[extropy-chat] Plastic promises dense data store

Eugen Leitl eugen at leitl.org
Sat Nov 29 18:34:47 UTC 2003


On Fri, Nov 28, 2003 at 10:52:38PM -0500, Dan Clemmensen wrote:

> True. But this is a multidimensional problem. In the classical 
> informatin management
> domain, we can use hierarchical memory techniques to drive the effective 
> latency down.

Memory hierarchy is a neat approach. Completely breaks down
when the data locality assumption is violated, though.

Stuff is CPU speed when it's in the registers.
Stuff is slower when it's in 1st level cache.
Stuff is slower still when it's in 2nd level cache.
Stuff in memory is slower still, and if the access
pattern is high-surprise (i.e., there is no pattern) it
breaks down futher still.

Even so, 50 ns is a long, long, long long way from 5 ms.

It *might* be okay if I already know for what I'm looking, and
just need to fetch a chunk of bits with mostly contiguous-movement
of the head.

It's a harsh constraint, and it ruins the domain of algorithms
which we can run within reasonable times.

> At the TB level, we can employ a bunch of disks in parallel to reduce 
> the effective latency.

How much disks can I afford though, and how does this scale? Even
if it's 90% sure the bits are on the same cylinder, you still have
to wait for them passing through under your head. Even at 10 krpm,
that's an effective eternity.
 
> The point is that you can actuall create a TB storage system for 
> <$1000US, quantity 1, today.

Sure, but what do I do with it? It's a great movie library, but
how does it help us with designing nano, or creating AI?
 
> I don't think that "plastic mem" will operate in the ns range. Memory in 
> this range is specialize,
> e.g. ZBT, FCRAM, or RLDRAM, and it is not hi-density. DDR mem can reach 
> 1GB/s throughput,

Right now you can have 6 GByte/s (3 GByte/s bidirectional), best-case
with prefetch/streaming. Of course, there are just not that many things
you can do in that access mode.

> but has latencies about 10ns.

If it streams, you're golden. If your fetches are PRNG driven, you'll do
lots worse than 10 ns.
 
> "Plastic mem" is likely to fill a gap in the hierarchy. Problems that 
> are amenable to a hierarchial
> memory may benefit.

I don't think it's any good. It's write-only, that niche is movies,
and we already have DVD killing CDR, and blu-ray is about to ship.

It would be interesting if they could fit a TByte in a cm^3,
and have some >>10^6 r/w cycles, at flash speed or better.
 
> As a separate issue, massive storage requirements are generally also 
> "write-once" instead of RW.

Yes, unless you're looking at a large physical simulation.

> To he extent that this is true, a printer is already capable of creating 
> cheap extensible storage.
> assuming a 20% overhead for ECC, we have 1200dpi can store  a mabagit 
> per square inch.
> This is one megabyte fo eight square inches, or 10 megabytes on an 
> 8.5"x11" page.

-- Eugen* Leitl <a href="http://leitl.org">leitl</a>
______________________________________________________________
ICBM: 48.07078, 11.61144            http://www.leitl.org
8B29F6BE: 099D 78BA 2FD3 B014 B08A  7779 75B0 2443 8B29 F6BE
http://moleculardevices.org         http://nanomachines.net
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 198 bytes
Desc: not available
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20031129/33c27757/attachment.bin>


More information about the extropy-chat mailing list