Quilted Dies of CPU & RAM was Re: [FoRK] Cheap Electronics Dissection Project

[Kragen Javier Sitaker]

On Wed, 18 Oct 2006 17:08:32 -0500, Shae Matijs Erisson wrote:
> ... the large octagon is RAM and the small square is CPU, and each
> edge is a commucations link to the chunk next to it, giving you a
> highly parallel computer where each CPU has links to four banks of
> RAM, and each bank of RAM has links to four CPUs and four other
> banks of RAM.
> http://www.missiontilewest.com/styles/nantucket/n32.jpg (octagon &
> square)

On Wed, 25 Oct 2006 12:38:52 +0200, Dave Long wrote:
> Traditionally, DRAM-style memory (tricky cells, regular layout) and
> CPU-style logic (simple cells, irregular layout) use different
> manufacturing techniques and lines, so it's relatively easy to get
> them in the same chip but relatively difficult to get them on the
> same die.

It seems pretty plausible that such a thing would work and be
relatively cheap per CPU, although if it were DRAM, it would probably
be more expensive than other DRAM, for the reason Dave cites.  (I
assume "same chip" here means "same plastic or ceramic package".)

Shae writes:
> I got this idea from reading about the Cell cpu and what drops the
> yield on silicon dies. One of the loss factors is diamond saws
> cutting into the chips accidentally when they saw the die apart. So,
> why saw it apart, and why not sell an entire die that mostly works?
> 
> If kerf ('saw off space'[1]) is currently left around the edges of
> dies, taking that out of the design would also save a bit more
> space.
> 
> So, mount the entire die on a socket and price it according to how
> much works.

While it might be relatively cheap per CPU, it would still cost a lot
per part, so it would only make sense when you want a lot of CPUs.  So
far, that's still a niche market.

> That way nearly every die could be sold for some money, and nearly
> everyone could afford a computer, as low yield dies would be cheap,
> and high yield dies expensive.

I suspect "low yield" usually means either 0 or 90+%, almost never 5%.

Dave writes:
> But I get plenty of hits for "embedded DRAM", so maybe this is now a
> solved problem?

I don't know; embedded DRAM might still be bigger than normal DRAM in
order to deal with processes that aren't optimized for DRAM.

On Wed, 25 Oct 2006 13:44:31 +0200, Eugen Leitl wrote:
> In an industry that only clocks progress on linear semi-log plots
> true innovation has become rare -- it's far too risky at the high
> end.

Right --- it's the anti-Innovator's-Dilemma.  As long as focusing on
making stuff smaller can get you 50+% improvements every year, you'll
beat all the Teras and Symbolicses just by doing that.