Choosing a new PC for GEMPACK simulations

Updated March 2021

We are sometimes asked, what is the best sort of PC to buy for running GEMPACK simulations? Often the inquirer is running larger models and has experienced problems such as over-long solve times or shortage of memory (RAM). If your GEMPACK model solves fairly quickly (<2 minutes) and does not complain about shortage of memory, any new Windows PC will work fine [but we'd suggest at least 8 or 16GB of RAM]. If you intend to run larger GEMPACK models, the advice below may be helpful.

Executive summary

Choose a CPU with 4 or more cores from the Intel I7 or I9 range, or from AMD's Zen 3 or Threadripper range.
Choose 64-bit Windows, and insist on at least 8GB or 16GB of RAM (or maybe more).
A solid-state hard disk (SSD) sized 512GB or more will speed up simulations.

Advantages of a multi-core CPU

A good CPU has 4 or more processors (aka cores) in one chip. That allows you to run several simulations at once, or to speed up one simulation by using parallel processing (see below), if you have enough RAM. For value, choose the 2nd or 3rd fastest models that have plenty of on-chip L3 cache.

You can manually launch several simulations at once; for recursive dynamic models the RunDynam shell program will automate the processs. In this case, no one simulation is faster, but the total job is done sooner.

In contrast, GEMPACK 10 or later allows for parallel processing to speed up a single simulation: it splits up one simulation into separate processes. You can specify how many additional jobs (called 'servants') are created. In each case:

Two processors will do two tasks in about the same time as one processor would do one task. There is little point in launching more tasks than you have processors.
Each running simulation needs its own memory. So 3 simulations require 3 times the memory (plus the background requirement). If you are using 32-bit Windows, you might well hit the limit (around 3.5GB) that it imposes on total (all-processes) RAM use. 64-bit Windows is able to use all available memory.

Hence, to take advantage of your multi-core CPU, you need ample memory.

GEMPACK 12 exploits multiple cores even if you are only running one simulation without servants: it launches multiple threads during the BLAS part of the linear solve.

Get plenty of RAM

When memory (RAM) is short, Windows is able to use hard disk space instead (called "paging") -- at the cost of a severe speed reduction. If your GEMPACK program is starved of RAM it will run very slowly. Also, the operating system and background processes consume quite a bit of memory. Use Task Manager, as described here to investigate how much memory is required to meet background needs and to run several simulations at once.

64-bit Windows is better

You need 64-bit Windows to make use of more than 4GB of RAM. As explained above, 64-bit Windows (Win64) has 3 key advantages over 32-bit Windows (Win32):

Under Win64, total RAM use (by all running programs) is limited only by installed RAM (under Win32, overall limit is 3.5GB).
Win64 can allow a single 32-bit EXE to use 4GB of RAM (under Win32, limit is 2GB).
Win64 can run a 64-bit EXE which can use all available RAM (Win32 does not run 64-bit EXEs).

64-bit Windows 10 is now the normal OS for home and work environments. Older 32-bit programs still run fine under 64-bit Windows. To reap the advantages above, you should have at least 8GB of RAM.

Other

Consider also purchasing plenty of screen space -- either one very large LCD monitor, or a pair of smaller ones. This can be be a real productivity booster.

Hyper-threading

Many Intel CPUs feature "hyper-threading": the chip pretends to have twice as many cores as it really does. On a quad-core PC, Task Manager will show 8 (virtual) CPUs ! Unfortunately, there is still limited advantage in launching more than 4 simultaneous GEMPACK jobs (if you have just 4 cores).

High-performance desktop

Both AMD and Intel offer a range of workstation/server CPUs which are distinguished from their consumer grade CPUs. The workstation CPUs typically:

cost more, and use a different socket design, requiring a more expensive motherboard.
have slightly slower single-thread performance than the fastest consumer CPUs.
offer faster memory access, and support more RAM.
offer more cores.
tend to evolve more slowly.

Lenovo Thinkstation Hence for most CGE modellers a CPU from the top end of the consumer range will offer better value. A workstation/server CPU is only needed if you need very many cores or an enormous amount of RAM.

For best performance, a 512GB or larger SSD should be installed, large enough to hold the Windows OS, programs, and currently active projects. Inactive projects may be stored on a conventional HDD.

Bragging rights:
In Feb 2021 CoPS took delivery of 4 Lenovo Thinkstation workstations, each containing a 32-core AMD Threadripper Pro CPU, 256GB of octa-channel RAM, a 1TB SSD and a 4TB HDD. These powerful PCs can simultaneously run many large multi-period simulations.

What about notebooks?

As for desktops, you should look for 64-bit Windows 10 with at least 8 or 16 GB of RAM and an Intel I7/I9 or recent AMD CPU.

To extend battery life, notebook manufacturers choose less power-hungry components -- which also run slower. Less memory can be installed than on a desktop. Again, notebooks often have no separate video RAM (instead borrowing some of the main RAM). Hence notebooks usually run slower than comparably-priced desktop PCs -- or cost more for the same performance. However this gap is narrowing. An I7 notebook with a 500GB SSD will run nearly as fast as its desktop equivalent. See this comparison.

Brain-power or computer-power?

Some users could halve their model's solve time simply by exploiting GEMPACK more skilfully. For example they could:

Use a TABLO-generated EXE rather than GEMSIM to solve the model.
Switch to GEMPACK 12, which solves some models up to 10 times faster.
Experiment to find an optimal condensation.
Use the CMF file to supply an estimate of MMNZ (memory to allocate for the LU phase).

Read more about these ideas here. Other non-hardware ways to save time include:

Reducing the number of sectors or regions to what is really needed. As a rule of thumb, 10% fewer sectors would lead to 20% less RAM use and 30% less solve time.
Running less accurate (fewer steps) simulations in the preliminary stages of a project. Quite often, models which seem to require very many steps to solve accurately, are in fact suffering from a faulty closure, anomalous data, or poor theoretical specification.
Running no more than one antivirus checker. It may also help to exclude your GEMPACK and project folders from real-time checking -- see here.

Thinking and computer-power are indeed potential substitutes -- although to solve really big models you will need plenty of both.

Watch this space !

Recent developments have included:

Competition in the CPU market. Previously AMD competed with Intel only in the middle and budget parts of the CPU spectrum. Now AMD dominates in the very-many-core segment and matches Intel in single-thread performance.
Widespread adoption of 64-bit Windows 10.
Bigger RAM modules. To install memory in a PC, you push a DIMM (memory module) into a slot. Notebooks often have just 2 slots; ordinary desktop motherboards have 4 or 6 slots; while some server boards have 8 or 12. Standard DIMMs contain 8 or 16 GB RAM, meaning that many notebooks are limited to 32GB total RAM (and desktops to 64GB). Recently, 32GB DIMMs have become available, doubling these limits.
Solid-state hard drives (SSDs) promise a large performance increase but were previously rather expensive. This is changing -- cheaper SSDs are now common, especially in laptops or tablets, and are much faster than conventional HDDs. Your SSD should be large enough to store currently-used data and work files, as well as Windows and your software -- this means 512GB or more.
Large (30" or more) monitors have become more affordable.

Go back to
GEMPACK FAQ Page