Are those supercomputers still a good idea compared to building dedicated hardware for each single algorithm that those computers will run ? When i look at bitcoin hashing, the trend seems to be to not use gpu anymore, but rather asics
Are they going to run such a wide variety of algorithm that they need general purpose gpu ?
Yes, eg for physics simulations (of which there exist a great variety that have been optimised for HPC architectures consisting of nodes with CPUs and GPUs, for example for CFD). Special purpose hardware also exists, eg for molecular dynamics ("Desmond") but that is not common across domains.
We recently had some real discussions about this. It’s worth remembering that designing, building and testing a custom chip runs, for a relatively old density, order 100 million dollars. Add to that the fact it would only be useful for a small subset of the problems your users want to use the machine for and it becomes pretty clear why supercomputer centers go generic as much as feasible. Also bitcoin is a bit of a poor example because the memory requirement for each unit is incredibly small. That’s rarely the case for any kind of meaningful simulation.
There’s a lot of variance between codes that run on a HPC system like this. You’d expect to see molecular dynamics, fluid dynamics, weather forecasting codes, density functional theory etc. etc., and the first two especially have a lot of variance in the codes and methods used to do those simulations.
I can’t remember the figures off of the top of my head but I read a report those use cases dominated ARCHER (UK national supercomputer) compute time a few years ago.
This is a supercomputer meant for academic and research use primarily. The computer thus must support a wide array of users, who'll be working on an even wider array of software, much of it very much not optimized for anything in particular. Now, the FZJ that hosts this machine has a department that aims to help researchers optimize (and professionalize, i.e. tests and CI and such) their software. We can help optimize your tool for our particular machine, or some other hardware, and we make some tools that help you use supercomputers (some of our users, perhaps many, are not trained software developers at all, let alone habitual HPC customers).
So, the hardware needs to be general purpose, because the software it's going to run are not known yet, and much of it will be barely executing ("research software" ;)).
> Now, the FZJ that hosts this machine has a department that aims to help researchers optimize (and professionalize, i.e. tests and CI and such) their software.
When I read articles celebrating “European HPC”, I always wonder about if systems like this are being an effective use of European taxpayer money, since nearly all of the cost goes to US chip manufacturers. What I understand is that Bull basically builds racks around those chips.
I’d imagine the Japanese Fugaku system which is based on ARM and built by Fujitsu a more effective way to directly support a nation’s economy and engineering capabilities.
I could be wrong, of course, nevertheless I find it puzzling.
Although there are tons of non-EU chips going into these machines, there are tons with EU involvement too (think networking, memory). Also, while chips are a significant share of the expense, it is certainly not most of it. Housing, cooling, infra, racks; those are all things that can be done 'locally'.
I couldn’t find any information about the interconnect of the new machine in the article, there is a comparison to LUMI, which uses Cray/HPE Slingshot and JEWEL, which is Mellanox/NVIDIA Infiniband. I couldn’t find anything concrete about JUPITER, maybe there is a chance it could use the Bull interconnect, should it be a Bull machine.
Agreed on the physical infrastructure and that those things will be done locally, I do not consider those things as a competitive advantage on the global market, compared to strengthening the local semi industry by doing an “Euro ARM” chip or some such.
No doubt that some of that money will benefit the European economy, with regards to the effectiveness and long term strategic view I do have my doubts, especially compared to what the US/Japan have been able to pull off.
The alternative is having to pay for HPC (Azure offers this) and not developing any competences around the planning, operation and maintenance of such an operation.
Europe has tons of supercomputers, this is just the latest crossing a particular size that sounds cool in outreach and promotion. The research done on these machines (particle physics is a major consumer, see CERN) means there is a very healthy and large community around such tooling, and work is being done to onboard new fields as well. This is a meaningful outcome of such computers, even if they don't contain local free range chips ;)
12 comments
[ 3.0 ms ] story [ 38.0 ms ] threadAre they going to run such a wide variety of algorithm that they need general purpose gpu ?
I can’t remember the figures off of the top of my head but I read a report those use cases dominated ARCHER (UK national supercomputer) compute time a few years ago.
So, the hardware needs to be general purpose, because the software it's going to run are not known yet, and much of it will be barely executing ("research software" ;)).
Sounds like a fun job
I’d imagine the Japanese Fugaku system which is based on ARM and built by Fujitsu a more effective way to directly support a nation’s economy and engineering capabilities.
I could be wrong, of course, nevertheless I find it puzzling.
Agreed on the physical infrastructure and that those things will be done locally, I do not consider those things as a competitive advantage on the global market, compared to strengthening the local semi industry by doing an “Euro ARM” chip or some such.
No doubt that some of that money will benefit the European economy, with regards to the effectiveness and long term strategic view I do have my doubts, especially compared to what the US/Japan have been able to pull off.
Europe has tons of supercomputers, this is just the latest crossing a particular size that sounds cool in outreach and promotion. The research done on these machines (particle physics is a major consumer, see CERN) means there is a very healthy and large community around such tooling, and work is being done to onboard new fields as well. This is a meaningful outcome of such computers, even if they don't contain local free range chips ;)