What is AVX? AVX or Advanced Vector Extensions is a set of CPU instructions introduced by Intel in 2008 (with Sandy Bridge processors) that can accelerate performance for workloads and usages such as scientific simulations, financial analytics, deep learning etc. Since 2008, Intel has been steadily developing AVX and the latest iteration is called AVX-512, first introduced in 2016 for Xeon processors (Intel's enterprise line). In 2017, Intel also introduced AVX-512 for its high-end desktop line called Skylake-X. Since I've recently built a new BOINC machine based on Skylake-X CPU (i9-7900X), the question is obviously: can it make a noticeable difference in real-world BOINC performance?
To determine that, I have joined Asteroids@home BOINC project. Asteroids@home is a volunteer distributed computing project developed at the Astronomical Institute of Charles University in Prague, studying physical properties and parameters of asteroids in the Solar System. With the help of thousands of volunteers who donate their computing power to this project, Asteroids@home team recently published a new scientific paper, defining new shape models for 662 asteroids, significantly enlarging our knowledge of physical properties of those celestial objects (usually mentioned in the news when they pass "dangerously" close to Earth).
Asteroids@home calculates the shape of asteroids in the Solar System, based on the photometric data gathered by professional observatories, NASA satellites and even backyard astronomers (i.e. lots of data). Shown here is the artist's concept of asteroid named Oumuamua, known for its unusually elongated, cigar-shaped body.
Asteroids@home is also well-known for its numerous computing applications, supporting SSE2, SSE3, AVX and CUDA. Unfortunately, their CUDA app doesn't yet support Nvidia Volta GPUs which introduced CUDA with Compute Capability 7.0 and the latest version supported by Asteroids@home is 6.1, used by Nvidia Pascal GPUs (like GTX 1060, 1070, 1080 etc). That means I can't crunch Asteroids@home CUDA tasks (yet) with my Titan V (it's erroring out immediately upon start) so I had to limit this analysis to CPU tasks only.
Asteroids@home current Top Hosts list, meaning fastest computers, crunching most data. If your PC is there, then it's a powerful one, by all means. Of course, more data crunched -> more science done -> more Gridcoins earned (so, many would say it's not only for prestige).
If you check Asteroids@home Top Hosts list, you will find my new PC at the 3rd place (quite high, considering it's competing there with enterprise CPUs with far more cores). Almost all top machines are crunching AVX (and CUDA) tasks only, so any CPU without AVX support would certainly be at a disadvantage here. So, evaluating only by total output, one could say Skylake-X really does make a difference. But how does Skylake's AVX-512 compare with older AVX implementations, like AVX2 introduced in 2013? I would say difference is small to negligible. Although my average runtimes for Asteroids@home tasks are quite good (compared to the rest of top hosts), if you check machine ID 521233 (currently placed 11th on the top hosts list), you will see that this machine's average runtimes are almost equal to mine. And that's an AMD Ryzen processor, which doesn't have AVX-512 and is limited to older AVX2 instruction set. Also, browsing through Asteroids@home message boards, I came across this post which states that Asteroids@home needs only AVX2 for maximum performance. Therefore, to use full benefits of AVX-512, I guess that Asteroids@home AVX app has to be updated first (that will probably happen when AVX-512 gets more widespread).
Of course, extra performance provided by AVX is by no means 'free lunch'. CPU consumes more power and produces more heat when using AVX units, to the point it can become very difficult to cool it with air. More about that in my next post.
