Compute In The HiSeq X
My new old sequencer turned up today! It’s a HiSeq X! Believe it or not the HiSeq X is still supported by Illumina and will be until March of next year. But I doubt many users are still using them. As such, they’ve been turning up pretty cheap on eBay… seemed like a good chance to grab one and see what’s going on inside.
I’ve already pulled it apart and paid subscribers can access an image dump after the break below. Over the next few weeks I’ll be looking at individual components in more detail (so please subscribe to get those updates).
Today I’m going to talk a little bit about the HiSeq X device computer, contrasting this to the MiSeq compute. To briefly review. The MiSeq uses a regular industrial PC motherboard with 16 Gb of RAM and a 2.9GHz i7-7700T 4-Core Intel CPU. A single 1Tb SSD is used for software, storing acquired images and basecalled datasets.
In contrast to the MiSeq the HiSeq X uses an external compute server:
This comes direct from Intel (both the motherboard (S2600CP), CPUs and chassis itself). It has the following specs:
CPU: 2× Intel Xeon E5-2697V2 64 BIT 2.7 GHz CPU 30 MB Cache
Memory: 128 GB RAM
Hard Drive: 8 × 1.0 TB 3.5” HDDs, 5 × 400 GB SSDs
Inside the chassis we’ve also got a Firebird CL Cameralink card from Active Silicon and a RAID controller.
Storage
The MiSeq has something like 1% of the throughput of the HiSeq X1. Even so it has a relatively spacious 1Tb of storage to play with. The HiSeq X on the other hand has 3 volumes in various RAID configurations of: 5Tb, 2Tb and 1Tb.
Given the HiSeq X is ~100x faster this seems like a relatively modest increase in storage size! However, the MiSeq can keep full image sets around, and I suspect this is more challenging on the HiSeq X.
Given the throughput increases shown on the NovaSeq 6000 and NovaSeq X (see below). We’d need much larger storage arrays on these instruments. Getting into the 100Tb range if we increased storage to match the throughput increase.
This isn’t totally infeasible. But probably suggests that more image processing is done online and the images discarded as soon as is feasible (or not retained on disc at all).
CPU/Compute
In terms of CPU performance the HiSeq X has ~4x the compute performance of the MiSeq. This is in part because my MiSeq is from 2019, as such it has a significantly newer CPU than that used in the HiSeq X (first released in 2014).
A modern CPU like the AMD EPYC 9754 would probably provide about 3x the compute performance than that in the HiSeq X in a single CPU.
Notably, the GPU in this system is very basic, and there are no other accelerator cards.
Looking at the specs of the NovaSeq 6000 and the HiSeq X. The NovaSeq 6000 is about 7 times faster than the HiSeq X2. A couple of AMD EPYC 9754 seem like might be up to the job.
But what about the NovaSeq X? This seems to be about twice as fast as the 60003. Suggesting this would require about the CPU performance of 4 AMD EPYC 9754s. This is starting to get pricey (but nothing you can’t fit in a $1M budget!). And would also start to dissipate quite a bit of heat…
These and other issues may well make GPUs or other compute solutions more attractive here. Hopefully 10 years down the line we’ll find a NovaSeq X on eBay for $1000 and we can find out!
Until then… you can find more HiSeq X images linked after the break. And subscribe for more updates on the HiSeq X teardown.