Unless I see some detailed techiestuff on just how this works, or see a big-brand server manufacturer (IBM would do) or memory maker (Kingston perhaps) start selling these new super-DIMMs, I shall suspect it to be snake oil.

"The direction over the next decade is clear," Weber told us. "We're moving to multi-core chips that do more software threads. These processors will require much more memory bandwidth."

OK, you can double - or quadruple - the memory capacity of a server relatively cheaply. But that's not what the above quote is talking about. So, how does MetaRAM increase memory bandwidth? You penalize all memory accesses with increased latency and cram 2-4x as much data through the same bus. This does not seem like an improvement to me.

... you know how the rest goes.

Dang!

I meant to say Blade Servers too!

Basically, anywhere DIMMs slots are limited.

Hippster

Sounds a lot like the old Quarterdeck RAM Doubler for Windows 95, except in hardware.

The Quarterdeck product worked by compressing data before saving it to RAM (and decompressing it on the way out). It worked in theory but in practice was terribly slow and ended up being little more than a novelty.

Other than this one being on a chip and marketed towards servers rather than software for desktops, what's the difference?

These days, you could fit a fast CPU and 64MB of SDRAM on the same chip with a pile of hypertransport or PCIe links. Link 16 chips together, and you have a 1GB 16 core machine with low latency and high memory bandwidth.

I think you have misread.

This doesnt increase the capacity of an existing DIMM. It allows more DRAM chips to be fitted onto a single DIMM.

This means that, for the same capacity DIMM, you can buy more, lower capacity, cheaper DRAM chips. Alternatively you could buy the higher capacity chips and fit more of them on the DIMM, giving a higher capacity DIMM that would otherwise be possible.

With the products listed in the article, you can see an 8GB DIMM for £100!

I actualy looked into a design for something similar myself, but it was horribly complicated and not worth it for me (I would have been using an FPGA, performance would've been shonky!)

So, there are 2 important questions: 1) is there any way to shoehorn 4 of the 8GB modules into my server (an AM2 athlon64), and 2) can I scrape the £400 together to do it?

It slows main memory bandwidth, sure. It's much faster than if you have to swap to disk because of your capacity. Chances are, too, that if they're bright enough to pull this off they might just have other RAM-related technologies up their sleeves in the future. He might be alluding to something coming.