Oakridge Frontier 1.5 Exaflop Supercomputer

Al Geist presented a history of the Exascale Project and How Frontier Solves the Four Exascale Challenges. This is part of the Advanced Technologies Section (ATS) series of the National Center for Computational Sciences at ORNL.

Chip and Network vendors greatly improved the reliability of their systems to push this problem beyond the Exaflop scale.

GPUs enabled more general supercomputer capability and GPUs hides a thousand to ten thousand parallelism. This enables an exaflop with 9000 nodes instead of millions of nodes.

Oakridge Frontier 1.5 Exaflop supercomputer appears on track for delivery by the end of 2021. Other US exaflop supercomputers were delayed by delays with Intel chips. Oakridge Frontier will use AMD chips. China’s exaflop supercomputers have been delayed because of trade restrictions and sanctions.

ATS Seminar Series: Al Geist from OLCF on Vimeo.

4 thoughts on “Oakridge Frontier 1.5 Exaflop Supercomputer”

  1. What new capabilities will exaflop machines enable? Will weather be predicted a month in advance? General AI? Unemployment of human coders? Video conversation Turing test achieved? World domination?

    Reply
  2. You throw GTP-4 at this or some of the stuff Tesla is doing. Could get real interesting real quick. If a program can "fake" total natural language or driving a car better than 99% of humans 99.999% of the time. You have some real game changers.

    Reply
  3. It's fun to compare this to the human brain. If we have 10^15 synapses, running at 1.5 kFLOPS each, then this supercomputer is equal to a single brain.  

    However, we don't really know how many synapses we have (probably somewhere between 10^14 and 10^15), nor how many FLOPS it takes to be equivalent to one synapse (probably 1 kFLOPS), nor whether there's any significant computation going on other than in the synapses (Penrose has some claims about quantum processing in the cell nucleus, but they're probably nonsense).

    In any case, our brains are highly interconnected, in a way that probably vastly exceeds the internal bandwidth of a supercomputer like this. So it's likely the supercomputer still hasn't quite reached what a human brain can do. But it's getting closer.

    Reply

Leave a Comment