Xeon E3, E5, or i7 for long duration statistical algorithms

[bookman3]

Hi all,

Been around since the 90s under various usernames, but tend to only post when I'm in the market for something new and lurk when I'm not.

I'm about to start a new job and was wondering if you could help me put together a top notch build.

I'm just starting my career as a statistician in biomed. Currently I'm doing research in endocrinology that requires long running MCMC simulation studies. Most of my day-to-day work for MDs will be fairly standard analyses (mixed models and the like), but I will also be continuing my research into biomedical applications of spatial bayesian and machine learning methods.

I have 3k to spend including monitors. So far I've looked at 4-core, high clock speed cpus -- Xeon E3-1241 v3, E3-1271 v3, E5-1630, and the hard to come by i7-6700. Tied up in the cpu choice is ecc rdimm vs udimm, ddr4 vs ddr3, a single m.2 ssd vs multiple smaller ssds on a raid.

GPU just needs to power two 1440p monitors over DP (though I'm tempted by UHD).

I'll be running linux mint or debian as my primary os and Windows on a virtual machine when needed.

Any and all suggestions would be much appreciated.

Thanks!



[Oh, and I'm currently waiting for an answer as to how much of a hassle it will be to build a pc versus go through their contracted supplier (Dell...). If it's a big hassle, I'm probably looking at a precision t5810 or t1700, but let's assume I'll build it.]

 

[i7Baby]

Budget?

See http://www.tomshardware.com/forum/261222-31-build-advice - to see if you left anything else out.

For number crunching, more CPU cores is better.

 

[NTMBK]

If it's for your job, go with Dell/HP/one of the other big name workstation vendors. If something goes wrong with your workstation, you will have a Dell engineer on site the next day. We've had full motherboard replacements next day. If your self-built workstation develops a hardware fault, you're in for days of debugging, followed by days of waiting for parts to get shipped to you (unless you stock replacements for everything, doubling the cost of your build).

 

[Geforce man]

If it's for your job, go with Dell/HP/one of the other big name workstation vendors. If something goes wrong with your workstation, you will have a Dell engineer on site the next day. We've had full motherboard replacements next day. If your self-built workstation develops a hardware fault, you're in for days of debugging, followed by days of waiting for parts to get shipped to you (unless you stock replacements for everything, doubling the cost of your build).

100% agreed. With the coupons and deals dell has regularly, it doesn't make sense to go for self built for a daily driver machine that is necessary for you to make $. Go with a high end workstation. 2500$ for the machine, 500$ for monitors should be decent I would think. around 2700$ (if you stretch a bit) nets you a dual xenon machine with 16gb ECC ram, 845watt PSU, and like mentioned above, peace of mind from warranty. (Dell T7810 workstation) Find yourself some coups and you could have the machine for less. Or you can configure it with non dual processors for less also probably.

Wow, just realized you could do a dual 8cHT chip for just over 3 grand... thats 16 cores and 32 threads, up to 3.2Ghz, for sub 3100$. Thats a metric **** ton of processing power for the money.

 

[bookman3]

NTMBK & Geforce-- good point. I have a tough time swallowing $500 for an ssd and even more for 32gb of ram. Is it worth getting either of these separately? Likely through a different it contract (cdwg)? And where would I find coupons, just search the net?

 

[bookman3]

I7Baby -- here you go

Approximate Purchase Date: ordering asap

Budget Range: $3000 max, colleagues have tended not to use full budget

System Usage from Most to Least Important: 1) standard statistical analyses in R, some SAS
2) MCMCs written in C, likely new work in C++
3) machine learning in R, Python to start

Parts Not Required: mouse, keyboard - bringing own das keyboard and logitech mouse

Preferred Website(s) for Parts: Dell, CDW/CDW-G

Country: USA

Parts Preferences:
-Intel xeon e3 or e5, or i7 - high clock speed, care about physical core count, hyper threading doesn't do my code much good.
-32GB DDR4
-512GB M.2 NVME SSD
-Cheapest GPU that will power the monitors


Overclocking: No

SLI or Crossfire: No

Monitor Resolution: 2x Dell U2515h 2560x1440 or 1x P2715q 3840x2160 -- a ton of screen real estate is a high priority

Additional Comments:
-biggest question is which of the wide range of Xeons should I be looking at and is ecc/registered ram imp
-In my daily work I push 1-15gb datasets between ram and disk, so respectable SSD is a priority

 

[zir_blazer]

Desktop Core i5/i7s do NOT support ECC. Xeons E3 do NOT support Buffered (Registered) RAM, they work with either Unbuffered (Standard, the most common one) or Unbuffered+ECC, which is the intended one. Xeons E5 work with all types, but best used with Buffered+ECC.
You also need a proper Motherboard to use ECC. Look for a Workstation or Server class Motherboard with a C Series Chipset. If you're serious, go Supermicro.

Desktop Core i7/Xeons E3 have an IGP that can drive 3 Monitors, you may not need a discrete Video Card at all. You may, however, need adapters (DVI-D to DisplayPort or whatever), since Motherboards usually bring one of each connector type instead of 2 or 3 of the same.
Xeons E3 that don't finish in 5 or 6 do NOT have IGP. The models you said do not have IGP. Xeon E3-1241V3 will not work without a discrete Video Card (Or a Motherboard with IPMI support, since they include a very basic GPU), E3-1246V3 does. You also need a C226 Chipset to use it (C222 and C224 don't support IGP. All Desktop Chipsets do).

Current Xeons E3 V3 are based on Haswell. Supposedly in one month the Skylake based Xeons E3 V5 will officially launch. You may want to wait for them.

Use a proper PCIe SSD like the Intel 750. I don't take M.2 serious for Desktop usage.

 

[Ken g6]

-biggest question is which of the wide range of Xeons should I be looking at...

How well does your code scale to multiple cores? I have some code that scales perfectly, so I can just multiply cores by GHz. I have some other code that doesn't, which makes GHz more important and cores less-so. And don't forget the case where you might be memory bandwidth-limited.

 

[Enigmoid]

How well does your code scale to multiple cores? I have some code that scales perfectly, so I can just multiply cores by GHz. I have some other code that doesn't, which makes GHz more important and cores less-so. And don't forget the case where you might be memory bandwidth-limited.

This. A multicore monster is useless if you code runs on a single core.

 

[Phynaz]

If it's for your job, go with Dell/HP/one of the other big name workstation vendors. If something goes wrong with your workstation, you will have a Dell engineer on site the next day. We've had full motherboard replacements next day. If your self-built workstation develops a hardware fault, you're in for days of debugging, followed by days of waiting for parts to get shipped to you (unless you stock replacements for everything, doubling the cost of your build).

This. Running a home built computer for work can cost you your job if something goes wrong. Not worth the risk.

 

[bookman3]

The simulation studies scale perfectly since its n independent sequential executions of the code. This will be a 1/5 or less of my daily work when I start. I hope it will become a larger proportion over the first year or two.

My daily work in R can be parallelized somewhat (e.g. sets of regression fits, bootstrapping, or cross-validation), but a lot of it (data cleaning and latex compilation) is naturally sequential.

Machine learning is a new area for me, so you know as well as I do.

If I were to stick with my initial intuition of 4 or 6 high clock rate (3.4ghz+) cores which e3 or e5's should I look at? I'd prefer ddr4, which probably then requires an e5

 

[i7Baby]

A build your own under $3k -

CPU: Intel Xeon E5-2630 V3 2.4GHz 8-Core Processor ($629.99 @ SuperBiiz)
CPU Cooler: Corsair H100i GTX 70.7 CFM Liquid CPU Cooler ($109.99 @ Amazon)
Motherboard: ASRock X99 Extreme6 ATX LGA2011-3 Motherboard ($213.98 @ Newegg)
Memory: G.Skill Ripjaws Series 32GB (4 x 8GB) DDR4-2400 Memory ($199.99 @ Newegg)
Storage: Plextor M6e 512GB M.2-2280 Solid State Drive ($453.62 @ Amazon)
Storage: Seagate 4TB 3.5" 5900RPM Hybrid Internal Hard Drive ($171.99 @ Amazon)
Video Card: PowerColor Radeon R9 295X2 8GB Devil 13 Video Card ($653.98 @ Newegg)
Case: Corsair 750D ATX Full Tower Case ($138.98 @ Newegg)
Power Supply: SeaSonic X Series 1050W 80+ Gold Certified Fully-Modular ATX Power Supply ($161.98 @ Newegg)
Optical Drive: Asus BW-16D1HT Blu-Ray/DVD/CD Writer ($50.99 @ Newegg)
Operating System: Microsoft Windows 10 Pro OEM (64-bit) ($139.75 @ OutletPC)
Wireless Network Adapter: Asus PCE-N53 802.11a/b/g/n PCI-Express x1 Wi-Fi Adapter ($13.06 @ Newegg)
Total: $2938.30

 

[wolf_squad]

most machine learning stuff is typically done in CUDA, so, if you can figure out what algorithms and frameworks you're gonna use, you can check if they're CUDA-accelerated, and, if so, grab a TITAN or something of the sort. neural network-based ml stuff (such as deep learning, lstms etc) done on GPU via CUDA is orders of magnitude faster than on CPU, and CUDA is de facto standard for that right now.

 

[zir_blazer]

A build your own under $3k -

That's is a gaming build and nothing else.

Motherboard: You DO NOT USE a X99 enthusiast Motherboard for a serious Workstation, you use a proper C612 based one.
For Single Processor, you may want something like a Supermicro X10SRA or a X10SRA-F (Has IPMI for Remote Management).
For Dual Processor, Supermicro X10DAi or X10DAX. These are EATX size and you need a Case to fit them in.

Processor: Putting a Xeon E5 2xxx in a Single Processor Motherboard is possible, however, Xeons E5 1xxx usually are cheaper and have higher Frequency but less Cores, on 2xxx you may find some models where the Frequency*Core value makes sense, plus you can use two of them.

Memory: For Haswell-E you can use DDR4. Buffered+ECC, obviously, at least 8 modules so you can do Quad Channel in each Processor.

Storage: For SSDs, M.2 is a format intended for Mobile, and usually the controllers have very high operating temperatures since they're intended for burst usage and not sustained. Money being no issue, I would pick the Intel 750 PCIe any day of the week.
I would use two RAID 1 HDs from Hitachi, since BackBlaze can't kill them fast enough.

Video Card: ONLY if he is actually going to use CUDA or OpenCL, and doesn't have a Motherboard with IPMI. And chances are that he needs a FirePro or Quadro, not a gaming card. With no IPMI, you still need a cheap one either to drive a Monitor.

Optical Drive: What for? Doubt he is going to use that Xeon as a Home Theather and reproduces Blu-rays in it.

 

[tynopik]

The simulation studies scale perfectly since its n independent sequential executions of the code. This will be a 1/5 or less of my daily work when I start. I hope it will become a larger proportion over the first year or two.

My daily work in R can be parallelized somewhat (e.g. sets of regression fits, bootstrapping, or cross-validation), but a lot of it (data cleaning and latex compilation) is naturally sequential.

E5-1650v3

best balance between high clocks, moar cores and affordability

https://en.wikipedia.org/wiki/Haswell_(microarchitecture)#Server_processors


E5-1630v3 is also a reasonable choice with slightly higher clock and lower price

 

[bookman3]

So I think its settled. I haven't been a big Dell fan in the past, but I may be a convert. The support/warranty could be a much bigger deal than I had thought. With a little help, I got an estimate close enough to my budget to make it work.

Dell T5810
Xeon E5-1650 v3
32GB DDR4 ECC R
512GB SATA SSD
FirePro W4100
and two solid monitors

The E5-1650 seems like a great compromise and I'm holding off on the pci-e ssd. The AMD gpu was the only option in my price range, so I'll have to do a little more research to see if my usual tools can use OpenCL instead of CUDA. Overall I'm quite happy with the specs we put together.

Thanks everyone for the advice.

 

[bshor]

I'm putting together a workstation for MCMC computation via R as well and I'm wondering how well the build you put together has worked out for you. I'm looking at dual 2630v4 Xeons.

 

[Phynaz]

Dude, it's been nearly a year since the op has logged into this site. He's not going to answer.

 

[corkyg]

I'm putting together a workstation for MCMC computation via R as well and I'm wondering how well the build you put together has worked out for you. I'm looking at dual 2630v4 Xeons.

That post is too old to expect a response. You might try a PM to the poster, but chances are slim.