New RAM tech quadruples capacity!

[f0dder]

This is pretty cute, if you have massive memory needs . Sure, you'll (currently) be stuck with DDR2-667 speed, but you can get a system with 256 gigs of memory and possibly avoid going to the harddrive for your database...

"We had to make our chip look like a DRAM to the memory controller, and like a memory controller to the DRAMs," said Suresh Rajan, the MetaRAM co-founder whom I talked to about the company's technology. This memory traffic routing messes with the DDR2 DRAM timings quite a bit, so the MetaRAM chipset's dynamic command scheduling circuitry ends up doing a kind of "out-of-order execution" with the flow of reads and writes so that the DIMM can operate at a full 667MHz without any glitches.

[wreckedcarzz]

If that works for traditional desktop PCs, I want some, and I want some BAD. Plus, just imagine what we will be asking for when we buy new PCs:

"Yes, I want just 500TB of RAM- I need to keep costs low"

[tinjaw]

This should help boost the adoption of virtual machines. I bet we will see 4 x Quad CPUs with this type of memory. You can pack a lot of computing powers in a little bit of physical space with such a configuration.

[f0dder]

This should help boost the adoption of virtual machines. I bet we will see 4 x Quad CPUs with this type of memory. You can pack a lot of computing powers in a little bit of physical space with such a configuration.

Yup, Virtual machines could benefit a lot from this, also by having more aggressive disk caching and avoiding harddisk thrashing. And database servers - even moderately large databases could be kept mostly in-RAM (obviously you need some good UPS to prevent loss on power-off ).

For home systems, you're better off with less but faster RAM, though. I have 8 gigabytes in my own box, and that's totally overkill and I hardly ever push that limit

[Lashiec]

* Lashiec orders a few motherboards

[Carol Haynes]

It looks really inventive and a novel approach to an old problem but if the PC 'sees' them as standard DIMMs how does it map the new memory size? Surely if the mobo controller allows a maximum of 1Gb per slot (like most mobos do) how can the same hardware even know of the extra RAM if you shove in an extra 8Gb card? Even if the memory board additional controller fast swaps the RAM in and out of existence the application will still only actually see 1Gb.

Or am I missing something?

[Deozaan]

Is 667Mhz that much of a bottleneck for RAM? If you're talking about servers with 500GB databases, sure. But for the home user aren't we more limited right now by capacity than by bandwidth?

[f0dder]

Carol: obviously motherboards need to support this, but the technology is targetted at server class hardware, not end-users (where the common limit right now seems to be 2GB per slot, four slots, and thus 8 gigs of memory max).

Deozaan: I honestly don't know where the bottlenecks are - core2 CPUs have a FSB speed of 1066MHz or 1333MHz for the new 45nm CPUs - thus, theoretically, you'd want the RAM to be just as fast. Dual-channeling ram does give a performance boost (although probably not 100% linear scaling), but you might be able to feed a 1333MHz FSB with DDR2-800 in dual-channel mode? But then the large & aggressive core2 caches come into play as well, combined with application memory access patterns, so superfast ram might not even be necessary...

I honestly don't know how it all sums up, and I haven't been able to find benchmarks that say anything really conclusive. I ended up opting for DDR2-800, since it was still pretty cheap (though not dirt-cheap as DDR2-667), and much cheaper than DDR2-1066 (not to mention the current horrible DDR3 prices, ugh!). So far my system has been screamingly fast, and I don't really want to know if it has any bottlenecks

I also dunno if memory speed would be a bottleneck for those huge databases, they'd already run insanely much faster with a massive memory system like that, instead of constantly going through a disk subsystem. DDR2-667 can push a whopping theoretical 5.3GB/s already, which is much faster than any RAID system I've heard of

And I wouldn't say home users are capacity-capped, really... you might be able to utilize 4GB of memory under peak situations, and 8GB if you're doing extreme stuff,  but as the current date (almpost-March 2008, should anybody not know ) I can't see where a relatively standard workstation user could utilize more than 8GB of memory...

[Lashiec]

Ah, I read a benchmark time ago, and the differences in performance between DDR2-667 and DDR2-800 were something like 1%. I can't remember where it was, though.

Anyway, I rarely go over 600-700 MB at the end of the day, my peak is around 1.5 GB when playing demanding games for some time. Still, it would be nice to have 4 GB, don't know why, bragging rights maybe

[f0dder]

Lashiec: 1% difference for that particular benchmark - they probably tested something that was mainly CPU bound. Video compression, for instance, does a lot of calculations... and many speed-critical algorithms have already been optimized for "stripmining", processing chunks of data that fits in CPU cache, so you don't end up re-requesting memory and go through the slow RAM, but can work in the fast cache...

So, what RAM throughput limited algorithms be? Hm. Searching through memory comes to mind... probably compression as well, when you want to find longest-match strings. And when you add CPUs with more cores, you might start struggling for RAM bandwidth as well.

Btw., to utilize 4GB you really do need a 64bit OS - you could more under a 32bit OS if Microsoft hadn't decided not to take full advantage of PAE on non-server OSes for greedy market segmentation compatibility reasons, but oh well. And you would be able to utilize it, even though each 32-bit app would only be able to use 2GB (or 3GB for large-memory-aware apps) at once, either for running multiple apps, or filesystem caching.

I opted for 8gigabytes because I could , it's cute being able to assign 5 gigabytes to a RAM disk and then manipulate DVD ISOs without going to the harddisk. But heck, under regular use, I don't go much above 2gig that often. Long live overkill!

[Lashiec]

Oh, what Google can't do

Yep, actually one of the reasons to buy 4 GB (well, it's on the upgrade list almost since I bought the computer) is Vista, if SP1 proves to be as good as it should be, I'll get another pair of sticks and move to the x64 version. Finding XP 64 here is almost impossible, and it's not an OS I would consider anyway. It won't happen anytime soon, I'll let other users to be guinea pigs experience Vista while I wait for 2 GB modules to come down in price (I hope the downward trend in memory prices continues for some months...).

[f0dder]

XP64 works nicely for me so far, but as I've stated a couple of times, I don't run funky hardware or peripherals... and I make sure to find cracks enhancement patches for the games I buy. The only issue so far is that I can't install MSN Live Messenger, because the Live downloader refuses to run on Server OSes and XP64 - no idea why, but Microsoft is fond of artificial limitations (btw. for people on 32bit OSes with 4gig memory frustrations: pre-SP2 there was a limit on 4GB of physical RAM, with SP2 the limit was changed to 4GB of address space. Subtle difference, but makes a large difference).

Games probably aren't a good benchmark test for RAM speed, they should be largely GPU bound, with some CPU stress as well when advanced physics and AI are put to play. For a better benchmark, you'll want something that can keep a quadcore busy... but still doesn't do too heavy calculation on the data, but keeps it streaming through. Perhaps MD5-summing?

Vista might be usable around SP2...