Ryzen Threadripper Gaussian16 Benchmark

Share This:

In our last post, we showed how to build a PC with Ryzen Threadripper.

Here we gonna show the gaussian16’s benchmark. In this post, we used only one molecule for benchmarking, but will make more accurate and diverse benchmark in the near future.

Based on our calculation, Threadripper is good for gaussian calculation.
First of all, the cost performance is amazingly great. It costs less than $1000 and has 16 cores! And Threadripper is also compatible for ECC and overclocking. There is no reason to pay a lot of money for Intel Xeon CPUs.

日本語版Threadripper の Gaussian16 ベンチマーク


Setting

We used the following PC parts.

CPU: AMD THreadripper 1950x, 16 core, 3.4 GHz
m/o: MSI Gaming AMD Ryzen ThreadRipper Extended-ATX Motherboard (X399)
SSD: SanDisk SSD PLUS 240GB Solid State Drive (SDSSDA-240G-G26)
RAM: Corsair Vengeance LPX 16GB (2x8GB) DDR4 DRAM 3200MHz (CMK16GX4M2B3200C16)
GPU: Radeon x550 2GB
Cooling: ARCTIC Liquid Freezer 240

We have installed Linux Fedora26 as OS, and installed Gaussian16 AVX2.
Memory is overclocked to 3,200MHz, and SMT (Intel calls this Hyper Threadding) is disabled.

For the comparison, we used the Desktop PC with Intel core i7-7700, 64 GB RAM and Gaussian16 (AVX2), and our PC cluster with Intel Xeon E5-2667 v2 (8cpu-64core), Gaussian09 E.01.

Calculation Set Up

We choose a natural product,vomilenine for our benchmarking test.
I draw the 3D structure on Gauss View and carried out Optimization and Frequency calculation. This compound has 47 atoms and made of C,H,N, and O, which is average size molecule for the QM calculation.

We used the following input file.

For Intel core i7-7700, calculations were performed with 4 cores, calculations were performed.
For Intel Xeon E 5-2667 v 2, calculations were performed with 4 or 8 cores.
For AMD Ryzen Threadripper 1950X, calculations were performed with 4, 8 or 16 cores.

Memory allocated 1 GB for 1 core. That is 4 GB for 4 core, 16 GB for 16 core (Actually, this amount of memory allocation is not necessary…).

As we don’t install gaussian16 on Intel Xeon E5-2667 v2, we run gaussian09 with the keywords “int=grid=ultrafine” and “scf=tight”.

Result

The result of benchmarking is as follows.

As Xeon E5-2667 v2 uses g09, which doesn’t show elapsed time at the end of log file, therefore we divided cpu time by the number of nprocshared.

Next, we made a graph with unit of time in seconds. First of all, let’s compare the performance difference among three kinds of CPUs.

For the 4 core job, the best performanced CPU was core i7-7700, and the second best CPU was Threadripper 1950X. These results are consistent with the PassMark Software’s single thread benchmark.

It was said that Ryzen is very slow for the scientific calculation, but it seems that there is no significant difference between Intel’s CPUs and AMD’s. Someone might think that the speed difference between Xeon E5-2667 v 2 and Threadripper is from the difference between g 16 and g 09, but the we confirmed that it does not effect much. In addition, as described later, Threadripper does not improve its calculation performance by AVX2.

Low Parallelization Eficiency

[2017.09.15]
After posting this article, we succeeded in improving parallelization efficiency.
See alsoThreadripper Parallel Efficiency improved ! Gaussian16 Benchmark

In the above chart I realized that the calculation speed of Threadipper and Xeon is reversed when using 8 core. Therefore, I examined the parallelization efficiency of Threadripper and Xeon E5-2667 v2.

Sorted by CPU time.

For Threadripper, even if the number of cores is doubled from 4 cores to 8 cores, the calculation speed became only about 1.6 times. Also, when doubling the number of cores from 8 cores to 16 cores, the calculation speed became only 1.3 – 1.4 times.

It was found that the parallelization efficiency was very bad when using all 16 cores. So we think that it is better to operate with 1 job with 4 cores or 8 cores. However, I think there is still room for improvement </ strong>. If there is a good way to increase parallelization efficiency, please let us know in the comments section or by email etc.

However, although the parallelization efficiency is low, when 16 cores are used, since it ends in a shorter time than when 8 cores are used. So if you are in the urgent situation like
“The revise of the thesis is close! I bet everything to this one job!”
Use 16 cores!

No Acceleration by AVX2?

It has been reported that Ryzen series CPUs are compatible for AVX2 but not accelerated. Here we tested SSE4, AVX and AVX2. In terms of calculation speed, there are no difference between them. Moreover, it seems. Rather, it seems that SSE 4 seems to be slightly faster! though it may be an error level

Conclusion

Our benchmark result shows good agreement with PassMark Softwares’ single thread benchmark.

Before benchmarking threadripper, we were very nervous because we heard bad rumours that ryzen is not good for scientific calculation which is not true. But actually, for the use of gaussian, threadripper works well!

In addition to it, the cost performance is really great! You can build 16-core node less than $2,000.
When we bought Intel’s Xeon E5-2667 v2 that has only 8cores, it costed $2,600. However Threadripper costs only $999, has 16 cores, and has almost the same single thread score with Xeon E5-2667 v2!

Though Parallelization Eficiency is very low, it works for 4-core jobs. If you run gaussian on Threadripper, don’t submit the 16-core job, but submit four 4-core jobs.
Personally, I strongly recommend Threadripper for gaussian!!

Related Articles

  1. How many kcal/mol ?
  2. Why atomic units (a.u.) are used in QM calculation?
  3. Building a PC with Ryzen Threadripper 1950x for Scientific Computation

One comment

  1. Nice comparison statistics! Another cpu chip less expensive than the Xeon is a Core i7-6950X. With nprocshared=16, we found that vomilemine optimization with Gaussian09 required 22 min as recorded on the wall clock (6h 13m CPU time). See job 2255 on the University of Alaska computational chemistry site guest user account (https://chem2.uaf.edu/facilities/WebMO).

Leave a Reply