I wonder how a Xeon 8180 or a Xeon Phi 7290 would do. The 8180 is probably the fastest x86 and the Phi, with its 72 dumber cores and, IIRC, 144 threads is almost always surprising.
The Phi... definitely slower. The single-core performance is very low, and there are usually at least a few linking stages in any source code compilation that collapse the process down to a single thread - you can't beat Amdahl's Law here.
Why? the Phi uses customized Atom cores but a lot of them, it's not designed to execute normal x86 code particularly fast.
It's designed to be 100% code compatible, you can boot windows/linux on it and run your programs with the specialized SIMD instructions for actual throughput.
You get code compatibility, cache and branch management from x86 CPUs and near GPU speeds (or even faster for FP64) for SIMD instructions.
The cores are individually slow and non-parallel code will not be fast, but the Airmont architecture it's based on is an out-of-order core, so it's not that bad as earlier versions of the Phi, plus it has up to 16GB of very fast HMC memory near the cores, helping make up for the slow cores with less costly cache misses than an ordinary Xeon. It also has four threads per core without the penalties of earlier Phis. Single-thread stages will suffer but that may end up being compensated by ridiculously parallel compilation stages.
But yes. The Phi is designed to shine with code that uses its AVX512 units. Compiling the Linux kernel is a "create misuse" of its technology. One I'd like to try one day ;-)
I am working on some Threadripper Linux comparison benchmarks/article currently and will have it in the next day or two on Phoronix, this was just a quick finding shared on Twitter.
I have many Epyc/TR benchmarks coming up... Some PostgreSQL and SQLite though is about it for SQL. List of possible tests @ http://openbenchmarking.org/tests/pts
Could add other SQL tests if being pointed to a decent set of scripts for setting up given test and execution in an automated manner so it could be incorporated into the Phoronix Test Suite test profiles.
I also included PSKReporter in my dataset but the owner doesn't include public download options for that.
I've got this database set up on a PC on my bosses desk and no one at my University thinks new hardware is a priority. You get supplies by scrounging at colleges.
It seems like nowadays mainstream dev workstations are all intel-powered, which are most developers use in the companies.
I did a simple search but didn't find any ThreadRipper powered dev workstations. Is it because AMD doesn't care about the market or that is too hard to enter?
Only for OEM products and afaik its only till end of year not for a year after release. Thinkmate indicated they will have workstations out in a month over email.
Ryzen was similarly lagging in compilation benchmarks. Ian Cutress (AnandTech Senior Editor) said that Zen performs poorly with compiler benchmarks due to its L3 cache being a victim cache, and that "you can't prefetch into a victim cache" (though Google results suggest that IBM might have a patented technique?).
Semi related, but Threadripper is not just about speed. It is about PCIE lanes. Intel has been taking advantage of its monopoly for decades, artificially limiting PCIE lane count, among other things.
This may be the start of a new era in computing, allowing GPGPUs to be used more readily, not to mention specific neural network and deep learning applications. I'm personally making a high risk bet that this is the first in a series of step that will bring AMD and NVIDIA closer to wresting the monopoly from Intel.
24 comments
[ 2.1 ms ] story [ 49.0 ms ] threadBut also substantially more expensive
The Phi... definitely slower. The single-core performance is very low, and there are usually at least a few linking stages in any source code compilation that collapse the process down to a single thread - you can't beat Amdahl's Law here.
See? Always surprising.
It's designed to be 100% code compatible, you can boot windows/linux on it and run your programs with the specialized SIMD instructions for actual throughput.
You get code compatibility, cache and branch management from x86 CPUs and near GPU speeds (or even faster for FP64) for SIMD instructions.
But yes. The Phi is designed to shine with code that uses its AVX512 units. Compiling the Linux kernel is a "create misuse" of its technology. One I'd like to try one day ;-)
ThreadRipper/Epyc sounds like a great platform for mass-NVME database systems.
Could add other SQL tests if being pointed to a decent set of scripts for setting up given test and execution in an automated manner so it could be incorporated into the Phoronix Test Suite test profiles.
You might try mysqlslap for a basic test.
http://wsprnet.org/drupal/
And
http://www.reversebeacon.net/main.php
I also included PSKReporter in my dataset but the owner doesn't include public download options for that.
I've got this database set up on a PC on my bosses desk and no one at my University thinks new hardware is a priority. You get supplies by scrounging at colleges.
I did a simple search but didn't find any ThreadRipper powered dev workstations. Is it because AMD doesn't care about the market or that is too hard to enter?
Also mind power consumption.
This may be the start of a new era in computing, allowing GPGPUs to be used more readily, not to mention specific neural network and deep learning applications. I'm personally making a high risk bet that this is the first in a series of step that will bring AMD and NVIDIA closer to wresting the monopoly from Intel.