Yes, you can use a single stick of DDR5 for bioinformatics and data analysis

INTRO

DRAM prices skyrocketed 171% in 2025 [1], and this trend looks like it will continue into 2026 unless there is a crash in demand for hardware for GenAI applications. This leaves bioinformaticians and other data analysts in a pickle, as most applications we use require a lot of RAM.

To keep costs low, we might consider using a single stick (aka Dual In-line Memory Module: DIMM) of RAM for a new workstation build, which is something that has been tried with reasonable success for low budget gaming setups [2]. So in this post we will look at whether using a single stick of DDR5 DRAM will cause a dramatic reduction in computational throughput as compared to the normal two-stick setup. We will also examine whether stock memory configuration (4800MT/s) is any slower as compared to the tweaked settings (EXPO 6000MT/s with low latency and high bandwidth support).

SETUP

The tests I will use include:

  1. A synthetic CPU test using stress-ng
  2. Single end RNA-seq human (STAR)
  3. Single end RNA-seq human (kallisto)
  4. Paired end RNA-seq zebrafish (STAR)
  5. Paired end RNA-seq zebrafish (kallisto)

The datasets are:

  • Zebrafish SRR36480541 21,807,697 read pairs.
  • Human SRR6467479 44,859,175 single end reads.

The system we are testing includes:

The DRAM kit being tested today.

RESULTS

Stress-ng runs various CPU stress test operations, which are measured in bogo ops per second, with more ops per second being better. As the Ryzen CPU has 12 cores and 24 threads, these tests were repeated to utilise 4-24 parallel threads. The results showed little to no difference in CPU throughput when using just one stick (DIMM) of DRAM (Table 1). In fact, some tests showed a slight speed-up in throughput when using the one-DIMM configuration, and this trend appeared stronger and more consistent with the BIOS-tweaked settings which was a surprise to me. Moreover, more aggressive memory speed settings led to slower CPU throughput in both the one- and two-DIMM configurations (Figure 1).

Table 1. Stress-ng CPU throughput with different memory configurations. We tested various parallel threads from 4-24.

Figure 1. Tweaking memory settings slightly reduced CPU throughput.

These are synthetic benchmarks and only stress the CPU, while the other read-mapping tests are intended to stress out other aspects like memory and I/O throughput. In the kallisto human test, the slowdown due to using only one DIMM was ~3.2%, and ~2.1% with tweaked settings (Table 2). Interestingly, in this scenario there was a significant benefit in tweaking memory settings, with ~14.9% and ~16.3% speedup with tweaked settings as compared to the stock memory settings in dual and single DIMM configurations respectively (Figure 2). 

Table 2. Time elapsed in seconds for Kallisto to process the human dataset with 4-24 parallel threads.

Figure 2. Tweaking memory settings increases data processing speed for Kallisto with human data.

The STAR human test was interesting, because the STAR index is large, close to the 32GB capacity. With one 32GB DIMM only installed, the swap space was being used with 8 or more parallel threads, and the job was killed altogether with 16 parallel threads (Table 3). At 4 parallel threads, the slowdown with using just one DIMM was 7.8% at stock settings and 16.6% with tweaked settings, but I have doubts about the accuracy of that as the index was using nearly all the available RAM.

Table 3. Time elapsed in seconds for STAR to process the human dataset with 4-24 parallel threads.


Figure 3. STAR struggled to complete with 32 GB RAM installed.

And this is why I replicated these benchmarks with the zebrafish dataset. It has a smaller genome, so the STAR index can comfortably fit in the 32GB offered by one DIMM. The slowdown for using only one DIMM was 0.63% with stock settings and -0.48% for tweaked settings (Table 4; Figure 4). The speedup for using tweaked settings was also really small. Taken together, this result shows that memory configuration does not impact STAR throughput that much as long as there's sufficient capacity available. This is probably due to STAR being more CPU bound.

Table 4. Time elapsed in seconds for STAR to process the zebrafish dataset with 4-24 parallel threads.

Figure 4. Memory configuration did not strongly impact Kallisto with zebrafish data.


For curiosity and completeness, I repeated the kallisto benchmark with the zebrafish data. It showed a substantial slowdown when using the one-DIMM config, being 12-15% at 12 parallel cores (Table 5; Figure 5). The speedup due to memory settings was also significant, in the range of 12-15%. One thing I did notice, was that these benefits tend to vanish or even reverse when using all available CPU threads, so a good approach to optimising workloads is to maintain them at 50 to 70% CPU utilisation.

Table 5. Time elapsed in seconds for Kallisto to process the zebrafish dataset with 4-24 parallel threads.

Figure 5. Tweaking memory settings increases data processing speed for Kallisto with zebrafish data.

CONCLUSIONS

So the take home message is that memory configuration and BIOS settings don't make much of an impact with CPU constrained workloads like STAR and stress-ng, but when the data throughput is higher, like when fastq files are being read from the disk, the benefits of using a two-DIMM tweaked setup is noticeable.

These results also show that tweaking the memory settings is just as important as the number of DIMMs. The biggest variation caused by the one-DIMM configuration was ~15%, so if you can tolerate a minor slowdown like that for your workflows, I would say that using a single DIMM is totally okay given the ridiculous prices being asked for computer parts at the moment. This slowdown can be somewhat mitigated by using the tweaked memory settings.

The importance of memory speed brings into question pre-built companies like DELL, which are still selling their overpriced workstation systems with relatively slow DDR5 at 4400MT/s [3], which are just not competitive with other offerings on the market. 

The other take-away from this is that running STAR for human data is just not reliable on a system with 32GB RAM, especially if you intend on processing the alignments afterwards. It may work at lower parallel thread counts, but that will slow down the job substantially. Consider alternatives like kallisto that use less RAM, unless you really need alignments to intergenic regions.

REFERENCES

1. https://www.tomshardware.com/pc-components/dram/dram-prices-surge-171-percent-year-over-year-ai-demand-drives-a-higher-yoy-price-increase-than-gold

2. https://www.youtube.com/watch?v=quHvWdmeQeQ

3. https://www.dell.com/en-au/shop/cty/pdp/spd/precision-t3680-workstation

Notes: No AI was used to prepare this post.

Location: Werribee VIC 3030, Australia

Popular posts from this blog

Data analysis step 8: Pathway analysis with GSEA

Image
In our RNA-seq series so far we've performed differential analysis and generated some pretty graphs, showing thousands of differentially expressed genes after azacitidine treatment. In order to understand the biology underlying the differential gene expression profile, we need to perform pathway analysis. We use Gene Set Enrichment Analysis ( GSEA ) because it can detect pathway changes more sensitively and robustly than some methods. A 2013 paper compared a bunch of gene set analyses software with microarrays and is worth a look. Generate a rank file The rank file is a list of detected genes and a rank metric score. At the top of the list are genes with the "strongest" up-regulation, at the bottom of the list are the genes with the "strongest" down-regulation and the genes not changing are in the middle. The metric score I like to use is the sign of the fold change multiplied by the inverse of the p-value, although there may be better methods out there...
Read more

Uploading data to GEO - which method is faster?

If you have had to upload omics data to GEO before, you'll know it's a bit of a hassle and takes a long time. There are a few methods suggested by the GEO team if you are using the Unix command line: Using 'ncftp' ncftp set passive on set so-bufsize 33554432 open  ftp://geoftp:yourpasscode@ftp-private.ncbi.nlm.nih.gov cd uploads/your @mail.com_ yourfolder put -R Folder_with_submission_files Using 'lftp' lftp  ftp://geoftp:yourpasscode@ftp-private.ncbi.nlm.nih.gov cd uploads/ your @mail.com _ yourfolder mirror -R Folder_with_submission_files Using 'sftp' (expect slower transfer speeds since this method encrypts on-the-fly) sftp  geoftp @s ftp-private.ncbi.nlm.nih.gov password: yourpasscode cd uploads/ your @mail.com _ yourfolder mkdir new_geo_submission cd new_geo_submission put file_name Using 'ncftpput' (transfers from the command-line without entering an interactive shell) Usage example: ncftpput -F -R -z -u  geoftp  -p "yourpasscode...
Read more

Generate an RNA-seq count matrix with featureCounts

Featurecounts is the fastest read summarization tool currently out there and has some great features which make it superior to HTSeq or Bedtools multicov. FeatureCounts takes GTF files as an annotation. This can be downloaded from the Ensembl FTP site . Make sure that the GTF version matches the genome that you aligned to. FeatureCounts it also smart enough to recognise and correctly process SAM and BAM alignment files. Here is a script to generate a gene-wise matrix from all BAM files in a directory. #!/bin/bash #Generate RNA-seq matrix #Set parameters GTF=/path/to/Mus_musculus.GRCm38.78.gtf EXPTNAME=mouse_rna CPUS=8 MAPQ=10 GENEMX=${EXPTNAME }_genes.mx #Make the gene-wise matrix featureCounts -Q $MAPQ -T $CPUS -a $GTF -o /dev/stdout *bam \ | cut -f1,7- | sed 1d > $GENEMX The data are now ready to analyse with your favourite statistical package (DESeq, EdgeR, Voom/Limma, etc). Consider attaching the gene name to give the data more relevance. To do that, first ...
Read more