I really need a bit of advice before spending a bucket of my boss' money, and hope that someone will be willing to do me a big favor and take a look at the specs below.
I'm trying to get a reasonable workstation configuration for some analyst-data crunchers. Their preferred spec comes out at around $10K, but we can afford max $4K per machine. I had to make some major tradeoffs--we will be buying several of these machines, and they all have to get the same rig (otherwise they complain that so-and-so has a better machine).
They called me an "IDIOT" for the spec I've come up with: does anyone care to confirm their allegation or defend me?
Use case: mostly data analysis (R/SAS/Stata/Python), ESRI GIS, MySQL/Posgre plus some analytical app development (light front-ends, if any in C++, C#, Java). They analyze small to medium-size datasets (never TB-class, and rarely GB-class). No CAD or 3D rendering, light use of Adobe CS.
To their disappointment, I've come up with the following (we must by from HP or Dell, can't build):
Model: HP Z420
CPU: Xeon E5-1650 (6C@3.2GHz)
Graphics: FirePro V4900 (to support their 2-3 1900x1200 screens)
Memory: 16GB DDR3 (4x4GB for option of going up to 32GB using all 8 available slots)
HDD: 256GB SSD boot + 3 x 7.2KRPM SATA 1 TB in RAID 0
They really wanted: 64GB, 3x600GB 15k SAS arrays, an E5-2600 series processor and Quadro 4000 graphics.
*The SAS drives are expensive, and I can't find any performance justification for their use case.
*I can't see any benefit to a E5-2600 series processor in a single CPU system. I realize there are more upgrade possibilities if we go with a system that has two CPU slots (e.g. HP z820 or z620), and there would be a marginal benefit of going to 8 cores (but this might be offset by the lower clock speed in the affordable E5-2600 8c CPUs, and we are on a 3yr upgrade cycle anyway).
*They don't do any heavy graphics, so a high-end graphics card is out of the question unless I'm just not understanding.
So, am I an idiot? Have I made the right trade-offs, or do I really just not get it? Any suggestions?
It would seem to me that when using a computer for data crunching the main focus is on the cpu , just like in gaming the main focus is on the video card. So my take on the setup would be that an eigh core cpu would be a good choice and with hyperthreading you then have sixteen cores and that would make the HP-Z620 or 820 a good choice. I do think that you have to get past the point of thinking your spending your own money and with this being a company the end result is the most important senario and not whether or not your an idiot. Besides the company gets to write off the expense of equipment where someone building a gaming computer does not. You Boss most likely like all Booes wants to get the most results out of the people working for him and will provide the tools to do so. You would be taking a huge risk trying to go with something that is not as efficient and your Boss finding out that you could have gone with something better but you wanted to be different?
THANKS inzone, I appreciate it . I've already had to push the boss to let me get up to a little over $4K per system, and he won't go higher (I don't mind the name-calling, it was just banter anyway).
If we go with a z620 and a single E5-2650 (8c @ 2 GHz), the price goes up and they get about 1GHz less--I could get the money from another pot and get it done if the extra cores dominate clock speed for their use (which it sounds like you are saying). I just don't have any experience with the new Xeons to say, but the benchmarks (FWIW) don't seem to show any benefit. They really wanted something like a 2687W (8C @ 3.1 GHz), but that's crazy expensive.
The biggest gripe was with moving away from 15k RPM SAS drives. I guess they will need to choose either the faster drives or the additional cores, as the price jump is similar.
Question is what will you benefit more from? Lots of slow(er) cores or fewer faster cores? If you're in charge for the build you should investigate and provide this data to your boss if possible so that he can see that this makes sense.
Again, I don't know what software you're going to use exactly but I don't see how will a Quadro / FirePro card can assist with data crunching? I believe it's only for CAD applications and stuff like that.. If you don't know the answer then you should definitely check that out as getting a mid / low range Radeon 6xxx / 7xxx series will save some money that can be used for more processing power, and almost all of these cards can support up to 6 displays.
If you were assigned to come up with a build you should really investigate, as it's your responsibility if these machines perform below expectation and your reputation on the line.
The questions I would ask myself are:
- Cores vs. speed.
- Queue depth required. If it's high, you may actually need a large SSD. If not (and by your description I believe it's not), maybe Hard disks will be sufficient.
- Is this work sensitive to HDD speed? If so, will it require a boatload of storage area (then you might want to get HDDs and raid them, and maybe a small SSD) or will it fit into smaller storage (then consider a large(r) SSD). In case it's not sensitive to HDD speed, a small SSD might be sufficient.
- SAS vs. SATA . Maybe SAS drives are more reliable, operate faster (15k does seem like they're much faster),maybe they have certain features that normal SATA drive don't (look at Intel for example, their SB-E line consists of processors that are 6-8 cores and are almost identical to the Xeon line at a lower cost, but lack certain features such as ECC memory support to prevent servers from getting those).
- Graphics card required. Do you need a workstation graphics? If you do, (maybe for the Adobe CS or w/e), how important is it? Meaning, how much work will it do? Once a week? Once a day? All day long? If it's only for Adobe CS and it's rarely used, maybe you can get an entry-level card such as the FirePro V3900 (review - http://www.tomshardware.com/reviews/firepro-v3900-revie...). Then again, I believe that Adobe CS is currently optimized for Quadro, and it's up to you to find out.
There are a ton more questions you should ask yourself and a whole lot of data you should gather. My suggestion is that you do some homework (if you haven't done so already) and also present all the data and maybe some actual benchmarks to your boss so that he can understand the reason behind every decision you've made and as an added bonus he'll see that you're taking your job really seriously.
P.S. That's the kind of work I do when people ask me to make a build for some special tasks, so I believe you MUST do it as well to avoid making some really bad decisions.
I'm the money guy, not the computer guy (I have a CS degree, but that's no help here). I got a request for 9 systems that were, in total, more than twice what the group had in their computer budget. The usual procedure is to deny the request and let them resubmit, but it is the end of the fiscal year, and they snuck the order in right before new requests were locked out. If I deny it, they lose their computer budget (it goes into the general fund), and they won't have accumulated enough $ for a new order for at least 2 years.
So, instead of stamping "denied", I spent 3 days (plus now my entire weekend) pouring over benchmarks for hundreds of configuration permutations, bothering good people like you and interviewing the analysts on their work to get up-to-date so as to give them the best system possible.
I very sincerely appreciate the expert advice I've gotten here. From your responses, it looks like they really do need the components they asked for (z620+ with one of the faster 8 core E5-2600s and nVidia graphics). I hate to do it, but it sounds like the trade-offs necessary to get the systems within the budget will just be too harsh to give them a justifiable spec, and nobody will be happy. So, I'll have to just deny the request after all. :-(
All of the benchmarks I looked at showed this spec to be very competitive witth their request (it was the one that gave up the least in everything but 3D graphics), so I'm assuming the benchmarks don't really do a good job for real-world use.
Thanks again for your help! I appreciate the guru-level advice, even if it did confirm that they cannot afford what they really need.