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

[DisEnchantment]

Speculate at will

 

[ToTTenTranz]

Sort of. 3.5 actually works.

It supposedly has better voltage-clock curves, which is why they're putting an additional 2 WGPs in there without increasing power consumption, but IPC-per-WGP should be similar between the two.

Of course, Strix Point's GPU should be even more bandwidth bottlenecked than Phoenix's because there's 33% more WGPs and 13.8% higher RAM bandwidth (plus a CPU with more cores = more demanding for the memory controller).

 

[adroc_thurston]

It supposedly has better voltage-clock curves, which is why they're putting an additional 2 WGPs in there without increasing power consumption, but IPC-per-WGP should be similar between the two.

Yeah, but clocks is where RDNA3 struggled.

Of course, Strix Point's GPU should be even more bandwidth bottlenecked than Phoenix's because there's 33% more WGPs

Phoenix wasn't bandwidth-bound much, and neither was RMB.
STX is, though, MALL would've been real handy for actual video games

 

[fastandfurious6]

CPU: AMD Ryzen AI 9 HX 170
Min/Max/Avg: 3545/3673/3653 MHz
Codename: Strix Point
Single: 2544
Multi: 14158
https://browser.geekbench.com/v6/cpu/6418984

This is extremely disappointing



my GB 6.3 results on 7840HS (same as 8845HS)

single 2396
multi 12477

LENOVO 82Y4 - Geekbench

Benchmark results for a LENOVO 82Y4 with an AMD Ryzen 7 7840HS processor.

browser.geekbench.com

and that's with 3 browsers open, 50 tabs active and another 600 inactive, several apps open and music playing in background

YIKES!!!!!!!!!!!!!!!!

 

[Mahboi]

Literally every single leaker has been proven wrong and every single leak has been shown to be untrustworthy.

MLID is now the most trustworthy source on the internet.

So more than ever WAIT FOR REVIEWS before throwing

This is extremely disappointing
my GB 6.3 results on 7840HS (same as 8845HS)
YIKES!!!!!!!!!!!!!!!!

 

[gdansk]

The 7840U I tested scored 2642/11352 in GB6.
But the model name, clock rates and/or clock rate reporting don't seem to be correct for that Strix example.

 

[fastandfurious6]

are you saying the strix point GB bench is fake?

also the theory that "linear 5.1 will be 3500+" is plain wrong

just did 7840hs gb6.3 @ base freq with boost disabled

1975 single
10848 multi
~20% difference

which means at very best case strix point at max boost (?) has around 2800 single 15000 multi

 

[gdansk]

Not fake but too questionable to extrapolate from.

 

[fastandfurious6]

if fire range (mobile 9950x) manages ~3500 single GB6 at acceptable temps then that's a success

but probably gonna be around 3000

let's see

 

[adroc_thurston]

But the model name, clock rates and/or clock rate reporting don't seem to be correct for that Strix example.

They rebranded it last minute and it's either running the silent mode or it's clock-locked using the PBO available on pre-prod BIOSes.

 

[Thunder 57]

They rebranded it last minute and it's either running the silent mode or it's clock-locked using the PBO available on pre-prod BIOSes.

Tom's says silent mode.

 

[Abwx]

This is extremely disappointing



my GB 6.3 results on 7840HS (same as 8845HS)

single 2396
multi 12477

LENOVO 82Y4 - Geekbench

Benchmark results for a LENOVO 82Y4 with an AMD Ryzen 7 7840HS processor.

browser.geekbench.com

and that's with 3 browsers open, 50 tabs active and another 600 inactive, several apps open and music playing in background

YIKES!!!!!!!!!!!!!!!!

With browsers opened those numbers are meaningless, the higher score at NBC is 2721 for the 7840HS, wich is 13.5% higher than your score, so your CPU didnt clock at 5.1GHz in ST, rather at 4.6-4.7GHz.

https://www.notebookcheck.net/AMD-Ryzen-7-7840HS-Processor-Benchmarks-and-Specs.680876.0.html

 

[therealmongo]

AMD, it turns out, likes to make money just as much as Intel and Nvidia.

Dont post much, but wow, shock horror, what a revelation this is

 

[therealmongo]

Zen5 @ 5.7ghz = ~3950points according to that scaling
Almost seem too good to be true 👍

i hope this is true only to see the reaction on that eeediot website, what was it called again, oh yeah userbenchmark ...

 

[Hitman928]

Zen5 @ 5.7ghz = ~3950points according to that scaling
Almost seem too good to be true 👍

GB doesn't scale linearly with clock speed across a wide frequency range, but it usually isn't too far off. It can be accurately approximated as linear if the frequency difference is relatively small, but with multiple GHz, it's going to be less than linear.

 

[Hitman928]

are you saying the strix point GB bench is fake?

also the theory that "linear 5.1 will be 3500+" is plain wrong

just did 7840hs gb6.3 @ base freq with boost disabled

1975 single
10848 multi
~20% difference

which means at very best case strix point at max boost (?) has around 2800 single 15000 multi

GB says your higher run was run at between 4.6 - 4.7 GHz. So from base, that is approximately a 22% increase in frequency for 21% more performance, pretty close to linear.

Edit: Comparing your base freq. run to the STX point result, the STX CPU has a 29% higher score while running at a ~4% slower frequency. Not sure why you think that is extremely disappointing. If it ran at 5.1 GHz, it should score at least 3350 pts, if not higher (3554 would be perfect scaling but it won't quite get that).

 

[adroc_thurston]

Comparing your base freq. run to the STX point result, the STX CPU has a 29% higher score while running at a ~4% slower frequency. Not sure why you think that is extremely disappointing. If it ran at 5.1 GHz, it should score at least 3350 pts, if not higher

Bingo.
Now, if you look at the subtests, that's where it gets more interesting.
The INT side is pumping the score.

 

[inf64]

GB says your higher run was run at between 4.6 - 4.7 GHz. So from base, that is approximately a 22% increase in frequency for 21% more performance, pretty close to linear.

Edit: Comparing your base freq. run to the STX point result, the STX CPU has a 29% higher score while running at a ~4% slower frequency. Not sure why you think that is extremely disappointing. If it ran at 5.1 GHz, it should score at least 3350 pts, if not higher (3554 would be perfect scaling but it won't quite get that).

I was about to reply to that post but you beat me to it again. The scores are looking good so far, but we need to wait for granite ridge leaks to be sure.

 

[H433x0n]

So it’s official AT canon that Strix does 3300+ GB6? So we’re going to have the hype train derail twice?

 

[Mahboi]

This thread is turning into a lesson into what NOT to rely on.
Nothing's official until reviews. NOTHING. More than ever there isn't a single leak or projection that should be trusted.

 

[inf64]

So it’s official AT canon that Strix does 3300+ GB6? So we’re going to have the hype train derail twice?

Well nothing is certain at this point. Same goes for Lion cove copium/hopium that it can get noticeably faster in Arrow Lake

 

[gdansk]

So it’s official AT canon that Strix does 3300+ GB6? So we’re going to have the hype train derail twice?

Go be annoying elsewhere.

The official canon is Phoenix times 1.16 until proven otherwise.

 

[HurleyBird]

Go be annoying elsewhere.

The official canon is Phoenix times 1.16 until proven otherwise.

Could be 1.4 in GB. It's still going to be ~1.16 across a range of tasks.

 

[invisible_city]

Dont post much, but wow, shock horror, what a revelation this is

I was trained by Jack Handey

 

[invisible_city]

Well nothing is certain at this point. Same goes for Lion cove copium/hopium that it can get noticeably faster in Arrow Lake

Intel said it could. It’s not hopium except to hope Intel stopped redacted.


esquared
Anandtech Forum Director

 

[gdansk]

Could be 1.4 in GB. It's still going to be ~1.16 across a range of tasks.

It could be anything but they gave us a score comparison to SDXE (with inaccurate footnotes lol) that suggests it is closer to 1.16x in GB6 too.