[tgma]

Impressive, of course; but not quite that impressive.

Only true if all you're running is matmul (supercomputer has general purpose CPUs so more flexible than M1 GPU) - also those flops are probably FP64 in supercomputer ratings and FP32 in M1.

As a smart man I knew used to say, supercomputers are about I/O not raw compute. Those have terabytes of RAM not 8GB.

[reverius42]

I do wonder, how fast is the RAM on a 2000 era supercomputer vs. the disk on a 2026 MacBook Pro?

[tgma]

Your question hits directly at latency vs. throughput distinction. Depends on which you mean by "fast."

Throughput-wise, the supercomputer is competitive because it has a lot of local RAM connected to lots of independent nodes, which, in aggregate, is comparable to modern laptop's RAM throughput (still much more than disk) with a caveat, that you can only leverage the supercomputer bandwidth if your workload is embarrassingly parallel running on all nodes[1]. Latency-wise, old RAM still beats NVMe by two or three orders of magnitude.

[1]: there's another advantage that supercomputer has which is lots more of local SRAM caches. If the workload is parallel and can benefit from cache locality, it blows away the modern microprocessor.