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

[DisEnchantment]

Speculate at will

 

[Glo.]

yeah Linux ... is a mess ...

Fixed it for ya.

 

[adroc_thurston]

Fixed it for ya.

No it works fine, just that for benchmarking you have to equalize OS setups to a tee.

 

[H433x0n]

Something I haven’t seen mentioned. GB5 scores on Linux tend to be ~5% higher than they are on Windows. They also have wild swings in either direction artificially inflating / deflating the result.

 

[tamz_msc]

The simpler explanation is that the score in the morning is while running a hypervisor, and the score in the afternoon is running without hypervisor.

 

[gdansk]

Something I haven’t seen mentioned. GB5 scores on Linux tend to be ~5% higher than they are on Windows. They also have wild swings in either direction artificially inflating / deflating the result.

Or more: https://browser.geekbench.com/v5/cpu/compare/22070686?baseline=22048770

But that's why we were trying to compare only Linux results to Linux results.

 

[H433x0n]

Or more: https://browser.geekbench.com/v5/cpu/compare/22070686?baseline=22048770

But that's why we were trying to compare only Linux results to Linux results.

Have we gotten a Phoenix test at 2.0ghz in Linux yet?

 

[CouncilorIrissa]

Or more: https://browser.geekbench.com/v5/cpu/compare/22070686?baseline=22048770

But that's why we were trying to compare only Linux results to Linux results.

Everyone's favourite meme distro, Clear Linux. Surprised to see it still being maintained.

You shouldn't have dug that up still.

 

[gdansk]

Everyone's favourite meme distro, Clear Linux. Surprised to see it still being maintained.

You shouldn't have dug that up still.

But using Intel's optimized Linux distribution on AMD hardware? I had to do it.

Have we gotten a Phoenix test at 2.0ghz in Linux yet?

Someone else will have to do it, I can't get it to work at 2000MHz.

 

[adroc_thurston]

Surprised to see it still being maintained.

An Intel project isn't it? Pretty good actually for what it's worth.

 

[CouncilorIrissa]

An Intel project isn't it? Pretty good actually for what it's worth.

Yep, some of its patches regularly trickle down into the mainline kernel, apparently.
Pretty unusable outside of this, though, or at least it seemed to be last time I tried it (~4 years ago).

But using Intel's optimized Linux distribution on AMD hardware? I had to do it.

Oh, those are your results. Oh god, what are you doing with your life. lol

 

[gdansk]

Oh, those are your results. Oh god, what are you doing with your life. lol

I'm not insane: I only ran it from USB to run some benchmarks.

 

[StefanR5R]

The *timestamps* are there on the entries

The timestamps of the firmware/ microcode?

 

[yuri69]

The problem is where is no universal "Linux" as a platform. There are completely different kernel + user space setups and their versions. Besides, anyone can configure and build their own kernel - optimize for latency, throughput, powersaving, or something else? Also the compiler flags can result in a huge performance delta - compare a binary built by a "rice" setup which fits to a specific CPU family and has all -Ofast/LTO/PGO/* enabled to a generic i686 binary.

Windows has a build number and that's it.

 

[tamz_msc]

The timestamps of the firmware/ microcode?

No, timestamps of when the results were uploaded online.

 

[Gideon]

Someone else will have to do it, I can't get it to work at 2000MHz.

The x3d chips tend tp be.more locked down, but you should still be able to do it im BIOS, it's just more work

 

[tamz_msc]

The problem is where is no universal "Linux" as a platform. There are completely different kernel + user space setups and their versions. Besides, anyone can configure and build their own kernel - optimize for latency, throughput, powersaving, or something else? Also the compiler flags can result in a huge performance delta - compare a binary built by a "rice" setup which fits to a specific CPU family and has all -Ofast/LTO/PGO/* enabled to a generic i686 binary.

Windows has a build number and that's it.

FWIW, both these ES and my TGL laptop are running kernel 6.8

 

[Mahboi]

Something I haven’t seen mentioned. GB5 scores on Linux tend to be ~5% higher than they are on Windows. They also have wild swings in either direction artificially inflating / deflating the result.

Don't know how much of a benchmark it is, but the same differential is found in Factorio.
Just chalk it up to generally superior Loonix memory management.

No, I have not yet been released from the Factory.

 

[Mahboi]

Everyone's favourite meme distro, Clear Linux. Surprised to see it still being maintained.

You shouldn't have dug that up still.

What makes it more memetic than say, 90% of non-standard distros?

 

[H433x0n]

So far the most apples/apples comparison we have is against Tiger Lake:

Both are at same frequency
Both are running same Linux Kernel
Both are using the same build of GB5

The results show Strix is +35% ahead at iso frequency. What does this mean for IPC? How far ahead of Tiger Lake is Golden Cove and Zen 4?

 

[Mahboi]

So far the most apples/apples comparison we have is against Tiger Lake:

Both are at same frequency
Both are running same Linux Kernel
Both are using the same build of GB5

The results show Strix is +35% ahead at iso frequency. What does this mean for IPC? How far ahead of Tiger Lake is Golden Cove and Zen 4?

That's not a valid sample.
The valid sample looks more like 70% above TGL.

 

[Hitman928]

So far the most apples/apples comparison we have is against Tiger Lake:

Both are at same frequency
Both are running same Linux Kernel
Both are using the same build of GB5

The results show Strix is +35% ahead at iso frequency. What does this mean for IPC? How far ahead of Tiger Lake is Golden Cove and Zen 4?

I already did that math, here. It shows that the sample in question is highly unlikely to be a real representation of Zen 5 performance. Most likely a neutered ES on an immature platform.

 

[Hitman928]

That's not a valid sample.
The valid sample looks more like 70% above TGL.

More like 100% above TGL if true.

 

[tamz_msc]

That's not a valid sample.
The valid sample looks more like 70% above TGL.

Why is one valid while the other isn't? How convenient that the "invalid" result is the one where it is massively underperforming expectations.

 

[FlameTail]

Will Strix Point iGPU be able match/exceed RTX 3050 mobile?

This might be a hot-take. In my opinion, a 128 bit APU should be able to match atleast a 2 generation old RTX xx50 dGPU.

 

[adroc_thurston]

Will Strix Point iGPU be able match/exceed RTX 3050 mobile?

how much timespy pts is that

Why is one valid while the other isn't? How convenient that the "invalid" result is the one where it is massively underperforming expectations.

you don't have to cope (especially not here).
You can always go back to the Intel thread and pray for nova lake!