[Cryptography] Nvidia- one Tflop could change the rules quicker than quantum...

[D. Hugh Redelmeier]

| From: Tom Mitchell <mitch at niftyegg.com>
| Subject: [Cryptography] Nvidia- one Tflop could change the rules quicker than
|     quantum...

TFLOP grates since FLOPS means FLoatingpoint Operations Per Second.
Of course English speakers instinctively think that the S is to make
it plural, but that s isn't in the acronym.  It sure would be clearer
if the industry had chosen / in place of P.

Floating point operations would not seem to be the most useful kind
for crypto.  Of course GPUs can do integer operations too.

There's something funny about how all the PR about this board only
talks about "deep neural networks".  Does that mean it isn't so good
at other things that GPUs have been used for?

| Little machines like this from nVidia could change the computational
| rules for crypto quicker than quantum computing.

Right now desktops with GPUs are cheaper per TFLOPS.

| When the price comes down it may be a productive adventure
| to explore how this can be harnessed for more than a drone.

These might be good at floating point operations / joule.

I suspect that the interconnect between the ARM CPU and the GPU is
faster than a PCIe bus.  This might be important.

This card is probably less interesting than the chip.  Not sure
how well a pile of them can be interconnected.

For some reason, bitcoin miners preferred AMD GPUs (before they
switched to ASICs).  I understand that AMD cards were faster than
NVidia cards for the crypto operations.

AMD APUs have been widely available for some time and AMD has been
pushing a shared address space between the CPU and GPU.  NVidia has
only recently gotten a CPU core to fuse with their GPU.  Their first,
the Tegra K1, uses a 32-bit ARM and that's not so great for heavy
crunching.  The Jetson TX1 uses the Tegra X1 chip with a 64-bit ARM core.