Latency seems to go up as x, while storage size goes up as x^2 (or even x^3), where x is 1/featuresize. Now the constant factors aren't so great, but perhaps caching helps for a lot of this type of arithmetic (e.g., Benford's Law). Also, you can spend some of these spare cycles *interpolating*, since you will likely get back a large number of nearby points after the long latency wait. Unfortunately, I've forgotten how to interpolate certain types of tables -- e.g., modular & EC calculations. At 09:56 AM 4/2/2017, Tomas Rokicki wrote:
Probably latency is much more important than bandwidth for this application.
The latency of these cards is typically 100 microseconds for a random read (this is for a fast M2 card; secure digital cards and SATA are much slower).
At 3GHz that's 300,000 cycles.
Intel's XPoint (now branded Optane) promise much better latency, as low as 9 microseconds, but even this is 27,000 cycles.
With a GPU and 9 microseconds you can do an awful lot of computation.
On Sun, Apr 2, 2017 at 9:46 AM, Simon Plouffe <simon.plouffe@gmail.com> wrote:
Hello,
a math table lookup on a SSD micro card ?: why not ...
ok but the reading speed of those are something like 90 megs per second on a 'class 10' card,
on the other hand there are those new ssd that can be plugged into a PCI express 4x slot (if you have one on your pc),
http://www.materiel.net/ssd/samsung-serie-960-pro-m-2-pcie-vnme-2-to-136308....
the reading speed is 3500 megabytes per second. That's another animal, the drawback is of course the price and your electricity bill. ;-). and you have 1 TB of space to put tables. the price is about 650 us dollars.
best regards,
Simon Plouffe