I’m presenting a paper on millicomputing at the HPTS workshop in October http://www.hpts.ws and at CMG07 in San Diego in December.

Cheers Adrian

Harold,

An Intel person confirmed to me that the FB-DIMM controller consumes five watts. That is why the FB-DIMMs need the extra “heat spreaders”. The original idea was to use standard high-volume DIMMs, presumably to save money, but the addition of the controller and heat spreaders has eliminated that advantage. Adding the 5 W penalty per stick just adds insult to injury.

Dave,

I don’t follow AMD, but all of Intel’s post-Netburst CPUs are based on the mobile architecture that the Intel Israel came up with. In a very real sense, we are all using mobile architectures now.

Also, I wouldn’t expect an IT guy to piece together an energy efficient server. Vendors should. I suspect Intel has the jump on HP, IBM and Supermicro due to the Google contract for – among other features – energy efficient servers. Whether they can figure out how to communicate the benefits is an open question.

Thanks for the pointer to VIA. A couple of folks mentioned them and now I’m curious.

Adrian,

Good to hear from you. Are you still with Sun?

I love the milli-computing concept. Gordon Bell’s concept of a new generation of computing every 10 years is bumping into human-scale limits. Rather than make ever-smaller computers, at some point we reach a reasonable size and price point and build as much as we can into it. A cluster in the palm of your hand!

Wes,

I *think* I’m seeing 2.5″ drives beginning to close the gap with 3.5″ drives. Like all prior drive migrations, the SFF drives don’t need to match the $/GB to win, just get close enough so that their other advantages close the deal. As I noted last year, flash drives are a very power-efficient alternative. Drive vendors see the writing on the wall: the 3.5″ drive won’t live forever.

Marc, I couldn’t agree more!

Jonathan, thanks for flagging that. I think I’ve fixed it. Let me know if there is still a problem.

Robin

I have been taking this general approach to the extreme with some work that I have called “Millicomputing” – I didn’t just look at mobile CPUs with low power disks, I have been exploring how to build useful (for some classes of application) enterprise servers out of CPUs designed for battery powered devices using flash based storage. These have an overall power per node of less than one watt, hence the name millicomputer, since they are measured in milliwatts.

Definitely a good article. As someone who’s done a lot of work with low power platforms I found it refreshing to see that there’s now a benchmarking tool available for efficiency, etc. One thing I’d like to point out is that the “Ideal System” makes use of a mobile CPU, not a consumer CPU. This does skew the envelope somewhat since the design of those systems is not something easily entered into by Joe IT User. There’s a couple of players, additionally, in that efficiency space that have more mature designs that, while not focusing on massive I/O for storage, easily enter into power efficient computation space.

The first contender is ALWAYS VIA. Their Eden platform (in mini-ITX and pico-ITX formats) is the class leader when it comes to power efficiency and thermals. Not only are they exceptionally cool and quiet, their onboard AES and Padlock encryption units make them ideal for security appliances and the like. Companies like Ainkaboot (http://www.ainkaboot.co.uk) have also developed 4U clusters using these processors and low power storage devices.

The second contender, while not currently leading the pack, is AMD. AMD actually has better power management (scalable clock multipliers, etc.) that allow a deeper idle state than Intel. This will only increase in coming months with separate power planes for the onboard memory controller and processor cores. Coupled with Hypertransport’s ability (ver. 3.x) to ungang and dynamically scale width and power requirements, this platform becomes very competitive. On the mobile front, this isn’t currently the case but, it should be noted for comparison purposes especially at the consumer level.

Anyhow, those are just some of the thoughts floating through the ether at this time.

I’d love to hear your thoughts.

cheers,

Dave

http://storagemojo.com/2007/05/22/intels-best-and-worst/

and that it’s not inconsequential in a “holistic understanding” ^_^.

Note what I assume to be the FB-DIMM penalty in the first “balanced server” system; I’ve just finished building three different systems (replacements for old ones for my father and myself) for which I did peak power budgets, so I have the numbers still in my head.

I just bought some Kingston (w/Micron DRAM) 2GB unbuffered ECC PC2-5300 sticks, and the data sheet says they weight in at 2.2 Watts each.

A little bit of Wikipedia and Googling indicates the CPU almost certainly uses FB-DIMMs, and chart says their 2GB sticks weigh in at 7.5 Watts each. About 5 Watts more each!

For someone like Don MacAskill for whom DRAM in his database servers is the single most important thing in selecting a motherboard, that’s going to add up to a bit of an ouch if he chooses to drink the FB-DIMM Kool-Aid. Only so much, 40 or 80W for 16 or 32GB (but that brackets this CPU’s TDP), he as I remember uses 15K disks by choice and would use faster ones if available, but it all adds up.

- Harold