| Index | Recent Threads | Unanswered Threads | Who's Active | Guidelines | Search |
 |
World Community Grid Forums
Category: Active Research
Forum: OpenPandemics - COVID-19 Project
Thread: CPU cache & performance |
| No member browsing this thread |
|
Thread Status: Active Total posts in this thread: 31
|
|
| Author |
|
|
fuzzydice555
Advanced Cruncher Joined: Mar 25, 2015 Post Count: 89 Status: Offline Project Badges: 
|
Since I have 3 generations of Ryzen CPUs running currently, I did a little comparison on per core peformance:
----------------------------------------Ryzen 1700 16MB @3 GHz: 23,7 points/hour/core Ryzen 2700 16MB @3.3 GHz: 25,48 points/hour/core Ryzen 3700X 32MB @3.6 GHz: 34,06 points/hour/core Ryzen 3900X 64MB @3.8 GHz: 47,49 points/hour/core The difference between 1700 & 3900X is absolutely huge. Per core performance is 2x, while total PPD will be around 3x for the 3900X. There is also a 37% difference between the 3700X and the 3900X, while the clocks are only 5% better on the 3900X. I suspect this is due to the huge cache of the 3900X, so I checked MIP as well, which is known to be sensitive to cache: Ryzen 1700 16MB @3 GHz: 33,78 points/hour/core Ryzen 3700X 32MB @3.6 GHz: 58,3 points/hour/core Ryzen 3900X 64MB @3.8 GHz: 72 points/hour/core Here it's the same, 210% performance compared to the 1700, 123% performance compared to the 3700X. I checked MCM, here the differences are much smaller: Ryzen 2700 16MB @3.3GHz: 34,42 points/hour/core Ryzen 3700X 32MB @3.6 GHz: 37,08 points/hour/core Ryzen 3900X 64MB @3.8 GHz: 45,63 points/hour/core Based on this, I'm selling the 1700 & 2700 and upgrading to a single 3900X   |
||
|
|
Henadzi
Cruncher Netherlands Joined: Oct 9, 2009 Post Count: 20 Status: Offline Project Badges: 
|
Good to know, thanks for your contribution!
A bit pity to know about such difference between 3700x and 3900x, I was planning to buy the first one for my new desktop PC...)) |
||
|
|
fuzzydice555
Advanced Cruncher Joined: Mar 25, 2015 Post Count: 89 Status: Offline Project Badges:
|
The 3900X makes the most sense performance/price wise for me:
----------------------------------------3700X: 294$ - 36.75$ / core 3900X: 410$ - 34,16$ / core 3950X: 719$ - 44,94$ / core However the 3700X makes more sense in some cases (gaming, casual use). It is still an extremely efficient processor. If you want to buy one, don't let the performance difference keep you! |
||
|
|
DrMason
Senior Cruncher Joined: Mar 16, 2007 Post Count: 153 Status: Offline Project Badges: 
|
Just a warning about MIP. Each workunit of MIP will use 4 MB of L3 cache. Even with its gargantuan 64 MB, 24 threads of MIP will still result in overall slowdown. I would probably recommend no more than 15 concurrent workunits of MIP on a 3900X. Not only would you be able to crunch the 15 MIP units way more efficiently, but you'd be able to crunch 9 other workunits from other projects as well. Only chips thus far that can run SMT while not running out of L3 cache are some Epyc Rome chips.
---------------------------------------- |
||
|
|
fuzzydice555
Advanced Cruncher Joined: Mar 25, 2015 Post Count: 89 Status: Offline Project Badges:
|
Just a warning about MIP. Each workunit of MIP will use 4 MB of L3 cache. Even with its gargantuan 64 MB, 24 threads of MIP will still result in overall slowdown. I would probably recommend no more than 15 concurrent workunits of MIP on a 3900X. Not only would you be able to crunch the 15 MIP units way more efficiently, but you'd be able to crunch 9 other workunits from other projects as well. Only chips thus far that can run SMT while not running out of L3 cache are some Epyc Rome chips. I limited MIP & ARP to 1 concurrent work units. Even though all of my CPUs should have enough cache, there is still a big difference. |
||
|
|
KerSamson
Master Cruncher Switzerland Joined: Jan 29, 2007 Post Count: 1684 Status: Offline Project Badges: 
|
In a near future, I plan to replace a 10 years old Phenom II x6, @2.8 GHz, by a Ryzen 9 3900X with 12/24 core/thread.
----------------------------------------The 3900X brings the advantage to have more cache available per thread comparing to 3950X (16/32). I operate already 2 Ryzen 7 2700X since one year and they do a very good job. Cheers, Yves --- PS: Covid-19 impacted negatively the planned replacement since all shops were close during the last 2 months. |
||
|
|
M-spec
Cruncher The Netherlands Joined: Jul 29, 2007 Post Count: 4 Status: Offline Project Badges:
|
Since I have 3 generations of Ryzen CPUs running currently, I did a little comparison on per core peformance: Ryzen 1700 16MB @3 GHz: 23,7 points/hour/core Ryzen 2700 16MB @3.3 GHz: 25,48 points/hour/core Ryzen 3700X 32MB @3.6 GHz: 34,06 points/hour/core Ryzen 3900X 64MB @3.8 GHz: 47,49 points/hour/core The difference between 1700 & 3900X is absolutely huge. Per core performance is 2x, while total PPD will be around 3x for the 3900X. There is also a 37% difference between the 3700X and the 3900X, while the clocks are only 5% better on the 3900X. I suspect this is due to the huge cache of the 3900X, so I checked MIP as well, which is known to be sensitive to cache: Ryzen 1700 16MB @3 GHz: 33,78 points/hour/core Ryzen 3700X 32MB @3.6 GHz: 58,3 points/hour/core Ryzen 3900X 64MB @3.8 GHz: 72 points/hour/core Here it's the same, 210% performance compared to the 1700, 123% performance compared to the 3700X. I checked MCM, here the differences are much smaller: Ryzen 2700 16MB @3.3GHz: 34,42 points/hour/core Ryzen 3700X 32MB @3.6 GHz: 37,08 points/hour/core Ryzen 3900X 64MB @3.8 GHz: 45,63 points/hour/core Based on this, I'm selling the 1700 & 2700 and upgrading to a single 3900X Besides the difference in cache size, did you check the difference in memory performance on those machines? Performance of my 3800X@4.2GHz@1.2V is around 20K(x7=140K) per day with 104/52 points/hour/(core/thread). That is about equal to your 3900X per core if it were to run at the same speed. With the same 32MB L3 cache as a 3700X. It runs with four ranks of DDR4 3733CL14. 4x single rank or 2x dual rank(2x16GB) for optimal perfomance with dual channel. Depends on the motherboard, t-topology or daisy chain layout. Swithing from previous 1800X CPU last year there was ~25% gain in performance overall. 64MB cache on 12 cores can have an advantage especially with other projects like MIP and ARP but thats not even enough running them max thread. I have seen performance up to max 10 of 16 threads with ARP with 32MB L3 cache before DDR4 DIMM temperatures max out. DIMM temps are also 6-8°C higher in case of ARP than running OpenPandemics. [Edit 1 times, last edit by M-spec at May 17, 2020 8:59:53 AM] |
||
|
|
fuzzydice555
Advanced Cruncher Joined: Mar 25, 2015 Post Count: 89 Status: Offline Project Badges:
|
Besides the difference in cache size, did you check the difference in memory performance on those machines? ... It runs with four ranks of DDR4 3733CL14. 4x single rank or 2x dual rank(2x16GB) for optimal perfomance with dual channel. Interesting! I run all my machines on the default 2133 setting, as RAM OC made some projects unstable when I tested it on the 1700. Is it possible that faster RAM could mitigate the difference in cache size, that is why your 3800X and my 3900X results are so close? Some benchmarks: mbw 4096 32 | grep AVG Ryzen 1700 (2x single rank 8GB): 6191 MB/s, 12428 MB/s, 8616 MB/s Ryzen 2700 (4x single rank 4GB): 6710 MB/s, 12418 MB/s, 9618 MB/s Ryzen 3700X (2x dual rank 8GB): 7597 MB/s, 13632 MB/s, 10714 MB/s Ryzen 3900X (2x dual rank, 2x single rank 8GB): 7125 MB/s, 13972 MB/s, 10057 MB/s Although there isn't a big difference, I'll switch around memory so all my systems have 4 ranks total. |
||
|
|
M-spec
Cruncher The Netherlands Joined: Jul 29, 2007 Post Count: 4 Status: Offline Project Badges:
|
Besides the difference in cache size, did you check the difference in memory performance on those machines? ... It runs with four ranks of DDR4 3733CL14. 4x single rank or 2x dual rank(2x16GB) for optimal perfomance with dual channel. Interesting! I run all my machines on the default 2133 setting, as RAM OC made some projects unstable when I tested it on the 1700. Is it possible that faster RAM could mitigate the difference in cache size, that is why your 3800X and my 3900X results are so close? Some benchmarks: mbw 4096 32 | grep AVG Ryzen 1700 (2x single rank 8GB): 6191 MB/s, 12428 MB/s, 8616 MB/s Ryzen 2700 (4x single rank 4GB): 6710 MB/s, 12418 MB/s, 9618 MB/s Ryzen 3700X (2x dual rank 8GB): 7597 MB/s, 13632 MB/s, 10714 MB/s Ryzen 3900X (2x dual rank, 2x single rank 8GB): 7125 MB/s, 13972 MB/s, 10057 MB/s Although there isn't a big difference, I'll switch around memory so all my systems have 4 ranks total. All Zen CPU's depend heavily on Infinity fabric and memory speed. Preferred 1:1 ratio. For example a Fabric speed of 1600MHz and 1:1 Mem at 1600MHz(=3200MHz DDR). You are missing 10-20% performance at DDR2133. That is the difference between a CPU upgrade. Four ranks (4xSR or 2xDR) can in some cases gain an extra 5-10% compared to two Dimms single rank. 1000-serie Zen CPU's do not support dual or four rank that well on the memory controller at high speed. Sweetspot is at DDR3000/3200. However, all Zen2 3000 CPU's will support DDR4 3733MHz, Fclk/Memclk 1866/3733. 1900/3800 depending on the CPU if you are lucky. |
||
|
|
|