- Jul 15, 2015
- 99
- 1
- 18,665
Hello all!
So, just a preface about why I'm even considering these crazy pricey co-processors. I have built several machines in the past, and my work entails the use of iterative calculations (pattern recognition). I found that increasing the number of processors vastly outstrips the performance of faster cores when running the number of jobs I do. Previously, code was written to utilize a single core per job (each step done in sequence vs in parallel), and that would take a month to complete (with limited parameter queries too). I reworked the code, enabled multi/hyper threading (yes, both for various tasks) and on a special machine I built with 40 cores (at slower speeds) it takes about two days for the same job. Thus, it became very clear that core number beat core speed easily. This code had been used for decades and the group logic (until me) was the core speed was more important (to be fair, my mentors all were raised on the awesome MHz beasts of old. They're legends of the tape reels).
Anyways, I understand very little about the Phi processors, but my work has accelerated from hyper/multithreading to the point where I'm doing stuff my clients thought was impossible for decades simply because I can try nearly every combination of parameters imaginable. I need more cores, plain and simple. I'm looking to both double my throughput and halve the time of the job, maybe more since the speedup seems to be exponential, not linear. And I noticed that if I could play with Phi processors I would be able to theoretically get over 100+ cores which makes me giddy to imagine what I could do with that time. In short: I already have proof that more cores would definitely be an improvement and not wasted.
Because of this, my client and boss are willing to entertain the idea of getting something silly. The idea is that I would build it by hand since we have a machine that is a exact copy of one I built and it cost twice as much since it was purchased. Thus, the reason I am interested in Phi processors since I figure I could get two of the 22 core Xeons then maybe 2 Phi processors. Again, I have no idea how those Phi processors actually "talk" with the Xeons, so that could be a waste where I only would need something like a i7 (or my FORTRAN scripts would only use the Xeons). It sounds like cooling is a new issue that would arise here based on my very limited understanding. I currently am handling 2 Xeon E5-2670 v2's just fine on 212 Evos, but more weight + heat I dunno how it would do just from a structural standpoint.
So I guess the three questions I have then are 1) is it a simple plug and play type installation where I drop in a Phi coprocessor and away I go 2) what would be some design considerations such as case and extra cooling I would need to consider (I'm used to building Tower machines of standard design, wildest thing is that 2 CPU rig I use now) and 3) if I am using FORTRAN 95 and an intel parallel compiler would I only need to recompile my code to get it to play nice with the Phi processor?
As always, thanks for any feedback. Toms is always where I go to get guidance on my next build and thus far myself, friends, and even co-workers have been thrilled with the results. You guys are the best.
So, just a preface about why I'm even considering these crazy pricey co-processors. I have built several machines in the past, and my work entails the use of iterative calculations (pattern recognition). I found that increasing the number of processors vastly outstrips the performance of faster cores when running the number of jobs I do. Previously, code was written to utilize a single core per job (each step done in sequence vs in parallel), and that would take a month to complete (with limited parameter queries too). I reworked the code, enabled multi/hyper threading (yes, both for various tasks) and on a special machine I built with 40 cores (at slower speeds) it takes about two days for the same job. Thus, it became very clear that core number beat core speed easily. This code had been used for decades and the group logic (until me) was the core speed was more important (to be fair, my mentors all were raised on the awesome MHz beasts of old. They're legends of the tape reels).
Anyways, I understand very little about the Phi processors, but my work has accelerated from hyper/multithreading to the point where I'm doing stuff my clients thought was impossible for decades simply because I can try nearly every combination of parameters imaginable. I need more cores, plain and simple. I'm looking to both double my throughput and halve the time of the job, maybe more since the speedup seems to be exponential, not linear. And I noticed that if I could play with Phi processors I would be able to theoretically get over 100+ cores which makes me giddy to imagine what I could do with that time. In short: I already have proof that more cores would definitely be an improvement and not wasted.
Because of this, my client and boss are willing to entertain the idea of getting something silly. The idea is that I would build it by hand since we have a machine that is a exact copy of one I built and it cost twice as much since it was purchased. Thus, the reason I am interested in Phi processors since I figure I could get two of the 22 core Xeons then maybe 2 Phi processors. Again, I have no idea how those Phi processors actually "talk" with the Xeons, so that could be a waste where I only would need something like a i7 (or my FORTRAN scripts would only use the Xeons). It sounds like cooling is a new issue that would arise here based on my very limited understanding. I currently am handling 2 Xeon E5-2670 v2's just fine on 212 Evos, but more weight + heat I dunno how it would do just from a structural standpoint.
So I guess the three questions I have then are 1) is it a simple plug and play type installation where I drop in a Phi coprocessor and away I go 2) what would be some design considerations such as case and extra cooling I would need to consider (I'm used to building Tower machines of standard design, wildest thing is that 2 CPU rig I use now) and 3) if I am using FORTRAN 95 and an intel parallel compiler would I only need to recompile my code to get it to play nice with the Phi processor?
As always, thanks for any feedback. Toms is always where I go to get guidance on my next build and thus far myself, friends, and even co-workers have been thrilled with the results. You guys are the best.
Facebook
Twitter
Instagram