely to get there though, if they stick to it.

 

[yuri69]

On a node better than AMD, with ram in the soc, with an arguably simpler instruction set, and can add instructions as needed since it owns the entire platform.

Oranges and apples mate 🙂

AMD's node advantage wasn't mentioned when Intel got an inferior node. Also AMD deliberately chose not to pick 3nm for the APU.

RAM on package is also a design choice.

So these SoCs can be compared.

 

[poke01]

I think it's obvious that there is a large market for someone to do Mx for Windows, but for some reason it can't be done. Qualcomm seems most likely to get there though, if they stick to it.

Thats Lunar Lake goal. Its the x86 version of M3.

arguably simpler instruction set

Thought the ISA didn't matter 😛

 

[Nothingness]

and can add instructions as needed since it owns the entire platform.

That doesn't affect SPEC results. AMD has the lead here with AVX-512 which will affect some of the tests (mostly the x264 one).

 

[Nothingness]

Try running your precious silicon at 5.5 GHz.

Isn't Strix Point supposed to run at a lower frequency of 5.1GHz?

 

[MS_AT]

It doesn't need to. They reach their perf goal with less frequency.

The point is that 4.0GHz was not the target frequency of this architecture. That of course doesn't change the fact overall Apple will take the lead, but lowering the frequency might undersell the core a bit. For example if compared Zen4 vs Zen 5 at 1.5GHz then I think Zen 4 would be ahead in Int at least based on the graphs shown.

 

[MS_AT]

That doesn't affect SPEC results. AMD has the lead here with AVX-512 which will affect some of the tests (mostly the x264 one).

That depends on who is running SPEC, since some people compile the test with default x86 baseline that depending on compiler version and might not even allow for AVX2, others are using march native that would allow AVX512 usage if compiler knew about the core but would fall back to default if it was not recognized, and some are ensuring specific sets are enabled (like AT with avx2+fma).

I think Geekerwan, Anandtech and David Huang are all using different compilers and different flags to boot.

Now it won't matter if SPEC has a specific code paths for each instruction set and is doing a runtime dispatch based on detected processor features but if we are talking about the autovectorization then I think only David is using flags that would make use of AVX512 possible.

 

[igor_kavinski]

Isn't Strix Point supposed to run at a lower frequency of 5.1GHz?

Yes but if Apple silicon is so great, why can't it scale all the way to 5.5 GHz and solve all the world's problems? Strix Point can be pushed further. AMD chose not to, to keep it within the efficient part of the v/f curve. We also don't know what Apple's design choices cost them. Using huge caches must not be good for their yields. They are not doing anything like V-cache so cache defects lead to discarding the entire die. They have some ways to go before they can be compared to "real world" CPUs that the majority of the world uses. Not select few with oversized pockets or tendency to get themselves into cc debt.

Here's what Apple needs to do before I would consider it fair for their CPUs to be used in these comparisons:

Make them price competitive with the competition.
Provide Linux drivers for their hardware so at least the community can work on creating decent Linux distros for their hardware.
Stop preventing users from upgrading their Macbooks by artificially soldering RAM and SSDs.
Give licenses to select few OEMs to use their chipset/SoC in their own laptop designs with beefier cooling and even loud jet fans, so users can look forward to more plugged-in performance.

Until the above is a reality, these comparisons are nonsensical and just fantasies. No one is going to switch to Apple based on these graphs.

 

[poke01]

Yes but if Apple silicon is so great, why can't it scale all the way to 5.5 GHz and solve all the world's problems? Strix Point can be pushed further. AMD chose not to, to keep it within the efficient part of the v/f curve.

I like how you answered your own question. Let me help: M4 can be pushed further. Apple chose not to, to keep it within the efficient part of the v/f curve.

Zen5 can't solve all the worlds problems too, I wish it did.