Australia teases binning x86 for Power CPUs in new supercomputer
Replacement for CSIRO's Bragg is required to hit a petaflop for just AU$4m
Australia's Commonwealth Scientific and Industrial Research Organisation (CSIRO) has issued a (regwalled) tender for a new petaflop-class supercomputer that can run either x86 or Power systems silicon.
But those keen to declare Intel's high-end reign is coming to an end should not get too excited: the tender may mention Power as a possibility, but also requires hardware compatibility with Windows Server 2012 R2 and Windows Server 2016. IBM once offered a tool called PowerVM Lx86 applications that translated x86 code on the fly, but Big Blue binned it back in 2013. This tender therefore looks like it is nodding to competition, not contemplating it seriously.
The new machine is a replacement for CSIRO's Bragg Accelerator Cluster, which was commissioned in 2009 and scored 156th place in the Top500.org list in June 2012, when it had 6,786 cores split among Xeon E5-2650 CPUs and NVIDIA Tesla 2050 GPUs. That rig was capable of boosting the super to 210.7 teraflops.
This time around CSIRO wants a petaflop-class machine but doesn't mind how it gets there: the tender calls for each compute node to “include at least two host processors with either a Power or Intel x86 64 bit architecture” but doesn't specify how many nodes are required.
Firm requirement include SUSE Enterprise Server 12.1 or later, at least 128GB of RAM and 500GB of local storage per compute and the presence of (unspecified) GPUs. Another requirement of interest ist that the new machine “shall be able to support existing science code that runs on an x86 architecture.”
The tender then says “If porting to another architecture is required associated costs should be provided,” which sounds like a wink to IBM as Big Blue recently updated its guidance on Porting from Intel x86 to Power systems running Linux.
But again, The Register feels safe in suggesting the possibility of porting is raised as a nod to competition, because the tender is for just AU$4m (£2.3m US$3m), which doesn't buy a lot of compute nodes. The tender also specifies that Bragg's existing InfiniBand network will be retained, as will the five racks Bragg currently occupies. Hanging on to racks hardly suggests a willingness to spend up big on code translation to make a move to Power, even if the architecture could meet the requirement of running WIndows Server.
The tender closes on December 19 and CSIRO says new machine “is expected to be up and running during the first half of 2017.”
The new “Advanced Accelerator Compute Cluster” will be tasked with “data analysis, modelling, and simulation in a variety of science domains, such as biophysics, material science, molecular modelling, marine science, geochemical modelling, computational fluid dynamics, and more recently, artificial intelligence and data analytics using deep learning.”
It's been a busy few days for Australia's HPC community, as last week the National Computational Infrastructure announced that a new Lenovo NeXtScale system will boost the power of its “Raijin” facility by 40 per cent. ®
UPDATE: In response to inquiries from The Reg, CSIRO has updated its tender and now says the Windows Server compatibility requirement does not apply to Power-powered servers. Some of the speculation above does not, therefore, apply. We've reported the change here.