This is the last part of the series, and if you have not read the other two parts(pt. 1,pt. 2), it might be helpful to glance back to see how I got here.

In the last part, I explained the thinking that led me to build my own workstation using reclaimed parts. Here is brief overview of what I did:

Note: this is not intended to be a build guide. If you have no experience assembling a computer system, please find a beginner’s guide for the basics.

Components

Before you run off to EBay to source components, it is worth taking a moment to evaluate your expected workload, specifically to decide if it will be CPU, memory, or I/O intensive. In an (overly) simple breakdown:

• CPU: These operations benefit from faster calculations, and (to over simplify) benefit from higher clock speeds on the CPU. You want to look for faster GHz ratings on the CPU.

• Memory: These operations involve moving or altering lots of data, and benefit from getting that data to and from the CPU efficiently. You want to look for larger cache (L3) and more memory bandwidth, more memory channels, etc.

• I/O: These operations involve reading, writing, or moving data around on the disk. You want to look for a system that can mount several drives in RAID or over faster interfaces (e.g., PCIE).

Again, there are tons of details in each of these, but taking a basic inventory can help you focus your parts search.

In my case, modeling with Stan is heavily influenced by memory bandwidth, but the models themselves do not typically use much memory total. Therefore, I looked for CPUs with large L3 cache, quad-channel memory, but was not terribly concerned about CPU clock speeds, core count, or storage.

My build

I focused on a dual socket build, using LGA 2011 Xeon parts. Right now, you can pick up some older CPUs off EBay for around $30 each, each with 25 MB of L3 cache. Used ECC RAM costs around $1/GB, so for around $100 you can easily get 32 cores, 32 GB RAM. Given the cost, I figure if one part dies, it can be replaced for a low sum.

However, the difficulty is in the rest of the build. Server parts are intended to be installed in servers, not workstations. Therefore, finding a case, motherboard, PSU ended up being much more of a pain. I found a dual socket LGA 2011 Intel workstation board for $200 refurbished, a case that could fit the EAXT (extended ATX) motherboard for $80, and a power supply with the necessary connections to drive 2 CPUs (Seasonic 650 Gold unit) for an additional $90. You read that right, almost all the other components cost as much as the CPU and RAM, which are typically what people focus on. Additional parts that were needed were CPU coolers and a SSD for the OS.

The total cost ended around $600.

Was it worth it?

For this comparison, I have my daily-driver laptop (Lenovo T460s, i5-6200U, 20GB RAM) which I purchased for ~$1k vs. the Chromebook (Lenovo 11e) and workstation combo which cost $770.

Computing power

Both the T460s and workstation are running Linux, hence UnixBench is the nature benchmark to compare them.

Single core

First, the laptop:

Not terrible. How did the $600 workstation with old parts do?

So, the laptop actually has better single core performance. This is not terribly surprising, since the laptop CPU is newer and has a higher clock.

Multicore

Again, first the laptop:

And now the workstation:

This is really where the workstation shines - notice that the benchmark is only running on 8 cores concurrently.

Benchmarks vs. real-world use

The workstation crunches on large Stan models without even breaking a sweat, even across 16 cores. Due to the memory bandwidth, even simple models are faster on the workstation compared to the laptop. I want to try some other computations at it (crop models, etc.), and I want to throw some GPUs in it for Machine Learning (it has 4 PCIE slots!). Until then, I can confidentially say that it out-performs any other $600 system I have seen.

One (important) thing to consider - I am not accounting for energy use here. The electricity needed to run the workstation and cool the room it is in should be taken into consideration.

So, what is the verdict?

The Chromebook is surprisingly nice for day-to-day cruising the web, answering emails, etc. When it comes to writing code, it is OK once set up (though the setup is not always fun). It is not great for actually running analyses, but the workstation helps patch up that deficiency.

I think if I only had ~$800 to spend, I would lean towards the Chromebook + workstation set up. But the limitations of the Chromebook when it comes to actually compiling code and/or running an analysis can be frustrating. While things might change, Chromebooks just are not intended for development.

Therefore, given everything I have learned over the last 2 years: If I were starting all over today, I would pick up a used Thinkpad (a X230 weighs approximately the same as the Chromebook, but with much more storage, processing power, etc.) and build a workstation for more intensive work.