Today we have an exciting announcement. The supercomputer Fugaku is being announced as the world’s top supercomputer. Fugaku is a joint development effort between Fujitsu and RIKEN. Instead of using off-the-shelf CPUs and GPUs, Fugaku utilizes a custom A64fx processor built with floating-point logic, Arm cores, and HBM2 memory to focus on HPC acceleration.
Supercomputer Fugaku by Fujitsu and RIKEN At #1
There are a number of notable firsts being attained with this machine:
- This is the first Arm-based server to be #1 on the Top500 list
- This is the first time Japan has held the #1 spot on the Top500 list since 2011
- This system is holding top spots in the Top500, HPCG, and Graph500 lists simultaneously
- This is the first system to have over 200 PFLOPS, 300 PFLOPS, and 400 PFLOPS of double-precision performance
For the six-year project cycle, the organizations have discussed real-world performance, not just linpack performance. Fugaku is designed to accelerate scientific as well as industrial workloads.
Just for some frame of reference, Summit with IBM Power and NVIDIA Tesla has been #1 on the Top500 list for some time at 148.600 rmax/ 200.795 rpeak PFLOPS. There is a lot of speculation that there are more powerful systems out there not being reported on the Top500 list but this is the benchmark list that the industry uses.
Final Words
While 415.53 PFLOPS is impressive, the Exascale series of computers are going to be coming out in the next 18-30 months. We know the US DoE has three systems (one Intel and two AMD) that are coming out in the next few quarters. Intel and Cray have said Aurora at 1 Exaflop, will be delivered in 2021. STH was on-site for the HPE-Cray El Capitan 2 Exaflop Supercomputer announcement. There, the US DoE set an early 2023 target for El Capitan which would put 2 exaflops on the June 2023 list. It is exciting to think we are getting a 5x performance boost in the next three years.
We are going to have more from ISC later this week on STH, however, we have some coverage that is being mixed in, such as the New NVIDIA A100 PCIe Add-in Card Launch. We will have our normal Top500 analysis in the next few days. There are a lot of interesting things happening. For example, AMD EPYC “Rome” CPUs are powering NVIDIA’s supercomputer at the #7 spot. This is great work by the Fujitsu and RIKEN teams so we wanted to get a quick congratulatory piece up to mark the occasion.
@Patrick:
“This is the first system to have over 200 PFLOPS, 300 PFLOPS, and 400 PFLOPS of double-precision performance”
In view of the ultimate statistic, surely the antepenultimate and penultimate statistics are redundant?
While the coming exascale machines will by definition be more than twice as fast when running HPL, it is possible that Fugaku will remain number one with respect to the HPCG benchmark. Since the high-performance conjugate gradient benchmark may be a better predictor of performance for practical computation, this is an important metric to use when comparing the computing power available to do science.