Top500 June 2018 Edition New Systems Analysis

[Patrick Kennedy]

We take a look at the new Top500 supercomputers list from June 2018 and dissect some of the trends we are seeing. There are some surprises such as the dominance of 10GbE as an interconnect which we analyze

The post Top500 June 2018 Edition New Systems Analysis appeared first on ServeTheHome.

Continue reading...

 

[mstone]

Nobody working in HPC considers more than the top 20, maybe the top 5-10, and that's been true for more than 20 years. Most of the list is just general purpose computers that someone ran linpack on, and that's always been true.

 

[Evan]

Nobody working in HPC considers more than the top 20, maybe the top 5-10, and that's been true for more than 20 years. Most of the list is just general purpose computers that someone ran linpack on, and that's always been true.

Too true, we have thousands of power8 cpu’s Even linked with fast network in addition to infiniband for some uses but it’s not running anything like a HPC workload, if we were to run the benchmark it would be well up the list.

And the list does not contain plenty of private company actual HPC systems that would be way up the list
Even if the vendor for example Fujitsu does a benchmark (and its required for commissioning and acceptance by us) we don’t allow them to disclose publicly we have the system so hence not on the list.

 

[sean]

Nobody working in HPC considers more than the top 20, maybe the top 5-10, and that's been true for more than 20 years. Most of the list is just general purpose computers that someone ran linpack on, and that's always been true.

I would say it's not quite that restrictive. It's more about representation in various subsets of the list. Top 100 is about the limit.

 

[mstone]

I would say it's not quite that restrictive. It's more about representation in various subsets of the list. Top 100 is about the limit.

By the time you get to #100 you're far from "High" performance.

 

[sean]

By the time you get to #100 you're far from "High" performance.

Sure. I'm going by what management says they use when looking at the Top 500.

 

[Evan]

All these big centralized systems are great but we have found to really support the big systems and dealing with data closer to the users and source we have been deploying a lot of small installs, just ~1500-2000 cores + some GPU’s in 2 racks with all storage and switching etc. much easier for power, cooling, cost of switches since much smaller infiniband setups.
Add them all up together and they are also significant.

Edge-HPC could be called.