SeaMicro pushes 'Atom smasher' to 768 cores in 10U box

The drumbeat of system upgrades continues apace at microserver startup SeaMicro, with the company launching its third server node for its SM10000 "Atom smasher" in the past nine months. The new machine is based on the same Atom N570 chip and related NM10 Express chipset from Intel that was used in the prior generation of …

[Michael H.F. Wilkinson]

Sounds very interesting

I can almost fit my entire image (1.5 terapixel remote sensing stuff) into main memory

;-)

seriously, really interesting machine given the performance and power draw

[GCom]

Again whats the point

Great box but how many use cases are there in the real world where it dosnt matter that the main memory is unreliable, best i can do is 5 all around media delivery but thats a pretty narrow market especially coupled with the virtulisation limitations

[Tchou]

main memory is unreliable?

When was the last time you saw RAM failure?

I might be extra lucky, but never happened to me!

[YARR]

You might just be lucky....

http://www.cs.toronto.edu/~bianca/papers/sigmetrics09.pdf

"In this paper, we analyze measurements of memory errors in a large fleet of commodity servers over a period of 2.5 years. ... We find that DRAM error behavior in the field differs in many key aspects from commonly held assumptions. For example, we observe DRAM error rates that are orders of magnitude higher than previously reported, with 25,000 to 70,000 errors per billion device hours per Mbit and more than 8% of DIMMs affected by errors per year. We provide strong evidence that memory errors are dominated by hard errors, rather than soft errors, which previous work suspects to be the dominant error mode."

[i_must1]

it always helps to read the fine print

To me, it sounds like they just got a bunch of improperly designed/installed DIMMs. From the paper:

Conclusion 1: We found the incidence of memory errors

and the range of error rates across different DIMMs to be

much higher than previously reported.

Conclusion 4: There is no evidence that newer genera-

tion DIMMs have worse error behavior.

Question is that do you want to pay Intel for ECC for these kind of errors or just make sure the servers DIMMs are with reasonable quality? I bet that reasonable quality DIMM is cheaper than paying tthe Intel ECC (i.e. so called server part) tax.

[Tchou]

Actually a fair bit more than that

"That's 50 per cent more oomph and memory for 43.6 per cent more money"

Computing power is not necessarily as linear as a price.

But hey, nice hardware is coming!

[Bill Neal]

nice box

looks like the raptor transport container from jurassic park. SHOOT HER!!!

[Anonymous Coward]

Very impressive

And it looks like Dell thinks so too...

http://seamicro.com/partners

[HyperG]

Practical?

768 nodes is very impressive indeed. Although how many useful applications there are for such a large set of cpus with each cpu of which is a N570 which isn't very powerful. Webservers maybe but not anything else?

Having to think twice about what applications to run on an atom based server almost makes me think I'd be better off with a regular xeon based servers - at least I can throw anything at it and it is guaranteed to not come to a crawl because of single threaded performance not being up to par (?). I'd be curious to see any benchmark or performance data on the SM10000.

[i_must1]

One example app

http://www.cs.cmu.edu/~fawnproj/

[Bob Bobson]

The unanswered question is...

What sort of workload would a box like this be handling?

[Anonymous Coward]

High demand online

services, tenth of thousands of simultaneous connections.

Need a highly scalable server on it like truslleap's GWAN : http://trustleap.com/

Sorry for the advertising but looks very hot!

[dontfearthecloud]

Workload examples

looking at the seamicro site, it looks like there has been success with web servers and hadoop workloads. Definitely seems like a niche product, but that niche is the fastest growing segment of the market.

[HyperG]

Smaller and lower power but at what performance reduction?

Web servers, hadoop sure seem like the sweet spot - that would mean the Googles, Facebooks would be the target customers. Both of these however have standardized on Xeons/Opterons across the board. It remains to be seen if they would take the trouble of deploying SM in a few targetted portions of their datacenters. It would make sense if the performance and cost was substantial . Notice they mention 1/4 the power and 1/6 the space both of which relate to cost but I am curious to see what the performance reduction is. In the end it is all about TCO..