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

[DisEnchantment]

Speculate at will

 

[Philste]

I wish the Bios was better, it's very basic. No fan profiles or anything, even in advanced mode.

You don't do these things in BIOS on Laptops. Its quite normal for them to have everything locked down. Try Fn+F or something.

 

[StinkyPinky]

You don't do these things in BIOS on Laptops. Its quite normal for them to have everything locked down. Try Fn+F or something.

Good call, that did it. Let's see if performance mode makes a difference.

 

[igor_kavinski]

You don't do these things in BIOS on Laptops.

Dell has multiple fan profiles in their Alder Lake Latitude laptop BIOS. I remember there was silent, power optimized, turbo and balanced.

 

[StinkyPinky]

15339 performance mode! With it plugged in. Fans were louder but not terrible by any means.

There is a "full fan" mode also...

 

[Philste]

Dell has multiple fan profiles in their Alder Lake Latitude laptop BIOS. I remember there was silent, power optimized, turbo and balanced.

Ok. Didn't know that. I know from Acer that Fn+f is power/Fan profiles and in BIOS you can't do anything. My 5 year old Swift 5 is barely thicker than Zenbook S16 and can cool ~21-23W with single Fan, so 28W shouldn't be a problem with 2 Fans.

 

[igor_kavinski]

so 28W shouldn't be a problem with 2 Fans.

Correction: ASUS IceCool Fan tech

 

[Philste]

15339 performance mode! With it plugged in. Fans were louder but not terrible by any means.

Can you open HwInfo or something during a run? For power draw.

 

[igor_kavinski]

What can I say? "It'll work great with our magic future compiler" is a pet peeve of mine.

For some reason.

The future is HERE NOW. Just get on over to Green Hills!

 

[StinkyPinky]

Can you open HwInfo or something during a run? For power draw.

HW Monitor shows 33.1W for the package and 28.2W for the CPU as max after a cinebench run.

Performance mode seems to be the sweet spot after doing some runs. Full fan mode gains like 5 points.

P cores max at 5.05 and E cores max at 3.32

 

[igor_kavinski]

Performance mode seems to be the sweet spot after doing some runs. Full fan mode gains like 5 points.

Upload a video review and recoup your investment with millions of views!

 

[SarahKerrigan]

View attachment 104018
The future is HERE NOW. Just get on over to Green Hills!

This is utterly irrelevant to your original point.

First, Coremark is literally the worst benchmark there is. Dhrystone is better.

Second, I never said some compilers aren't better than others. I said that aggressive optimization for one microarchitecture doesn't really accomplish much on modern OoO cores (none of which are tiny Cortex-M's anyway.) Of course some compilers are better than others. This is much of why there are compiler vendors to begin with.

At this point you're just tossing crap at the wall in the hope something sticks. I would encourage reading a book.

 

[Philste]

HW Monitor shows 33.1W for the package and 28.2W for the CPU as max after a cinebench run.

Thx. So it's 15.5K@33W vs 13K @28-32.5W for 7840U. Not particularly great with 2 cores/4 threads advantage.

P cores max at 5.05 and E cores max at 3.32

Cool, so David Huangs numbers were right. What is the clock during the run with 33W? How hard are the 5c cores already gimping? I bet the normal ones are at 3.8 or something.

 

[NTMBK]

Because there is no other developer here to help me out.

@NTMBK @eek2121 @quikah @repoman0 , HELP!!!!

Eh, the only specific microarchitecture we'd even think about targeting would be the PS5's Zen 2. Potentially sizing a few specific arrays to fit neatly into the L2 during a hot loop, would probably be the extent of intentional targeting we would do these days. But just profiling the code on that CPU and then optimizing based on that profile will generally lead you to accidentally prefer that uarch slightly- you'll hit the problem cases that crop up for that chip, and maybe miss other ones that might be more of an issue on other systems.

Outside of super hot algorithm loops (stuff like renderer code, physics engine etc), a lot of the time game code is more dictated by the flexibility you need to actually make a game. You could try to really maximise your register usage and minimise branch predictor misses on a chunk of code, but then game design ask you to make a big change to the feature and all your micro-optimization goes out the window and you're back at square one. Better to focus on getting big picture performance stuff right, and only go to the micro optimization right near the end of the project when things are locked down if you really need to squeeze the performance budget (and have time, and aren't too busy doing last minute bug fixes).

 

[Abwx]

Thx. So it's 15.5K@33W

He said 28W max CPU power, there s no average here, and the package in this case is eventually the whole device.

 

[Philste]

He said 28W max CPU power, there s no average here, and the package in this case is eventually the whole device.

My interpretation was 33W CPU Package 28W Core power.

 

[Abwx]

My interpretation was 33W CPU Package 28W Core power.

That s the case in DT but i noticed that for laptops it s sometime the package as a whole, that is, the whole laptop.

Other than this those numbers have all the characteristics of a chip that is somewhat overvolted to not only increase yields but foremost to guarantee an inconditional stability, that s more important for AMD than chasing even more perf/Watt advantage, they are already clearly in the lead so that s not as important as getting a good image.

 

[Geddagod]

You don't do these things in BIOS on Laptops. Its quite normal for them to have everything locked down. Try Fn+F or something.

The asus M16 is nice in that regard with how you can control PL1, PL2, and the fan curve without ever having to explicitly go into ur bios either.

 

[majord]

That s the case in DT but i noticed that for laptops it s sometime the package as a whole, that is, the whole laptop.

Other than this those numbers have all the characteristics of a chip that is somewhat overvolted to not only increase yields but foremost to guarantee an inconditional stability, that s more important for AMD than chasing even more perf/Watt advantage, they are already clearly in the lead so that s not as important as getting a good image.

its not the rop SKU either.

 

[Josh128]

I am not sure how to lock the frequency?

Hey! Maybe try to install the latest Ryzen Master. It may let you lock a low all core freq from within Windows (reboot will be needed). Even 3GHz would be fine.

Either that or turn off CPB (Core performance boost) in the BIOS-- that locks your CPU down to base frequency, at least on desktop Ryzen.

 

[Abwx]

its not the rop SKU either.

Right but it should still have better efficency numbers, i explained above to which extent.

If IPC is increased by 17% you can reduce frequency by 1.17x and have the same throughput at same core count.

At 1.25x the core count you can reduce frequency by 1.25X at same throughput, so with these two parameters combined the 365 can have 1/(1.17 x 1.25) = 0.68x the frequency at same throughput, wich imply less than 0.5x the power, yet here it s about 0.65x, so that could mean a voltage that is 14% higher than the 7840U at a same frequency.

This can also mean that voltage is not that much increased but that Zen 5 uarchitectural efficency is something like 10% lower than Zen 4 and combined with a slightly higher voltage.

 

[Mopetar]

Minor point, but at 17% IPC increase you'd only reduce the clock speed by 14.53%, (1 / 1.17).

Otherwise if you increased IPC by 100%, you'd have a very hard time getting similar performance if you reduced the clock speed by 100%.

 

[DaaQ]

There were new chipset drivers released today or yesterday.

 

[Abwx]

Minor point, but at 17% IPC increase you'd only reduce the clock speed by 14.53%, (1 / 1.17).

Otherwise if you increased IPC by 100%, you'd have a very hard time getting similar performance if you reduced the clock speed by 100%.

There s no error, you noticed that i wrote 1/(1.17 x 1.25) = 0.65x for the combined frequency reductions allowed by IPC and core count.

The rest of the math is that the typical Power/F curve of TSMC process is of the form P(f) = f^2.36.

So at 0.65x the frequency power is 0.65^2.36 = 0.36x at same throughput.

Even if the process is close to a theorical perfect mosfet curve this would still be 0.65^2 = 0.42x the power.

 

[majord]

Right but it should still have better efficency numbers, i explained above to which extent.

If IPC is increased by 17% you can reduce frequency by 1.17x and have the same throughput at same core count.

At 1.25x the core count you can reduce frequency by 1.25X at same throughput, so with these two parameters combined the 365 can have 1/(1.17 x 1.25) = 0.68x the frequency at same throughput, wich imply less than 0.5x the power, yet here it s about 0.65x, so that could mean a voltage that is 14% higher than the 7840U at a same frequency.

This can also mean that voltage is not that much increased but that Zen 5 uarchitectural efficency is something like 10% lower than Zen 4 and combined with a slightly higher voltage.

looking at the desktop Zen 5 leaks I don't think Zen 5 has much improved perf/watt in throughput.

but still, 7840U was the top silicon for the 15w -28w class so without testing halo SKU back to back t's hard to tell.. If the 370 and 365 were reduced core counts to differentiate themselves down the stack then you could probably rely on them being good quality silicon still, and assess them on a per core basis, but not so much when they're all 12 c parts (albeit with lower GPU CU on the 365) . they'll likely be both artificially limited in performance via boost clocks (st and NT) combined with higher yielding silicon with less favourable VF characteristics .


-Edit I made a mistake there , I forgot the AI 365 is 10c not 12 .

 

[Abwx]

looking at the desktop Zen 5 leaks I don't think Zen 5 has much improved perf/watt in throughput.

but still, 7840U was the top silicon for the 15w -28w class so without testing halo SKU back to back t's hard to tell.. If the 370 and 365 were reduced core counts to differentiate themselves down the stack then you could probably rely on them being good quality silicon still, and assess them on a per core basis, but not so much when they're all 12 c parts (albeit with lower GPU CU on the 365) . they'll likely be both artificially limited in performance via boost clocks (st and NT) combined with higher yielding silicon with less favourable VF characteristics .

Not easy to do an evaluation without knowing some parameters, FI if Zen 4 use the standard N5 process then N4P has 28% bettter efficency at same frequency, but if Zen 4 use a refined N5 then the difference is of course lower.

Beside we dont know what is the uarchitectural efficency of Zen 5 compared to Zen 4 but given that the core is substancially bigger it could be lower by an unknown amount.

Guess that reviews will clear all those numbers in a week or so.