As one might suspect, "14.7 GHz" is roughly what you get by adding up the individual clock speeds. The overall tone of the video is a little goofy and this was obviously not meant by the creator as a serious performance claim.
That being said, it's still a cool project and I'd recommend just watching the video rather than reading third-party articles. It's only about 10 minutes long.
Le sigh.... dude mounted a bunch of underpowered CH32V203 microcontrollers on a bunch of simple boards and connected them by GPIOs. It is not a cluster in any traditional sense, and it is not 14.7 GHz.
Whoever wrote the wccftech article has both never heard of Amdahl's law AND was too busy making clickbait titles to view the original video...
What's "underpowered" about it? 32 bit processor running at 144 MHz (so 144 MIPS, less branch or memory stalls). That's similar to a Pentium 100 or PPC 601/100. And it's using 20 mW, vs about 10 W for those, so a factor of 500 better in energy use.
Certainly faster CPUs exist, but you don't always need faster, and for what it does it is pretty efficient -- and for a suitably embarrassingly parallel task 256 of them are still low power (about 5W).
Saving web searches: 144 MIPS x 256 processors = 36864 raw MIPS, or about twice as fast as an Xbox 360, or fast as up to mid-90s supercomputer, if we utterly completely ignore obvious node interconnect problem.
There are useful things that don't need a lot of interconnect speed. For example all of those XYZ@home projects (SETI, primes ...), or finding the hash for bitcoin blocks. Those just need aggregate speed and low energy use per computation.
This particular CPU might or might not have instructions needed to make those particular computations efficient, of course.
this things still need "enough" per-node performance, which potentially involve multiple cores with fast interconnects and definitely involves having enough RAM and cache etc.
also while the latency for @home projects are higher the bandwidth of your internet upload/download might be faster then what that cluster has effectively (due to the shared bus)
through this doesn't mean anything in the end
the project wasn't about creating a faster DIY special purpose compute system at home or anything like that
it's about a cool challenging and inspiring tech hobby project
compared to a super computer cluster a small highly simplified cluster with a fraction of total practical compute power then even a single node of many other cluster system but still a cluster
and the way it has blades and a bunch with each having multiple cards on them and a single shared buss connecting all of them and then a higher level bus which interconnects the blades is (from a generic far away POV) quite similar to how super computer cluster might be build, just much much simpler
Just because it's a super simplified cluster compared to any real world project doesn't mean it's boring or that you don't learn anything from it. But it means it is affordable to do as a hobby project.
That being said, it's still a cool project and I'd recommend just watching the video rather than reading third-party articles. It's only about 10 minutes long.
reply