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

[DisEnchantment]

Speculate at will

 

[eek2121]

I would love to have Strix Halo in a sub 3lb laptop. Hopefully OEMs pay attention.

I mean they paired the Remembrandt CPUs with dGPUs when the actual upgrades is its iGPU. And do we know anything about Strix Halo's configs besides the highest end configuration with 16 Zen 5 and 40CU? I mean 8 + 4c for a lower spec configuration is nice too, but the iGPU CUs though...also I'm just mostly looking at Legions for best Linux compatiability anyways and that's why I'm worried.

Otherwise thanks? But the prices are really off for your speculations, and LPCAMM 2 from Crucial seems to be the only logical RAM module that would be paired with this APU, with 128 bit LPCAMM 2 9600 per module(for a full 256 bit with 2 modules), otherwise a slightly odd 24~48GB RAM config from SK Hynix

Some “OEMs” are nearly exclusively B2B.

Who cares? The content is from Andrei F.

Cinebench is a terrible general purpose CPU benchmark. One of the worst. It does not correlate with anything most consumers use.

Here's exactly what Andrei F said:

Dr. Ian Cutress has said similar things as well.

Also, I suspect folks would retract those statements given that GB6 changes how things are scored.

GB5 is a good benchmark, as long as you know about garbage-in-garbage-out.

Weird how people are so attached to a benchmark.

Yikes.

Cinebench does not correlate with SPEC. SPEC Is the industry's standard. SPEC correlates with Geekbench. We know Cinebench inflates AMD numbers. That's why it's so popular here.

It isn’t that it inflates AMD’s numbers, it just scales extremely well with more cores.

mikegg is not the first, nor the best to make this claim.

This has been said by more experienced and knowledgeable people a long time ago.

SPEC and Geekbench 5 put the CPU through similar types of workloads. IMO SPEC is more comprehensive, but Geekbench gives the results quicker.

 

[dr1337]

Both these benchmarks are good for users that care about software rendering. That probably is less than 1% of computer users. That looks like a poor choice to assess "real world performance". You need several benchmarks for that.

Actually doing something is more real world than a completely arbitrary benchmark. Usually you use your computer to make some kind of product yes? Like text input on a forum. Or probably something a bit more complicated if you get paid to use one. Synthetic benchmarks like geekbench aren't actually doing anything substantial, and its impossible to reflect its scores onto real world performance.

Like if zen 5 has 1000 higher score, does that mean documents will open 5% faster? What does a geekbench score actually mean in terms of "real world performance"?

 

[FlameTail]

Can't you guys (including the meme poster) give it a rest already and simply take individual benchmarks for what they are? And how about we all went back to the thread's topic for a change?

yo is that a reference to me?

 

[adroc_thurston]

And absolutely meaningless when it comes to measuring CPU performance.

What's with the cope?
GB6 nT test is meaningless since it's not a throughput workload and it has no relevance to the only client nT workload that matters anyway (gaming).
Cinememe is ok.

For AMD, they don't usually upstream compiler patches unless the CPU is launched.
Forget the patches, you wont even see the manual or the optimization guide.

Yeah man, money.
It's an issue they solved with money.

 

[Doug S]

So what benchmarks are better? Because as far as I can tell Cinebench is 2nd only to the Blender benchmark in terms of real world performance impact.

Microbenchmarking isn't without its merits but I have pretty much never found geekbench numbers to correlate much to anything I do on a computer/device. My iPhone gets higher geekbench scores than an old haswell rig, but when trying to do things like encode a video (yeah I tried to CPU encode on a phone, fight me), it's obvious the iPhone SOC was not even remotely close to the PC in terms of performance.

Video encoding is a specialist task, the performance of which depends heavily on SIMD instructions when it is done with the CPU alone. So why should anyone be surprised that Cinebench correlates with it, and Geekbench which does a wide variety of tasks of which encoding is only a small piece does not correlate with it as well.

If you pick some really narrow task that exercises only part of the CPU then yeah you will see correlations with benchmarks that depend heavily on that part of the CPU. That task won't correlate as well with benchmarks that exercise the entire CPU.

 

[desrever]

7zip is the best

 

[gdansk]

I determine CPU performance by compiling and running a specific 1T benchmark I wrote many years ago in university.
Anyone who disputes me is, in fact, wrong as this is the best benchmark.

I'm joking but I do still run it each time I get a new CPU. It's interesting to see the progress for code written by a normal human programmer who was not trying to be clever or fast. My phone is now 3.45x faster 1T than the Core 2 Duo I had at the time. Well, some of that might be compiler advances because I gave away that C2D long ago and can't rebuild & retest. And that's a problem with pretty much every benchmark.

 

[adroc_thurston]

Well, some of that might be compiler advances

A lot more than you think.
Modern compilers are truly magical things.

 

[APU_Fusion]

285 pages into zen 5 thread and we talking 7zip and compiling. Love these forums

 

[mikegg]

It isn’t that it inflates AMD’s numbers, it just scales extremely well with more cores.

It doesn't just scale extremely well, it also favors many slow cores over fewer fast cores. It also heavily benefits from SMT. It's the exact opposite of common applications that people actually use.

Most common applications, such as browsers or Excel, benefit from fewer very fast cores over many slow cores. SMT has almost no effect. And having many cores does not matter. This is why GB is a better benchmark, by far.

Cinebench is a terrible general purpose CPU benchmark. It's exactly what Andrei F alluded to in the Reddit thread.

It isn’t that it inflates AMD’s numbers, it just scales extremely well with more cores.

It inflates AMD's numbers in a way that it has now become the defacto CPU benchmark due to marketing and people think Cinebench results equate to actual CPU speed for other applications.

 

[mikegg]

What's with the cope?
GB6 nT test is meaningless since it's not a throughput workload and it has no relevance to the only client nT workload that matters anyway (gaming).
Cinememe is ok.

That's just silly. GB6 is a consumer benchmark designed to test CPUs for applications that most consumers use.

Cinebench tests how good a CPU is for Cinema 4D at best. I say at best because most Cinema users will render with the GPU, not CPU.

 

[Markfw]

It doesn't just scale extremely well, it also favors many slow cores over fewer fast cores. It also heavily benefits from SMT. It's the exact opposite of common applications that people actually use.

Most common applications, such as browsers or Excel, benefit from fewer very fast cores over many slow cores. SMT has almost no effect. And having many cores does not matter. This is why GB is a better benchmark, by far.

Cinebench is a terrible general purpose CPU benchmark. It's exactly what Andrei F said in the Reddit thread.



It inflates AMD's numbers in a way that it has now become the defacto CPU benchmark due to marketing and people think Cinebench results equate to actual CPU speed for other applications.

You assume everyone uses browsers and excel. Many people (my sons company for one) use CAD software, and many other people use encoders, compilers, etc.

Your logic is flawed.

 

[mikegg]

So what benchmarks are better? Because as far as I can tell Cinebench is 2nd only to the Blender benchmark in terms of real world performance impact.

What do you mean by real world performance impact?

Cinebench has weak correlation even with things like Blender rendering. It doesn't even correlate with parallel CPU code compilation.

There are better CPU benchmarks out there. GB is far better.

 

[mikegg]

You assume everyone uses browsers and excel. Many people (my sons company for one) use CAD software, and many other people use encoders, compilers, etc.

Your logic is flawed.

And yet, GB would correlate better with CAD software, encoders, and compilers more than Cinebench.

 

[Markfw]

And yet, GB would correlate better with CAD software, encoders, and compilers more than Cinebench.

This thread needs to get back to a Zen 5 discussion, not a benchmark discussion.

 

[FlameTail]

We need to revive that benchmark thread

Edit: Revived it. You guys can take the discussion there

 

[adroc_thurston]

CPUs for applications that most consumers use.

Like over half the subtests are server stuff.

And yet, GB would correlate better with CAD software, encoders, and compilers more than Cinebench.

No it doesn't.
Not even 1t test is a good proxy and the less we speak about nT the merrier.

Cinebench tests how good a CPU is for Cinema 4D at best.

It's a basic throughput test and it's a nice proxy for CPU perf that way.

 

[mikegg]

It's a basic throughput test and it's a nice proxy for CPU perf that way.

A good general purpose CPU benchmark correlates to performance of common applications. GB succeeds. Cinebench fails.

If you're going to use one benchmark to get a proxy for CPU performance, use GB over Cinebench. Period.

I will say no more in this thread since people want to focus back on Zen5.

 

[adroc_thurston]

A good general purpose CPU benchmark correlates to performance of common applications.

If it benched realistic worksets yeah.
GB isn't that.

Cinebench fails.

Cinememe is an alright throughput test, stop the cope already.

If you're going to use one benchmark to get a proxy for CPU performance, use GB over Cinebench. Period.

GB has no usable nT test so it gets disqualified to begin with.

 

[Doug S]

I determine CPU performance by compiling and running a specific 1T benchmark I wrote many years ago in university.
Anyone who disputes me is, in fact, wrong as this is the best benchmark.

I'm joking but I do still run it each time I get a new CPU. It's interesting to see the progress for code written by a normal human programmer who was not trying to be clever or fast. My phone is now 3.45x faster 1T than the Core 2 Duo I had at the time. Well, some of that might be compiler advances because I gave away that C2D long ago and can't rebuild & retest. And that's a problem with pretty much every benchmark.

Too bad you can't recompile, it would be interesting to know how much of that is compiler improvements vs hardware.

 

[DrMrLordX]

Too bad you can't recompile, it would be interesting to know how much of that is compiler improvements vs hardware.

That's one reason why benchmarking should shift towards open source wherever possible. Everyone can see and change the code, no questions asked, no upfront fees or binary blobs.

 

[itsmydamnation]

It doesn't just scale extremely well, it also favors many slow cores over fewer fast cores. It also heavily benefits from SMT. It's the exact opposite of common applications that people actually use.

Most common applications, such as browsers or Excel, benefit from fewer very fast cores over many slow cores. SMT has almost no effect. And having many cores does not matter. This is why GB is a better benchmark, by far.

Cinebench is a terrible general purpose CPU benchmark. It's exactly what Andrei F alluded to in the Reddit thread.



It inflates AMD's numbers in a way that it has now become the defacto CPU benchmark due to marketing and people think Cinebench results equate to actual CPU speed for other applications.

Do you even excel ? People who drive excel that actually use serious cpu will have more then one instance open, right now I have 15, the largest with like 300 sheets full of path loss and link budget calculations. The fun part is this spreadsheet is generated by Perl and takes a 16core 32 thread zen3 about 36 hours to run at 100 cpu utilisation.

 

[lightmanek]

I suspect our friend here has vested interest in promoting GB6.

Anyway, I've been benchmarking computers since C64 times, and never have I wanted a benchmark to only load my CPU 20% because it is "realistic". I want to see how processor does in tasks I do, but I also want to see how it works internally and how much it improved from previous generations.
I keep score database with things as old as SuperPi and WPrime, or even obscure fractal generators optimized for x87, not because it reflect web browsing experience, but because it tells the story of what improvements or regressions new architecture had to make to perform the way it does.

In the end it helps me to give a good advice to my customers of best options for the tasks they employees need to run on their PCs. Some CNC machines are still using control software from early 2000s and heavily rely on SSE or even old x87 instructions and a lot of CFD easily loads all cores and all threads of even 64 core (+) monsters.

Advocating for single benchmark to be some kind of oracle of architecture performance and perfect reflection of software stacks running on all the devices in the world is a bit ridiculous for me.

For one, when I get Zen5 I will be very interested to see how it runs unoptimized old codepaths.

 

[FlameTail]

Do you even excel ? People who drive excel that actually use serious cpu will have more then one instance open, right now I have 15, the largest with like 300 sheets full of path loss and link budget calculations. The fun part is this spreadsheet is generated by Perl and takes a 16core 32 thread zen3 about 36 hours to run at 100 cpu utilisation.

I didn't know Excel could be such a monster

 

[Nothingness]

Actually doing something is more real world than a completely arbitrary benchmark. Usually you use your computer to make some kind of product yes? Like text input on a forum. Or probably something a bit more complicated if you get paid to use one. Synthetic benchmarks like geekbench aren't actually doing anything substantial, and its impossible to reflect its scores onto real world performance.

Like if zen 5 has 1000 higher score, does that mean documents will open 5% faster? What does a geekbench score actually mean in terms of "real world performance"?

Geekbench isn't a synthetic benchmark. It even has parts of compilers, browsers and so on. You really should read about it before making such unsubstantiated claims: https://www.geekbench.com/doc/geekbench6-cpu-workloads.pdf