From: Eugen Leitl (eugen@leitl.org)
Date: Thu Aug 15 2002 - 03:19:12 MDT
On Wed, 14 Aug 2002, J Corbally wrote:
> Depends on the density, which depends on the cell size, which to a
Please notice that while desktops march towards GByte size memories, most
portable embeddeds operate on a 32-64 MByte footprint. MRAM will kill both
the flash and DRAM in these architectures, resulting in cheaper devices
with a much reduced power drain (a major issue with portables). Many
people are wishing flash would go away.
> large degree depends on the mask size, and ultimately its
> bandwidth/latency. May be that instead of using it as main memory,
> they stick in just enough to store all information necessary to do
> instant restarting. The main problem with "instant on" currently
No, on the long term MRAM will probably supercede DRAM in desktops. The
only major disadvantage I'm seeing that it might be harder to make
embedded MRAM than embedded DRAM. No telling when rotating magnetic bits
will have to go, though.
> isn't the memory, it's the OS. ACPI 2.0 is supposed to help a lot,
Not just the OS, it's the entire way we do things. If you go to persistent
object store and garbage collection (no pointers no more) you most
assuredly can't afford going to insane state every day or two and have to
reboot, even though you could resume at last known sane state dumped to
nonvolatile storage.
*nix does achieve uptime of years, though, so at the very least one would
be able to move it over without changing too much.
> but you'll still have old peripherals and old apps that'll go screwy
> cos they were never written to fully allow for suspension and power
> off modes. Windows is still a bit away from flawless "no power"
You know what the half life of hardware is. Old peripherals will die out
in a couple of years.
> resumption as well.
Well, who cares about what Windows does or fails to do.
> As for MRAM, I'm not familiar with it's performance characteristics, but
> it'll have to be at least up to current DRAM, with comparable room to scale up.
It looks that way. It has a lot to cover in designspace, though. DRAM
processes are very mature, MRAM starts from scratch.
> I've heard about this disk. What I'm concerned about is whether
> they'll sell it to the highest bidder in which case it'll be hamstrung
> like DVD, which suffers from the "too many CEOs, not enough engineers"
> scenario. The CD was invented to facilitate Digital music. The DVD
> was invented to protect company profits. And it shows. No rewritable
> version, discs that only work in certain countries and so on.
The sharks seem to be focusing on DRM systems these days. Even if you have
a violet version of DVD ROM with fat content people will rip it with
advanced compression. A DVD image compresses to less than a GByte, there
are lot of GBytes on a TByte drive. The only way to address this is
hardware DRM.
> PCI was the saviour of the PC, IMHO. Genuine high speed PnP even on
> pre-Win95 systems. I used to debug them, lovely system. Good to see
> they've been keeping it current. They've also created a High speed
> serial PCI standard from what I hear, fully backward compatible.
> Then again, motherboards have had what was essentially Serial PCI for
> a few years now.
Things are moving towards serial high-speed buses in general, thankfully,
though they're not packet switched yet (HyperTransport on crossbar is the
best you can buy). It should be rather doable to go 10 GBps on a serial
line as long as geometries are short, and you can certainly switch that
fast.
> I'm not too surprised to see a GPU that fast. Graphics cores tend to
> have small and highly optimized instruction sets, as well as
> architectures that can be better optimized as the type of work they do
> is pretty consistent in its nature. Since large caches don't provide
Much more interesting is the advent of embedded memory and more flexible
architectures. Things will become truly interesting if you can run e.g. a
neuronal code on a graphics accelerator (we're seeing CA demos on nVidia
hardware, though only 2d).
> the same kinds of performance gains that they do in CPUs, plus you can
> dump a lot of the predictive/predicative hardware, you've got much
Notice that with embedded memory you don't need crystal ball hardware and
caching, and also can dispense with all I/O but highspeed serial. As the
result you can shrink the die, reduce power density, or put more memory on
it.
> more die space for more hardwired functions and processing pipelines.
> Not to mention the very high speed dual ported RAM they slap on.
As long as the GPU is dumber than the CPU you need low-latency
high-bandwidth access. As soon as the game rendering logic resides
entirely within the GPU, and GPU doesn't look that different from a CPU
you don't need that.
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 09:16:08 MST