LL Doublecheck runs hotter than PRP?

[themoon123]

Hi.

I completed a PRP test a few days ago and noticed it ran at approx. 55c. This was considerably cooler than a PRP test on Primegrid with the same amount of cores(8 cores).

I just started a LL doublecheck yesterday and now my CPU is running at around 68c like it did on PG.
The LL test is also quite faster compared to the PRP test. The PRP test used 6M FFT and LL is using 3200K.

The exponents in question are M110913079 and M59990291
PRP was at 4.3 ms/iter and LL is 1.4. Is 4.3 the speed i should expect or is something wrong?
My CPU is an AMD Ryzen 7 5800X @ 4.2 GHz with 2x 8 GB 3200 MHz. I'm using prime95 30.6 build 6 on Windows 10.

I'm quite new to this community so apologies if i misunderstand some of these terms.

[axn]

I am assuming that you were doing a single WU using all 8 cores.

5800X has a 32MB L3 cache. a 3200K FFT takes 3200*1024*8 bytes or 25MB. That will fit within the 32MB L3.

Clearly the 6M FFT, which takes 48MB, won't fit completely within L3. Hence there is efficiency loss.

You can improve by a) using faster RAM (3600 or 4000), and b) using two dual rank DIMMs / 4 single rank DIMMs. The 8GB DIMMs are likely single rank.

These will help improve the memory thruput and give you faster timings (and hotter CPU) on the 6M test

[themoon123]

Originally Posted by axn

I am assuming that you were doing a single WU using all 8 cores.

5800X has a 32MB L3 cache. a 3200K FFT takes 3200*1024*8 bytes or 25MB. That will fit within the 32MB L3.

Clearly the 6M FFT, which takes 48MB, won't fit completely within L3. Hence there is efficiency loss.

You can improve by a) using faster RAM (3600 or 4000), and b) using two dual rank DIMMs / 4 single rank DIMMs. The 8GB DIMMs are likely single rank.

These will help improve the memory thruput and give you faster timings (and hotter CPU) on the 6M test

Thank you! That makes sense. I checked my RAM and it is dual rank 8 gb sticks. I will be looking at some faster 32 GB Quad channel RAM instead. This is enough right?

Appreciate the help!

[axn]

Originally Posted by themoon123

This is enough right?

That's the best we can do; we will never get the efficiency of running entirely from L3 cache.

Just so there is no confusion, there is difference between dual rank and dual channel. If you run matching pairs of sticks in the correct mobo slots, it will enable the dual channel mode in the CPU. That is always a must. If you're running in single channel mode, you have half the bandwidth and half the thruput.

On top of that, using dual rank sticks (quad rank is much rarer) allows the CPU to do simultaneous read/write on the same stick. This is a minor speed improvement - a few % at best. Still, no reason to give up that performance.

So, if the sticks you have are already dual rank, you probably won't get any additional benefit going with quad rank / 4 sticks. There is no quad channel as that is an attribute of CPU; normal Ryzens are dual channel, only Threadrippers have quad channel.

[themoon123]

I see. I guess i will try and increase the speed of the DIMM's.
I looked up the RAM i bought so i'm quite sure it's dual rank.

Thank you for the suggestions.

[Viliam Furik]

Originally Posted by themoon123

Thank you! That makes sense. I checked my RAM and it is dual rank 8 gb sticks. I will be looking at some faster 32 GB Quad channel RAM instead. This is enough right?

Appreciate the help!

There is a difference between rank and channel. Rank is an internal property of the memory module, which I think behaves like an internal channel. Channels themselves are tied to the number of memory slots the motherboard has and CPU supports. 5800X supports dual-channel memory, thus it has up to 4 slots on the motherboard. Quad-channel memory opportunity comes with Threadripper processors, which have 8 memory slots on their motherboards. As you may have noticed, there are usually two memory modules per channel possible.

-----
EDIT: I have now noticed that axn explained it too... But two agreeing sources of information can't hurt anything.