My guess is that it would be really slow for the majority of tasks, but be extraordinary at very threaded tasks. Pricing will be the primary consideration, I think. I'm not aware of any workstation applications that benefit from having more than 8-10 cores, so you might be better off with a dual Xeon setup or something so you can get high single threaded performance and probably faster multithreaded performance. (unless your applications use 50 threads...)
I'm sure there are applications/scenarios where this excels, I just can't think of any. Can't wait for benchmarks!
Well, most demanding tasks in HPC actually scale quite well, almost perfectly. Those usually divide the work and do it in multiple passes. However I am also curious about async massive multithreading scenarios, where scores of thread do their own thing and exchange results rather than working on parts of the same workload.
Don't expect screen melting geekbench scores. To take advantage of Phi codes need to be optimized considerably. Vectorization, data locality, memory and cache optimization. This is a true parallel processor that with the right code modernization, will melt your screen..
Yes, I'm eager to see how a high thread count, low IPC platform does across an array of tasks. Even though they're Atom based cores, I'm eager to see if performance is 'good enough' for the per thread aspect while exploiting the massive amount of parallelism for high throughput.
"atom based" doesn't mean much in terms of actual performance, at least the kind of performance I am interested in, actually comes from the SIMD units, and this thing has 512bit wide simd while atoms don't even have avx - that is they cap out at sse3, which suggests the simd units are 128bit wide.
So while this may have low IPC for scalar, this is by far not the case for vector execution, which matters for practically 100% of the computation intensive tasks.
A review of that workstation would be nice, at least we can see how Intel's Knights Landing actually performs (including workstation and server style workloads).
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
8 Comments
Back to Article
ddriver - Tuesday, August 23, 2016 - link
Superservers and superworkstations sound oddly childishly named.Curious how the phi would do in traditional workstation (not enterprise) tasks as a main processor.
nathanddrews - Tuesday, August 23, 2016 - link
My guess is that it would be really slow for the majority of tasks, but be extraordinary at very threaded tasks. Pricing will be the primary consideration, I think. I'm not aware of any workstation applications that benefit from having more than 8-10 cores, so you might be better off with a dual Xeon setup or something so you can get high single threaded performance and probably faster multithreaded performance. (unless your applications use 50 threads...)I'm sure there are applications/scenarios where this excels, I just can't think of any. Can't wait for benchmarks!
ddriver - Tuesday, August 23, 2016 - link
Well, most demanding tasks in HPC actually scale quite well, almost perfectly. Those usually divide the work and do it in multiple passes. However I am also curious about async massive multithreading scenarios, where scores of thread do their own thing and exchange results rather than working on parts of the same workload.mtroute - Thursday, September 8, 2016 - link
Don't expect screen melting geekbench scores. To take advantage of Phi codes need to be optimized considerably. Vectorization, data locality, memory and cache optimization. This is a true parallel processor that with the right code modernization, will melt your screen..Kevin G - Tuesday, August 23, 2016 - link
Yes, I'm eager to see how a high thread count, low IPC platform does across an array of tasks. Even though they're Atom based cores, I'm eager to see if performance is 'good enough' for the per thread aspect while exploiting the massive amount of parallelism for high throughput.ddriver - Tuesday, August 23, 2016 - link
"atom based" doesn't mean much in terms of actual performance, at least the kind of performance I am interested in, actually comes from the SIMD units, and this thing has 512bit wide simd while atoms don't even have avx - that is they cap out at sse3, which suggests the simd units are 128bit wide.So while this may have low IPC for scalar, this is by far not the case for vector execution, which matters for practically 100% of the computation intensive tasks.
beginner99 - Wednesday, August 24, 2016 - link
I would like to have that one and also know some scenarios were it would be great. Now I will have to wait 10 years till software catches up...BlueBlazer - Wednesday, August 24, 2016 - link
A review of that workstation would be nice, at least we can see how Intel's Knights Landing actually performs (including workstation and server style workloads).