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

[DisEnchantment]

Speculate at will

 

[branch_suggestion]

GCC 15 Compiler Showing Off Nice Performance Improvements On AMD Zen 5 - Phoronix

www.phoronix.com

A nice little bump from optimised compilers.

 

[DrMrLordX]

A nice little bump from optimised compilers.

Hmm, interesting. Did this compiler update help only Zen5?

 

[Tuna-Fish]

AMD should double the v-cache with Zen 6 and make the standard processors with the same amount of v-cache found on the 9800x3d chips. The dual v-cache sounds great but AMD has finally fixed their clock regression on Zen 5 v-cache. It makes sense to make v-cache standard and introduce a new x3D CPU with double the v-cache.

No, it does not, because vCache adds manufacturing cost, thus giving them an additional way to scale a single chip to different market segments. Low-end chips ship without vCache, making them cheaper.

It would probably make sense to manufacture additional sizes of vCache. The current size for mid-tier, double size (or more) for next flagship? There are probably diminishing gains, but I'd really like to find out where the limits are.

 

[igor_kavinski]

I have to chuckle because every time I mentioned in my posts that software may be rewritten or compilers may be optimized for Zen 5 to extract more performance out of it, some people scoffed at the idea and "explained" that nobody would go to the trouble of doing that. Yet, that's what has happened. Maybe because if you tell developers that there is "hidden" performance in some piece of hardware, they get excited enough to devote their time on unlocking it.

 

[eek2121]

No, it does not, because vCache adds manufacturing cost, thus giving them an additional way to scale a single chip to different market segments. Low-end chips ship without vCache, making them cheaper.

It would probably make sense to manufacture additional sizes of vCache. The current size for mid-tier, double size (or more) for next flagship? There are probably diminishing gains, but I'd really like to find out where the limits are.

I do somewhat agree, however, an easy solution is to add more cache to the XT parts and leave the non-XT parts without it.

 

[eek2121]

View attachment 116577

I have to chuckle because every time I mentioned in my posts that software may be rewritten or compilers may be optimized for Zen 5 to extract more performance out of it, some people scoffed at the idea and "explained" that nobody would go to the trouble of doing that. Yet, that's what has happened. Maybe because if you tell developers that there is "hidden" performance in some piece of hardware, they get excited enough to devote their time on unlocking it.

As someone who has done some really fun software development, I can say you are correct.

I wrote a gaming engine optimized for core count back when the 1950X was released. Why? It was fun and a cool tech demo, even though I never released it.

 

[eek2121]

I always use headphones. There's just one big problem... I use open-back headphones so fan noise is an issue. My T14 Gen 1 is noisier in average load than my desktop in full load.

I use open-back headphones as well, but my desktop has adequate cooling is is near silent under load.

You can actually build a silent desktop, even with hot/power hungry Intel parts.

As a bonus, if you don’t need a huge GPU, you can get the weight down under 2.5kg.

You will pay for all of that, of course.

A youtuber did a 9950X build with no fans at all, not even in the PSU, even running at stock settings (170/230W).

 

[igor_kavinski]

It was fun and a cool tech demo, even though I never released it.

Hope you can release the tech demo part of it at least

 

[eek2121]

Weight ? With power brick of course...
Noise ?
Not my type of machine I'm afraid....

$2.6k for the equivalent of full desktop 9950X3D + 5060 12gb PC with all peripherals + monitors etc

100% worth it
(though that specific Clevo chassis is probably not very worth it ergonomics-wise, there's gonna be better ones)




Best offerings gonna be around 2.5kg - just right
Others gonna be 3.5kg - on the heavy side
Power bricks around 700 to 999 grams depending on company, or get Slim-q 330w at ~400grams

2025 gaming laptop noise even at full load is always bearable. 100% fan speed in games are also still fine given there's game music/sounds (usually only under Custom power profiles, otherwise it's like 80% max which is totally fine)

2,6Kg power supply excuded, around 3-3,2Kg with it. Noise depends on profile, definitely you'll need headphones when gaming at full speed (this has even a detachable liquid cooling system). But there are people who need such things, like me.

Power brick weight/size is out of control, especially since USB-C PD can power most or all of these devices.

My alienware laptop is thin and light, despite having a laptop 3070 (which performs similarly to a desktop 3060). Sounds great until you see the power brick. Weighs almost as much as the laptop. If USB-PD on that machine supported 100-230W it wouldn’t need one, instead it is limited to something like 60W and requires the power brick when completely dead or the GPU is used.

 

[igor_kavinski]

You can actually build a silent desktop, even with hot/power hungry Intel parts.

Having a 12700K, umm...not too sure about that

 

[eek2121]

Hope you can release the tech demo part of it at least

If I can find the source. I hope to release a bunch of my older projects.

It was basically a primitive voxel engine. Load it up, it generates a world you can run/jump around in. It was still missing quite a bit, but technically you could build a single player game using it. It scaled pretty well with core counts and didn’t need a fast GPU since most things were done in software. Think Minecraft, but slightly better graphics and no water lol.

 

[eek2121]

Having a 12700K, umm...not too sure about that

IIRC VideoCardz had someone make one with a 13900k or 14900k. With big enough heatsinks and the right design you can cool quite a bit without a fan. Maybe that company should work on a nuclear reactor or something. 🤣

 

[igor_kavinski]

It scaled pretty well with core counts and didn’t need a fast GPU since most things were done in software.

Sounds great as a multicore CPU benchmark if you can script it and then it gives a score at the end.

 

[igor_kavinski]

Puget Systems Most Reliable Hardware of 2024

With 2024 wrapped up, we want to take the opportunity to call out specific brands and models that had exceptional reliability over the last year.

www.pugetsystems.com

SHOCK. SURPRISE.

Is AMD sprinkling gold dust on these CPUs?

Personally, I think that the fact that each TR CPU having a fuse blown when it is overclocked, is keeping their users from pushing them beyond comfortable levels.

 

[MS_AT]

I have to chuckle because every time I mentioned in my posts that software may be rewritten or compilers may be optimized for Zen 5 to extract more performance out of it, some people scoffed at the idea and "explained" that nobody would go to the trouble of doing that. Yet, that's what has happened. Maybe because if you tell developers that there is "hidden" performance in some piece of hardware, they get excited enough to devote their time on YourY

Not really. All that this test is showing is that gcc 15 is faster than 14. It might be because of new algorithms or bug fixes in the compiler itself that have nothing to do with Zen5. You would need to test with znver5 on/off while keeping instruction sets the same to claim there are any zen5 specific optimizations. It might be Zen4 and ArrowLake will see similiar gains.

 

[fastandfurious6]

Power brick weight/size is out of control, especially since USB-C PD can power most or all of these devices.

240W DC/USB C Charger

SlimQ 240W Gaming Laptop Adapter AC Port Power Adapter Type C Ports GaN Super Fast Charger Laptop Charging Travel Power Portable Compatible Dell HP MSI Razer Lenovo Asus Sony Acer Toshiba Aftershock and More Nice Style: The patented design of USB C 240W GaN charger offers a unique style and...

slimq.life

400g, pocket size, 3 ports also replace phone chargers

pricy but buy once keep very long time

330w / 550g model available for big dgpu 5080+ laptops

 

[igor_kavinski]

It might be Zen4 and ArrowLake will see similiar gains.

Fair point.

 

[lightmanek]

Puget Systems Most Reliable Hardware of 2024

With 2024 wrapped up, we want to take the opportunity to call out specific brands and models that had exceptional reliability over the last year.

www.pugetsystems.com

SHOCK. SURPRISE.

Is AMD sprinkling gold dust on these CPUs?

Personally, I think that the fact that each TR CPU having a fuse blown when it is overclocked, is keeping their users from pushing them beyond comfortable levels.

Naah .... It took me less than 30 minutes to take Threadripper 7985WX from baby performace at 350W stock to a cosy 485W with PBO and CO. On a 10h render in RelityCapture, it saved almost 2h for my client. Platform is rock solid and comes with 256GB of ECC DDR5 6000.
Who cares about fuses if there is time to be saved (time=money)!
PS. It is watercooled, Nocuna was there only for initial POST to make sure all is good.

 

[Nothingness]

View attachment 116577

I have to chuckle because every time I mentioned in my posts that software may be rewritten or compilers may be optimized for Zen 5 to extract more performance out of it, some people scoffed at the idea and "explained" that nobody would go to the trouble of doing that. Yet, that's what has happened. Maybe because if you tell developers that there is "hidden" performance in some piece of hardware, they get excited enough to devote their time on unlocking it.

Vectorization has improved. Too bad too little software can be vectorized which explains the aggregated gain is only 2% (which again is hidden by how you crop images).

I wonder if AVX-512 was produced for any of the tests that saw larger speedups.

I also wonder if any part of memcached can be vectorized (perhaps the hash function).

 

[igor_kavinski]

View attachment 116580View attachment 116581

Thanks for the pics! That's hardware I won't ever see in real life

 

[lightmanek]

Thanks for the pics! That's hardware I won't ever see in real life

Never say never!
I thought the same back in early 2000s and then in 2018 I've build myself a Threadripper 1920X as I needed more PCIE (crypto craze helped fund this little project).
Nowdays I would love to have 64C or 96C TR as a daily, but I really have almost no use for such compute power. Besides, the above machine is only 3 min drive from my home, so I can always go and admire it in person 😅

 

[gdansk]

Thanks for the pics! That's hardware I won't ever see in real life

It might be affordable in 5 years.

 

[igor_kavinski]

It might be affordable in 5 years.

My mind will have moved on by then

 

[Hans Gruber]

No, it does not, because vCache adds manufacturing cost, thus giving them an additional way to scale a single chip to different market segments. Low-end chips ship without vCache, making them cheaper.

It would probably make sense to manufacture additional sizes of vCache. The current size for mid-tier, double size (or more) for next flagship? There are probably diminishing gains, but I'd really like to find out where the limits are.

We all know v-cache works and without it, Zen 5 CPU's do not perform well. The easy solution is to make v-cache standard and add more v-cache to the x3D chips.

 

[Hitman928]

We all know v-cache works and without it, Zen 5 CPU's do not perform well.

Come again?