The HPE Apollo was installed on the ISS as part of this experiment: High Performance Commercial Off-The-Shelf (COTS) Computer System on the ISS (Spaceborne Computer) - 09.19.18
According to news reports, the computer will return to Earth on SpaceX CRS-17 in February or March of 2019 (e.g, this November 1, 2018 Fast Company report):
With three or four months before the system gets a ride back to Earth for more testing, NASA decided to put the system, an HPE Apollo 4000-series enterprise server, to work doing real science experiments on the ISS.
“We’ve been scheduled to return to Earth on SpaceX 17, which is in late February or early March,” says Mark Fernandez, lead developer for HPE’s Spaceborne Computer program. “Therefore we can open up the supercomputer on the ISS for advancing other types of space exploration.”
While this is considered a "high performance computing" (HPC) machine, it does not include GPUs. I sent several questions to HPE, including about possible GPU. Below is the reply that I received in August 2017, about the time the computer arrived on station:
HPE’s Spaceborne Computer does not contain GPUs. Spaceborne Computer uses un-modified Commercial-Off-The-Shelf (COTS) servers from the HPE
Apollo 40 family. These are 2-socket x86-based HPE Gen-9 compute nodes
featuring Haswell/Broadwell class processors, commonly used in HPC
Within Spaceborne Computer are two such compute nodes and a 56Gbps
high speed optical interconnect. An unmodified open-source operating
system (RedHat 6.8) is installed; and internationally recognized
open-source benchmarks are the workload for the duration of the 1-year
long experiment. Several software layers from HPE provide the
HARDening with SOFTware. Spaceborne Computer has a peak performance in
excess of 1TF.
So... no GPU. Standard RedHat Linux on high-end Intel processors; not rad-hard chips.
Basically, this was to see if a commercial HPC engine could be deployed in space. If so, a lot of experiments on the ISS might be able to benefit from it. Without this, heavy computation on the ISS would have to rely in ground-based computers, which would receive data over relatively low-speed links from orbit; in such cases, it often makes sense to do the computation after the experiment is over, and data is collected and shipped back to Earth by returning spacecraft.
In retrospect, there are a few reasons that I can think of for not including a GPU:
- Not having GPUs simplifies the configuration, making it less heterogeneous, and easier to repair or retain parts for. Putting systems in orbit is very expensive, and simplicity (or at least lack of unknowns) raises the chances of success.
- Use of a GPU requires special application programs using a software framework like CUDA or OpenCL. For smaller machines, this is generally useful if the computing system is designed for a particular application, like gaming or crypto-mining. These tend to be single-purpose machines.
- Making the system rather generic allows several applications to utilize the machine at the same time, possibly running simultaneously. In fact, multiple ISS users could connect to it over SSH to monitor their compute jobs. In general, it provides greater flexibility as experiments come and go.
So for the three or four months before it returns to Earth, the HP Apollo 40 can be used to support some ISS experiments. Then HPE will find out how machine fared in zero G, elevated radiation, etc. Beyond that, one can imagine an improved machine being sent later to the ISS. At that point, a proposal for an ISS experiment could factor in on-orbit compute time.