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

[DisEnchantment]

Speculate at will

 

[majord]

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

Is it possible to monitor it during r23 Run. to see what it settles at?

 

[gaav87]

I am not sure how to lock the frequency?

Universal x86 Tuning Utility

amdaputuningutility.com

Try this: Universal X86 Utility
And here is a video how to use it:

 

[Josh128]

Universal x86 Tuning Utility

amdaputuningutility.com

Try this: Universal X86 Utility
And here is a video how to use it:

Perhaps, but I would give Ryzen Master a shot too, being its an official AMD app. That or turn Core Performance Boost off in BIOS. But if those dont work, this might do the trick!

 

[Abwx]

Zen 5 ipc in cb is only on par with raptor lake.

You are confused by RPL s higher frequency, 13900K does 2262 pts at 5.8Ghz, 9700X does 2280 pts at 5.5 - 5.55 GHz.

 

[Josh128]

@Malachi - the R23 MT numbers posted by the ES user here do actually hit about 16.5% in MT in R23, as did the locked 5GHz run done during tech day. None of the ST scores or MT scores of auto-boosting SKUs do though- most likely because they arent reaching the same frequencies as their Zen 4 counterparts. Which is pretty shitty because that Donny W dude specifically said Zen 5 has better "clock residency" than Zen 4. From literally every single leak we've seen, perhaps other than this CPU-z leak, that has obviously not been the case.

https://twitter.com/x/status/1817249994233278572

 

[TESKATLIPOKA]

Perhaps, but I would give Ryzen Master a shot too, being its an official AMD app. That or turn Core Performance Boost off in BIOS. But if those dont work, this might do the trick!

Ryzen master is not working for laptop APUs, I already tried It with my Phoenix laptop.
The only option is that Universal x86 Tuning Utility.

 

[Philste]

Strix Embargo fell right now... lets read the articles

 

[FlameTail]

I am surprised the thread hasn't exploded already!

 

[Abwx]

Accprding to Computerbase the 370@33W match the 8945HS@75W in CB 2024, that s 0.44x the power at same perf, the estimations i did are right on the money.

AMD Ryzen AI 9 HX 370 „Strix Point“ im Zenbook S16 im Test

AMD „Strix Point“, die neue APU mit Zen-5(c)-Kernen, RDNA 3.5 und starker NPU, ist da. Im Test ist der Ryzen AI 9 HX 370 sehr effizient.

www.computerbase.de

 

[Philste]

Meh. Computerbase: 370@33W is 18% faster than 8945HS@37W. With 50% more Cores/Threads. iGPU is bugged af and Clocks below 2.5GHz, only 10-15% uplift in steel nomad light.

 

[Abwx]

Meh. Computerbase: 370@33W is 18% faster than 8945HS@37W. With 50% more Cores/Threads. iGPU is bugges af and Clocks below 2.5GHz, only 10-15% uplift in steel nomad light.

As fast at 33W than the 8945HS at 75W, that s the number to consider.

Beside 18% faster at 33W than the 8945HS@37W mean 24% faster than the 8945HS@33W.

 

[Philste]

As fast at 33W than the 8945HS at 75W, that s the number to consider.

No? That's just massive cherrypicking. If you run Phoenix that much out of sweetspot, it will probably lose to Rembrandt@35W, while being a full node ahead.

Beside 18% faster at 33W than the 8945HS@37W mean 24% faster than the 8945HS@33W.

Which is quite bad for 50% more cores/threads.

 

[Abwx]

No? That's just massive cherrypicking. If you run Phoenix that much out of sweetspot, it will probably lose to Rembrandt@35W, while being a full node ahead.

That s the other way around, chips should be compared at isoperf, why should a device be required to do more work in a given time for a comparison..?..

Which is quite bad for 50% more cores/threads.

That way good enough to have the best chip by far, what matters is that at 33W it perform 28% better than the 185H@46W, i let you imagine the tiny power amount required to just match the Intel chip, because the competition is Intel, not Hawk Point.

 

[Philste]

That s the other way around, chips should be compared at isoperf,

You can do that, but it still doesn't gell you anything on the archs efficiency.

because the competion is Intel, not Hawk Point.

Of course Hawk Point is competition. These Strix devices nearly cost double compared to Hawk Point ones. All for 20% CPU and 15% GPU Performance and that stupid NPU.

 

[igor_kavinski]

These Strix devices nearly cost double compared to Hawk Point ones. All for 20% CPU and 15% GPU Performance and that stupid NPU.

That's just the usual AMD early adopter tax. These will get discounted in a mere few months IF demand doesn't outstrip supply.

 

[FlameTail]

Anandtech review:

The AMD Ryzen AI 9 HX 370 Review: Unleashing Zen 5 and RDNA 3.5 Into Notebooks

www.anandtech.com

 

[Philste]

That's just the usual AMD early adopter tax. These will get discounted in a mere few months.

Months ago we said Strix will be a new Performance Tier and Kraken will be the true Phoenix successor? What do we say now. Strix is barely a normal generational uplift to Phoenix and Kraken is set to be destroyed.

Btw. look at the Anandtech Review. ZEN5 has 0 IPC in SepcINT and 15% IPC in SpecFP.

ZEN5 definitely looks like AMDs 2nd slip after RDNA3.

 

[Panino Manino]

Mixed feelings about Anandtech review.
ST performance is missing, IMO.
At least RDNA 3.5 really seems to be a real improvement.

 

[Ghostsonplanets]

Phoronix:

When taking the geometric mean of the 100+ benchmarks run for this launch day article, the Ryzen AI 0 HX 370 was overall about 10% faster than the Ryzen 7 7840HS and about 18% faster than the Ryzen 7 7840U... Not bad, considering those prior Zen 4 parts were all Zen 4 full fat cores compared to the Ryzen AI 9 HX 370 having four Zen 5 cores and eight Zen 5C cores. But across a variety of real-world workloads especially among creator workloads and more, this Strix Point laptop was delivering great uplift. The Ryzen AI 9 HX 370 also came in 43% faster than the current Intel competition with the Core Ultra 7 155H "Meteor Lake" processor.

Where things got really interesting with the Ryzen AI 9 HX 370 was the power efficiency of this Zen 5 laptop SoC. Across the span of all the benchmarks, the Ryzen AI 9 HX 370 SoC was pulling about 20.4 Watts with a peak of 34.2 Watts.... Meanwhile the Ryzen 7 7840HS had an average of 35 Watts and a peak of 60 Watts. The Ryzen 7 7840U had a 27 Watt average and a peak of 51 Watts. The Ryzen AI 9 HX 370 came out faster than those parts while consuming significantly less power. This Zen 5 power efficiency is very exciting and should carry over for desktop and server parts too. Meanwhile the Core Ultra 7 155H was consuming 29.6 Watts on average with a peak of 65 Watts.

AMD Ryzen AI 9 HX 370: 100+ Benchmarks Validate Zen 5's Captivating Power Efficiency & Performance Review - Phoronix

www.phoronix.com

Seriously impressive efficiency gains gen over gen.

 

[Hitman928]

Mixed feelings about Anandtech review.
ST performance is missing, IMO.
At least RDNA 3.5 really seems to be a real improvement.

Anandtech has never been good with laptop reviews, we need more data to really get a good idea of STX performance.

 

[Rheingold]

What I'm taking away from the first Strix Point reviews is that AMD was talking about its great AI capabilities and the NPU with 50 TOPS non-stop, and yet I find exactly zero benchmark results that use said NPU. Did somebody find anything where the NPU was actually tested?

 

[Philste]

yet I find exactly zero benchmark results that use said NPU. Did somebody find anything where the NPU was actually tested?

Because there is nothing to test it in. There's no reason for these NPUs right now.

 

[RnR_au]

Seriously impressive efficiency gains gen over gen.

Yeah but not really ready for usage under linux until later this year unfortunately...

From the Phoronix review;

Another item to be aware of is the heterogeneous core topology handling. AMD engineers have been working on AMD P-State patches for heterogeneous core topology CPUs but as of writing the patches haven't yet been upstreamed into the mainline Linux kernel. These patches should help with better placement of processes onto the Zen 5 cores rather than Zen 5C. But as these patches weren't picked up for the Linux 6.11 merge window, unless they end up being submitted as a "fix" for Linux 6.11, that means they won't be merged until at least the Linux 6.12 kernel cycle. Linux 6.12's merge window will be in September while that stable kernel won't be out until near the end of 2024.

Which is kinda fine for me as it allows me to have an excuse to wait for Strix Halo

 

[jpiniero]

Because there is nothing to test it in. There's no reason for these NPUs right now.

stonks

 

[Rheingold]

There's no reason for these NPUs right now.

Yeah, but then that's a complete fumble on AMD's side. Pushing AI in every single presentation and then having nothing to show on release. As usual, they did everything for Microsoft only for Redmond to completely ignore them regarding CoPilot+ support. But it would have been nice to at least work together with the developers of the Procyon suite to get XDNA2 support on launch day.