A 2nd generation of ZRAM is emerging

[Yousuf Khan]

"LONDON — Innovative Silicon Inc. (ISi), a startup pioneer of
"floating-body" memory, has come up with an improved version of the
technique that provides a factor of ten improvement in the 1 to 0 margin
and the data retention time. Advanced Micro Devices Inc., which took out
a license for the previous generation of the Z-RAM in December 2005, has
taken a license out for the second generation of the technology. "
http://www.eetimes.com/news/semi/showArticle.jhtml?articleID=196601127

 

[Tony Hill]

"LONDON — Innovative Silicon Inc. (ISi), a startup pioneer of
"floating-body" memory, has come up with an improved version of the
technique that provides a factor of ten improvement in the 1 to 0 margin
and the data retention time. Advanced Micro Devices Inc., which took out
a license for the previous generation of the Z-RAM in December 2005, has
taken a license out for the second generation of the technology. "
http://www.eetimes.com/news/semi/showArticle.jhtml?articleID=196601127

All well and good that the second gen. is on it's way, but did anyone
ever *USE* the first generation technology?

Still, it sounds reasonably promising for L3 cache in quad-core
processors. It looks like it will be far too slow for L2 cache, but
it might allow AMD to get by with 4 cores with 512KB or 1MB of L2 each
and then a large (16MB or 32MB) shared L3 cache made up of this ZRAM
stuff.

I'm not holding my breath though! This is not something that's going
to suddenly appear tomorrow. And let's keep in mind that Intel has
managed to do some rather incredible things with plain old SRAM (in
terms of performance, size and power consumption).

 

[Yousuf Khan]

I'm not holding my breath though! This is not something that's going
to suddenly appear tomorrow. And let's keep in mind that Intel has
managed to do some rather incredible things with plain old SRAM (in
terms of performance, size and power consumption).

What have they done with SRAM that's so incredible?

Yousuf Khan

 

[Tony Hill]

What have they done with SRAM that's so incredible?

Shrunk it down to extremely small sizes while keeping it's performance
high and the power consumption fairly low. They aren't going to match
ZRAM for size, but they'll be able to stay close enough that their
traditional manufacturing advantage (ie the fact that they can make
bigger dies cheaper than anyone else) should be more then enough to
keep them price-competitive for some time to come.

Basically what I'm getting at here is that ZRAM is a rather promising
technology (from what I've read of it at least, I certainly haven't
been involved in trying to BUILD the stuff!), however it's definitely
not about to displace SRAM. It looks to be a really good design for a
large L3 cache in modern processors. Performance-wise and cost-wise
it fills a nice gap between SRAM and conventional DRAM.

 

[David Kanter]

What have they done with SRAM that's so incredible?

Have you ever looked at the performance and density of their caches
compared to everyone else?

Just look at Montecito's cache hierarchy, it's darn near amazing, the
130nm P4 as well. The latter had a 4way 8KB L1D with 0.66ns latency.
The former uses a 16KB L1D with 0.625ns latency (can't recall
associativity). Montecito's L3...which is 12MB, has a 8.125ns access
time, which is amazing...when you remember that it was optimized for
density and not access time.

DK

 

[Yousuf Khan]

Shrunk it down to extremely small sizes while keeping it's performance
high and the power consumption fairly low. They aren't going to match
ZRAM for size, but they'll be able to stay close enough that their
traditional manufacturing advantage (ie the fact that they can make
bigger dies cheaper than anyone else) should be more then enough to
keep them price-competitive for some time to come.

It's not entirely certain that they can make big dies cheaper than
anybody else. Otherwise, Itanium would've been economical, whether or
not it would've been successful is still debatable. Also recently, AMD
took the lead in CPU gross margins for two quarters in a row.

Basically what I'm getting at here is that ZRAM is a rather promising
technology (from what I've read of it at least, I certainly haven't
been involved in trying to BUILD the stuff!), however it's definitely
not about to displace SRAM. It looks to be a really good design for a
large L3 cache in modern processors. Performance-wise and cost-wise
it fills a nice gap between SRAM and conventional DRAM.

Well, they're talking about displacing SRAM in large caches, like L3.
The latency seems to go up with larger SRAMs, whereas the latency goes
up less so with ZRAM. At some point, the latencies meet up and then ZRAM
gets less.

Yousuf Khan

 

[nobody]

It's not entirely certain that they can make big dies cheaper than
anybody else. Otherwise, Itanium would've been economical, whether or
not it would've been successful is still debatable. Also recently, AMD
took the lead in CPU gross margins for two quarters in a row.


Well, they're talking about displacing SRAM in large caches, like L3.
The latency seems to go up with larger SRAMs, whereas the latency goes
up less so with ZRAM. At some point, the latencies meet up and then ZRAM
gets less.

Yousuf Khan

We'll measure the latencies when it comes out - or rather leave the
actual benchmarking to the likes of Anand. However now it's a moot
point - there is nothing to measure but wapor.

NNN

 

[Tony Hill]

It's not entirely certain that they can make big dies cheaper than
anybody else. Otherwise, Itanium would've been economical, whether or
not it would've been successful is still debatable. Also recently, AMD
took the lead in CPU gross margins for two quarters in a row.

Obviously I don't have access to AMD or Intel's respective
manufacturing costs, but I think it's pretty safe to say that Intel
has an advantage. They have moved very effectively to 65nm production
and 300mm wafers, two things that AMD (and most other players in the
world) are still working on.

As for gross margins, certainly manufacturing cost is half of that
equation, but the other half is average selling price. Here AMD was
doing VERY well until Intel brought out their Core-based chips.

Well, they're talking about displacing SRAM in large caches, like L3.
The latency seems to go up with larger SRAMs,

The increase in latency is often due more to optimizations than actual
limitations of SRAM.

whereas the latency goes
up less so with ZRAM. At some point, the latencies meet up and then ZRAM
gets less.

Well no one at ISi seemst o want to give a straight answer as to just
what the latency of ZRAM is, except that it's slower than SRAM and
faster then DRAM. Given that those two technologies are roughly an
order of magnitude apart, that leaves a VERY wide range. Personally I
wouldn't hold my breath for any thing too close to SRAM speed. If
ZRAM is going to succeed it will be first and foremost due to higher
densities, not performance.