Can a Ryzen 2600-2700(X) owner run some benches please?

[Scu84St3v3420]

Don't you hate it when the dumb guy starts interjecting? Me, too, but I'm going to do it anyway.

I'm not sure that running the Ryzen at lower RAM speeds is "objective" at all. RAM speed is inherent to the chip's performance. It's a feature of the chip, so why is that an invalid component if you're measuring chip performance?

I'm not claiming to be smart on this subject (however don't think name calling is necessary), however from what I can infer his intent is to test a single variable. And when one system has ram at 3000 and the other has ram at 2666, those ARE additional variables in his testing. Part of Ryzen's performance is benefiting from faster ram, thus it is part of the chips inherent performance, just because Intel chips do NOT benefit from it in this manner doesn't make it irrelevant.

His initial statements on this subject in another thread suggested that he only wants to test the CPUs with/without smt and see how IPC and HT scales and what difference if any when reducing the number of active cores ryzen vs intel... Honestly this is stuff we all learned in basic high school science classes, variables are just that variables.

To me at least, this is equal to asking what the difference is between a 8700K and a 2700X in pure performance (stock), but then under clocking the 8700K to 4.3 Ghz and overclocking the 2700X to 4.3 GHz, even though the 8700K will hit 4.7 stock and then saying it is irrelevant. (unless clocking both to the same frequency to see how IPC and HT performance compare chip to chip purely because the 2700X can't hit the frequencies that the 8700K can, which there are many published tests of)

Perhaps I am the idiot here, but this is just how I am reading it.

 

[Scu84St3v3420]

On another note, I do own a 2600X and would be willing to run some tests if the OP would like. Just need some links to the software I need and exact variables intended to be tested.

 

[Alaric]

I was referring to myself as the dumb guy, not anyone else. Mackerel and ED are always leaving me in the dust with their knowledge of Things Computer., so I tossed in some self deprecation.

 

[Scu84St3v3420]

I was referring to myself as the dumb guy, not anyone else. Mackerel and ED are always leaving me in the dust with their knowledge of Things Computer., so I tossed in some self deprecation.

Haha, I get it now, yea I am quite green to the concepts, but basic scientific method applies in my eyes. Variables are variables, if attempting to test a single variable, then all other variables must remain constant.

 

[Alaric]

When mackerel gets on a roll I usually end up asking the wrong questions, so I figured I'd acknowledge the fact that I probably drive him nuts sometimes.

 

[mackerel]

On another note, I do own a 2600X and would be willing to run some tests if the OP would like. Just need some links to the software I need and exact variables intended to be tested.

The link to 3DPM is in the OP link via twitter. Cinebench R15 is well known but the download link is at https://www.maxon.net/en/products/cinebench/

In light of recent testing, I'd suggest a variation to the method:
Download the whole 3DPM file set in one go, not individually (dropbox sucks)
If necessary edit the batch file so the first number in it is a "3" for 3 runs. I don't know what the 2nd number does.
Reduce the system via bios to 2 cores.
Back in Windows, allow system to stabilise and finish whatever background stuff it might want to do. Close any unnecessary apps that might take CPU time. This includes hardware monitoring and even task manager.
Run the batch file and allow it to complete 3 runs. Either copy as text or screenshot the complete results. Note the save file only seem to have the average overall score. If run via windows, note window will close on a keypress so be careful. This obviously doesn't happen if run from cmd.
Repeat with SMT on/off.
With results saved, in another test run, check what clock the CPU is at when running these tasks. Alternatively, fix it to any known clock before starting testing. The new precision boost thing might make it more variable than previously...

For Cinebench R15, just run the multicore test 3 times with and without SMT, and let me know the highest of those results.

Some numbers to act as a sanity check, the overall score for the 1600 on 3DPM in 2 cores 4 threads was 591 and I got 423 in CB15. It would seem reasonable to expect the 2600X to score above that.

@ed on the taking a baseline thing, that's what that testing was. Again, for my eventual needs, the ram speed will be factored out as far as is practical. I will, at some point, dig out my 1700 system also. I think that had faster ram in it, and will serve as a double check of the 1600 results. Even then, it is only going to run at 2933. I don't know I have any ram that runs easily at 3200 on Ryzen.

I also ran into another problem during testing... it doesn't look like I can adjust core clock while limited to 2 cores. At least, if I attempt to do so in bios, it wont boot. So I was unable to do the underclock test. I know I could use Ryzen Master, but I find that a bit of a mess and extremely unreliable... The next-best thing I could do is run 4 cores and see if that scales closer to 2 or 6 cores. If I'm lucky, and it is closer to 2 core scaling, then I think that would be sufficient to declare the 2 core results as "good enough".

 

[Scu84St3v3420]

I will look into it on Wednesday, it's my day off... CB R15 I have and can do those tomorrow night after work.

 

[mackerel]

Above is a quick screenshot of the output of 3DPM that I'm after.

I should add, this is the Ryzen 1600 system, still at 2 cores 4 threads, but now with 2933 ram! Let's try and settle that once and for all... I remembered the kit I had in there, although rated at 2666, I had in the past accidentally overclocked when moving it between systems. I thought I'd try my luck again. With XMP still on, I changed the speed from 2666 to 2933, increased the voltage from 1.20 to 1.35 as is typical for the speed, and crossed my fingers. It booted! No stability testing whatsoever, this is just a quick bench run.

I compared the subscore results for 3DPM. Depending on subtest, the 2933 ram run was -0.2% to +0.3% faster, with an overall average change of +0.1% from the faster ram. Would we agree this is insignificant and within reasonable test variations? On a similar note, CB15 went up 1 point from 423 to 424, which is +0.2%. As mentioned before, I'd be happy with 1% tolerance on the results. Both tests were best of 3 runs.

 

[Alaric]

I compared the subscore results for 3DPM. Depending on subtest, the 2933 ram run was -0.2% to +0.3% faster, with an overall average change of +0.1% from the faster ram. Would we agree this is insignificant and within reasonable test variations? On a similar note, CB15 went up 1 point from 423 to 424, which is +0.2%. As mentioned before, I'd be happy with 1% tolerance on the results. Both tests were best of 3 runs.

I was wondering if the before and after SMT numbers would scale the same regardless of RAM speed. With my limited grasp that seems to be the case. Cool.

 

[mackerel]

A quick update on testing, which I forgot to post here. I did the 4 core testing also, to fill in the gap between 2 and 6 cores. Compared to 2 cores, both the 3DPM and CB15 scores at 6 cores were 2% lower per core per clock. The 4 core results came out 1% lower per core per clock. So... a kinda trend there. If it were possible to run 1 core maybe it'll be fractionally faster than 2 cores, but it isn't practical to set that up.

I also did some live data testing with GCW sieve on PrimeGrid BOINC project. This turned out not so great... I had overlooked that GCW sieve has various different work unit types, and they can take different durations. I tried both average (mean) and median to see if I can pick out some result, and it was consistent. The 1600 (6 cores, stock) showed 43% improvement with SMT on. The 7800X showed 38% improvement with HT on. I didn't test the 8086k as that was more important to be used for gaming I haven't tried to look at the absolute per clock per core between them yet.

 

[Alaric]

I had thought it was established that AMD's implementation of SMT was more efficient than Intel's HT? Wasn't that one of the big deals when Ryzen came out?

 

[EarthDog]

Yep. It is. Many reviews have shown this already. I think Mack is in his rabbit hole for his specific testing though.

 

[mackerel]

I had thought it was established that AMD's implementation of SMT was more efficient than Intel's HT? Wasn't that one of the big deals when Ryzen came out?

You could argue Ryzen sees a bigger SMT gain as it starts with lower single thread performance. However the core has more execution potential, so the 2nd thread per core is better able to make use of those extra resources. What I'm looking at is how much that is. This is just the start... and I need to add 2nd gen too.

 

[Alaric]

And I poked the HT/SMT bear several posts back, trying to keep up with the IPC testing methodology.

 

[EarthDog]

You could argue Ryzen sees a bigger SMT gain as it starts with lower single thread performance.

Explain plz...

If it is fast or slow, HT/SMT scaling doesn't have much (anything?) to do with IPC. Again more TOTAL work is being done with HT/SMT, but that isn't because IPC has increased (its the same for each thread, just more threads - remember the highway analogy?), just the efficiency of HT/SMT that comes into play.

 

[mackerel]

ED, just for you, I'm avoiding the use of those three letters Let's call it throughput?

It is generally accepted that Ryzen (at least 1st gen) has lower single thread performance than recent Intel at a given clock. That's not because the cores are weaker, just that a single thread isn't able to extract the potential. The Zen cores do have more general execution potential (outside of my special interest cases), and running two threads per core is better able to extract that. Thus, when SMT is enabled, you will see bigger increases.

If my understanding is correct, single thread per core, Zen should be slower than equivalently clocked Intel. Two threads per core, Zen should be faster than equivalently clocked Intel. Over time, I intend to build up a database of results and see if this is the case or not.

 

[EarthDog]

Don't do it for me, do it because to use the right terminology.

just that a single thread isn't able to extract the potential.

This is an assumption, no?

and running two threads per core is better able to extract that.

...also an assumption, yes?

I'm not sure why you are tieing it to a 'lack of potential' in single core based solely off of its IPC performance. Is it possible the efficiency of the arcitecture the reason for the apparent efficiency increase of SMT while also not being related to IPC? I have no idea how you can test this, properly. Why would sharing resources increase potential? Wouldn't there be more cache hits and other steps needed when sharing a resource like that? This is where the architecture comes in.

Really, I don't know either. I just don't find the same association you do with IPC/Throughput being slower has to do with SMT being faster when the resources are now shared. I think the increases have more to do with the infinity fabric or w/e they call it than it does 'starting off weak'. Your words are describing anything that doesn't use SMT threads are being held back. I don't buy that . If anything, AMD would have matched single thread performance and SMT with Intel for a better showing. But AMD, they love to go wide (more cores) than fast so... who knows really.

 

[mackerel]

Part of it is assumption on my part. I think I previously mentioned Agner Fog's microarchitecture guide. https://www.agner.org/optimize/ document #3 on that page. The document is aimed at a different audience, but it is still very readable and I do recommend checking it out. Note in particular the descriptions around Zen's execution potential, as well as the capability of Intel's FPU. I'd love to have the ability to poke around at such a low level to see how things really work, but I'm not at that level. The best I can do, is to throw workloads at it, isolate or eliminate variables that are not of interest, and see where I get that way.

Edit: thinking more, I guess I'm testing "what" the outcome is. Saying "why" is speculation on my part, unless I can back it up with external references.

 

[mackerel]

Back to the original reason for creating this thread, I wont need anyone else to run the bench now as I've ordered a 2600 and Asrock B450 ITX to go with it. There was a minor price drop on the 2600 and that kinda sealed the deal with my indecision. Don't think I can justify the gap to the 2600X now, and certainly not to the 2700 models.

I did go for the "free" shipping so I might not get the bits for a week.

 

[(G{in}[AK)TION]]

I just got my ryzen 2700 in the mail today. you still want me to run those tests? can do it on Sunday night, Monday or Wednesday