Dual Epyc 7742 and stockfish

[yorkman]

Could Patrick or someone tell me what kind of difference in NPS can I expect when running Stockfish chess engine with 2x Epyc 7742 cpu's in a Supermicro H11DSi-NT motherboard with DDR4-2666 MHz ram or DDR4-3200MHz ram?

I read that in an AMD it can be several percent (~5) faster in benchmarks, but not much with Intel. Buying 3200MHz is a lot more expensive than 2666MHz so I wonder if it's worth it.

Similarly I heard it's better to use all your ram slots and use 1R (single rank) ram than it is to use 2R (dual rank). Is this true, even with ECC ram?

In any case, Patrick got 275,000,000 NPS with 2x 7742 but I don't know what ram he used. If it's 3200 MHz then it's important for me to know if I'd lose a lot of NPS by buying 2666 instead??

 

[Patrick]

For benchmarking the EPYC 7742 we are using DDR4-3200 8x 32GB.

We do not have a H11DSi-NT and the EPYC 7742's never end up in systems with DDR4-2666.

Sorry, those CPUs are in huge demand in the lab and we only have two. It is very hard to free them for custom benchmarks.

 

[yorkman]

It's ok. I wasn't asking for a benchmark. Curious. Why do these cpu's never end up in systems with DDR4-2666? Who says and where do they not up in these systems?

 

[Psycho_Robotico]

Sorry, those CPUs are in huge demand in the lab and we only have two.

Time to throw more money at AMD then

I'd be interested as well regarding the performance impact of different RAM clocks on Naples/Rome. With lower core counts those massive caches might help mitigate slower RAM.

 

[Patrick]

It's ok. I wasn't asking for a benchmark. Curious. Why do these cpu's never end up in systems with DDR4-2666? Who says and where do they not up in these systems?

I was specifically talking about the two CPUs we have. On a relative scale DDR4-3200 does not cost that much more so we always use the top rated speed memory with them.

Now, the lower end chips like the 7232p, 7272, and 7282 are going to more often be found with DDR4-2666.

 

[yorkman]

Thanks Patrick. One more question. If I went with 8x 16 GB DDR4-3200 (still essentially same quad setup per cpu as your benchmark test only half the total ram) do you think I would get a slower benchmark result than the 275,000,000 NPS you got? This is already going to cost me about $1300 on Newegg.ca/Amazon.ca so I'd rather not double that cost right now since those two 7742's are going to flush my bank account already :-/ Because I won't be happy if I get anything less than the speeds you got.

 

[Patrick]

No idea really. We have a version of the engine we run in a controlled setup and have focused on Rome boards. Our goal is to control the test setups then show relative performance. If you are using different software, cooling, BIOS settings, platforms I cannot say the exact value you are going to get.

 

[yorkman]

Were you using the 2x 7242's in an AMD Daytona_X motherboard? I ask because I can't find ANYthing about this motherboard. There's absolutely no data. Nobody's selling them either. No reviews. Not even any news about it being released maybe sometime later this year. It's as if it doesn't exist yet I've seen a couple of people using it for their benchmarks. So I can't even compare it to the Supermicro H11DSi-NT to get an idea why the Daytona_X might be so superior (I doubt it in terms of pure calculation horsepower).

Also, when I get this system up and running I'll be trying to mimick your test. That means running the same chess engine (asmFishW_2017-05-22), at 3.40 GHz (max. boost) with HT on (256 threads) and 1024 MB hash. As for the rest I plan on using the same amount of ram. Not sure what BIOS settings you could've used that would give it so much performance boost other than making sure that all cores are enabled with HT on.

Can I ask what two coolers you used and the exact memory model for the 8x32GB Micron's you used?

 

[i386]

Not sure what BIOS settings you could've used that would give it so much performance boost other than making sure that all cores are enabled with HT on.

Check your bios for the performance profiles, power settings and stuff like that

 

[Patrick]

It is not a motherboard. It is a system complete with heatsinks. Technically it is an AMD system, manufactured by a major OEM, that they do not sell publicly. I am not supposed to disclose the OEM that made it for AMD.

I know that is a frustrating answer.

 

[alex_stief]

Similarly I heard it's better to use all your ram slots and use 1R (single rank) ram than it is to use 2R (dual rank). Is this true, even with ECC ram?

Not sure which circumstances led to that recommendation, but with fixed memory frequency and timings (as is usually the case with server-grade hardware), the opposite is true. Dual-rank DIMMs get a slight edge thanks to rank interleaving. ECC has nothing to do with it.

Also, with 15000$ worth of CPUs alone, I would not make any compromises with memory. Fill all channels with 3200MT/s DIMMs.
With 64 cores per CPU, having only 4 memory channels will become a limiting factor for a lot of workloads that are generally not considered to be bandwidth limited.

 

[yorkman]

Not sure which circumstances led to that recommendation, but with fixed memory frequency and timings (as is usually the case with server-grade hardware), the opposite is true. Dual-rank DIMMs get a slight edge thanks to rank interleaving. ECC has nothing to do with it.

Also, with 15000$ worth of CPUs alone, I would not make any compromises with memory. Fill all channels with 3200MT/s DIMMs.

With 64 cores per CPU, having only 4 memory channels will become a limiting factor for a lot of workloads that are generally not considered to be bandwidth limited.

Thank you for your feedback. I was told the opposite by someone else just prior to purchasing my dual E5-2696v3 system. Now you tell me that 2R is better than 1R. You sound more knowledgeable than him so I'll probably take your word for it but if possible I'd like some reference to that being the case.

As I mentioned earlier, this system will be solely used for running chess tournaments (you must think "what a waste") and it'll run 24x7x365 at 100% cpu (chess engines will always use max cpu usage). And always running calculations with about 75% ram usage. Any percentage of a performance boost helps in my case so that is why I also asked if it usually makes things worse by using all ram slots with either 16x8 GB (1R), 16x16 GB (2R) modules as opposed to 8x32 GB (2R).

From experience have you noticed such performance boost or degradation or read anything about these comparisons? If filling all slots with higher sized memory is faster than filling all slots with lower sized memory than that's what I'll try to buy. But this will depend on the cost as initially I won't be able to afford buying 16 modules at $240 CDN a pop. Maybe 8 at first then slowly add on the rest but even that's pushing my already depleted bank account.

 

[yorkman]

It is not a motherboard. It is a system complete with heatsinks. Technically it is an AMD system, manufactured by a major OEM, that they do not sell publicly. I am not supposed to disclose the OEM that made it for AMD.

I know that is a frustrating answer.

Under "Motherboard" it said AMD Daytona_X so I thought it was a motherboard. Ok. Can you at least tell me the exact ram you used? I know it's 8x 32GB (256 GB total) but can you provide the model number?