https://videocardz.com/press-release/am ... ercomputer
U.S. Department of Energy now expects El Capitan to reach 2 exaflops once it’s fully installed, which would cement its place at the top of the US’s supercomputer inventory. El Capitan comes with a $600 million price tag and is intended to ensure the US’s leadership in supercomputers in the exascale era. Lawrence Livermore National Laboratory will be using the system to replace Sierra, their current IBM Power 9 + NVIDIA Volta supercomputer. All told, El Capitan will be 16 times more powerful than the system it replaces. LLNL will be using it primarily for nuclear weapons modeling – substituting for actual weapon testing – while the system will also see secondary use as a research system in other fields, particularly those where machine learning can be applied.
On the CPU side of matters, AMD will be supplying a standard version of their Zen 4-based “Genoa” EPYC processor. As it’s still two generations out from AMD’s current wares, the amount of information on Zen 4/Genoa is limited, but AMD is promising support for next-generation memory, Infinity Fabric 3, as well as broad promises of both single and multi-threaded performance leadership.
Meanwhile on the GPU side of matters, AMD and Cray are continuing to hold their cards rather close. While the companies are confirming that this will use a next-generation AMD GPU using a new architecture, they aren’t naming the architecture or offering too much in the way of details about it. For now, what they are saying is that these GPUs will be using next-generation HBM for their memory, and that they’ll bring support for mixed precision compute for improved deep learning performance.
For the first time AMD is naming their Infinity Fabric 3.0, which will be used to connect the processors within each blade. Like Frontier, El Capitan will be running in a 4:1 configuration, with four GPUs hooked up to each CPU. For Infinity Fabric 3.0, AMD is promising further improvements to inter-chip bandwidth and latency. However the most interesting claim is that these IF 3.0 device nodes will support unified memory across the CPU and GPU, which is something AMD doesn’t offer today. Indeed even Frontier is only slated to offer coherency between the processors which is a step below a true unified memory model. The devil is in the details of course – a unified memory system does not necessarily mean fast access to other devices’ memory – but this stands to be a major leap for AMD as a unified memory system can improve both the ease in programming such a system, and improving its performance when running heterogeneous workloads.
Finally, as previously mentioned, tying together the nodes will be Cray’s own Slingshot interconnect. Among other things, Slingshot supports adaptive routing, congestion management, and quality-of-service features. The interconnect is capable of 200Gb/sec per port, with individual blades incorporating a port for each GPU in the blade so that other nodes can directly read and write data to a GPU’s memory.
El Capitan is slated to use less than 40MW of power – and we’re told it’ll be "fairly substantially under that" – however at this time the DOE isn’t disclosing the total number of cabinets. But to put things in comparison, Frontier is slated to use 100 Shasta cabinets, with a total power budget lower than El Capitan. So we wouldn’t be too surprised to ultimately find out that part of the reason that El Capitan is 33% faster than Frontier is due to the DOE throwing more hardware at it and ordering more cabinets. But whatever the number, it’s going to be enough that El Capitan will be using direct liquid cooling.
Overall, El Capitan marks an important second exascale supercomputer win for AMD, while Cray will now be involved in all three US exascale systems. So it’s a big win for both vendors, and a continuation of momentum for AMD, who only just scored its first big supercomputer win in a long while with Frontier last year. The fact that El Capitan is a derivative of Frontier also means that with all three exascale systems now locked in, it will be NVIDIA who finds themselves on the outside looking in for this generation. As we noted with the Frontier announcement, the Intel Aurora and the AMD Frontier/El Capitan systems are coming from full-service processor vendors that supply both CPUs and GPUs. Current-generation systems like Summit use mixed vendors – e.g. IBM + NVIDIA – so the move to integrated vendors is a big shift for these CPU + accelerator systems. And while it makes a lot of sense for LLNL to order a copy of one of the other exascale systems in the name of efficiency, it should be noted that US DOE supercomputer contracts are as much political as they are technical. The US has a vested interest in supporting a domestic supercomputer industry and ensuring there are viable competitors to help keep costs down (there used to be several), so with three major processor alliances/vendors in the US, someone was bound to end up the odd man out.
At any rate, El Capitan is scheduled for delivery in early 2023.