Author |
Message |
|
This current challenge is the first time in a long time that I ran WUs that would be bottlenecked by memory bandwidth on my systems. I have a 7600k and a 9700k, and normally see the 9700k have twice the throughput as the 7600k, which is to be expected given that the 9700k has twice as many cores as the 7600k.
The 7600k has 4x4=16GB single rank modules at 3200GHz. The 9700k has 2x8=16GB single rank modules at 3200GHz.
During this challenge, I was getting the following results for Woodall:
7600k: ~64000 second run time and ~228000 second CPU time
9700k: ~56000 second run time and ~404000 second CPU time
I would have expected the 9700k to be slightly more than twice as fast as the 7600k given the doubling in number of cores and the difference in generations, but that isn't the case here.
I am making the assumption that the difference is that the 7600k has 4x4 single rank modules, which in dual channel allows for the two sets of dual channel modules to be treated as two separate ranks and allows one to be written to while the other is read from. The 9700k is not able to do this, and reads will block writes and vise-versa.
I just ordered a second set of matching memory to expand my 9700k, and will report back in a week or so with my results. Hopefully it doubles the speed of the system on large tasks.
____________
|
|
|
|
Some timings from my systems for comparison. All CPUs are running stock apart from the 8086k which has turbo disabled as it is LLR unstable otherwise, and the 3700X which has a small positive voltage offset otherwise it is unstable. Running one task with threads = cores.
Timings are an eyeball average of longer Woo units.
Dual channel systems
8c 3700X 2x8GB SR 3600 28ks
4c 5775C 2x8GB DR 2400 59ks
4c 6700k 2x8GB DR 3200 50ks
6c 8086k 4x16GB DR 3200 41ks (turbo disabled)
Quad channel systems
6c 5820k 4x4GB SR 2666 45ks
6c 5930k 8x4GB SR 2133 41ks
12c 7920X 8x8GB SR 3000 16ks
SR/DR applies to individual modules. If there is 1 or 2 DPC can be seen from configuration vs channels.
I'd expect your 7600k to perform similar to my 6700k. There is some 5% clock difference but both CPUs seem to be running slower than comparable systems I have. Looks like you're also running GPU units? Maybe that is making up the difference. Bit late for the current challenge, but if you wanted to max the challenge points discontinuing GPU might have helped somewhat.
Anyway, in the past I did do some testing on 6700k comparing 3200 DR vs 3000 SR, I saw some 25%+ difference between them. I don't have the numbers to hand, but I also compared 3200 DR vs 3600 SR, and the 3200 DR easily won in performance. Doing DR or 2 SR modules per channel does help, but don't necessarily expect a doubling in performance. |
|
|
|
I have tested a lot of different configurations, and double rank is definitely better than single rank when a task requires main memory access. Of course, fitting everything in L3 cache is the best, but when you need memory, go for double rank. |
|
|
|
Ryzen is about 10% faster with dual rank which using currently available ram means you need to be using 2x16GB as 8GB ram is all single rank. |
|
|
|
Some of my setups are quad channel dual rank. That's the best for things like P-1 sieving. |
|
|
Nick  Send message
Joined: 11 Jul 11 Posts: 882 ID: 105020 Credit: 1,318,826,036 RAC: 0
                    
|
Ryzen is about 10% faster with dual rank which using currently available ram means you need to be using 2x16GB as 8GB ram is all single rank.
On my two 9960X which are quad channel, I have 4 x 8Gb ram.
I should swap those for 4 x 16Gb?
My 9980XE has 4 x 16Gb.
I know my 9900K (dual channel) has 32 Gb - I would need to look to find out if it is 2 or 4 sticks of ram. |
|
|
|
Ryzen is about 10% faster with dual rank which using currently available ram means you need to be using 2x16GB as 8GB ram is all single rank.
On my two 9960X which are quad channel, I have 4 x 8Gb ram.
I should swap those for 4 x 16Gb?
My 9980XE has 4 x 16Gb.
I know my 9900K (dual channel) has 32 Gb - I would need to look to find out if it is 2 or 4 sticks of ram.
First thing to do would be to confirm ram type. Grab a copy of cpu-z if you don't already have one and check the SPD tab.
Second, it may be obvious but if you're going to buy new ram make sure what you're buying is actually dual rank. That information can be hard to find pre-purchase. Not all manufacturers mention it.
|
|
|
|
Ryzen is about 10% faster with dual rank which using currently available ram means you need to be using 2x16GB as 8GB ram is all single rank.
On my two 9960X which are quad channel, I have 4 x 8Gb ram.
I should swap those for 4 x 16Gb?
My 9980XE has 4 x 16Gb.
I know my 9900K (dual channel) has 32 Gb - I would need to look to find out if it is 2 or 4 sticks of ram.
First thing to do would be to confirm ram type. Grab a copy of cpu-z if you don't already have one and check the SPD tab.
Second, it may be obvious but if you're going to buy new ram make sure what you're buying is actually dual rank. That information can be hard to find pre-purchase. Not all manufacturers mention it.
Is Crucial's Ballistix 2x8GB dual rank?
____________
SHSID Electronics Group
SHSIDElectronicsGroup@outlook.com
GFN-14: 50103906^16384+1
Proth "SoB": 44243*2^440969+1
|
|
|
|
Pretty much all modern 8GB modules are single rank now. Some were dual rank way back in the early days (2015 or earlier).
With 16GB modules, it is getting more complicated as I hear they are also going single rank, so it is not guaranteed just having a 16GB module is dual rank.
Equivalent to having dual rank modules is running two single rank modules per channel. Running 4x8GB on dual channel systems (or even 4x4GB if you can find smaller modules and don't need the capacity) may be a safer option than going random 16GB modules now. |
|
|
|
An update here after installing the additional RAM.
My 9700k went from
~56000 second run time and ~404000 second CPU time
to
~44500 second run time and ~336000 second CPU time
From just that change. It isn't as dramatic as I hoped, but it will have to do.
____________
|
|
|
|
A 20% improvement is massive. |
|
|
|
A 20% improvement is massive.
I was hoping for 50%. I should try running on just 4/8 cores and seeing how the times compare between the two computers to rule out time being lost due to the additional parallelism.
____________
|
|
|