Hacker News new | ask | show | jobs
32 cores of fast ARM in half-depth 1U case (baserock.com)
65 points by rjek 5045 days ago
10 comments

Calxeda has been building 16 core cards[1] for awhile now with about 18 fitting in a 1U-sized box[2], each card connected by 10GbE similar to this[3].

[1]: http://www.calxeda.com/technology/products/energycards/quadn...

[2]: http://semiaccurate.com/2011/11/03/calxeda-launches-a-4-core...

[3]: http://www.calxeda.com/wp-content/uploads/2012/06/EnergyCard...

Not to mention Tilera, who makes 36-core chips/cards [1], with 4 in a 1U box [2], with each processor with connected to each its four neighbors by a 250 Gb/s link.

[1] http://tilera.com/products/processors/TILE-Gx_Family [2] http://tilera.com/products/platforms [3] http://www.tilera.com/technology

But Tilera uses its own architecture, ARM on the other hand is nearly everywhere.

There were other companies like Tilera: Transputer, Connection Machine.......SiCortex was founded in 2003 with a similar idea and was shutting down by 2009

The problems with these architectures is cost: mass market tech like x86 and ARM offers more bang for your buck

Interesting that these can run a single OS as a huge multi-core machine.
It might be worth pointing out that TileGX isn't ARM.

It is an interesting system though.

HP has also been working on Project Moonshot which (I believe) is based on Calxeda.

http://h17007.www1.hp.com/us/en/iss/110111.aspx

And you can buy them where, exactly? :)
Yeah, I see they have the same sort of "talk to us for pricing and more info".

Important things to note is that each of the nodes in this Slab have their own local SATA solid-state storage, and the individual CPUs are quicker. The Boston thing has higher density though.

Not a 32 core computer. It's 8 4-core SoCs with independent storage connected by a 5 Gbps network switch. Not a useless machine (it's a cheap way to buy lots of DRAM bandwidth, I'll note), but not what the headline is selling as.
With the right software and if you don't look it from too close, it can be considered a single computer.

We've been doing things like these for decades.

http://en.wikipedia.org/wiki/VMScluster

Yes, I'm that old.

Did the headline change? The one I see now just says that it's 32 cores in a tiny box, not that it's one machine specifically.
The site itself is a little misleading. It says "The Baserock Slab is a multi-processor server". To me, calling something a multi-processor server implies that a single OS instance has access to all the hardware. Instead, this is eight independent systems in a single chassis. Which is still awesome for that suggested use cases, such as an ARM build system.
I didn't say it was factually incorrect, I said (implied, I guess) it was misleading. You're seriously saying you read that and understood it meant 8 (!) distinct machines on one board?
I read it and understood it was a lot of cores in a small box. I presumed they weren't actually one machine, yes; but I didn't feel the headline was trying to imply anything either way.
So each quad-core SoC appears to the outside world as its own system and only communicates via Ethernet? Not useless at all, but it's not clear how the SoCs are presented.
Yeah, it's eight individual systems with two very fast Ethernet links to an on-board managed switch.
Looks neat. It should have a redundant power supply though. With redundant power just one of these could be a good platform for deploying fault-tolerant services.
I think the point of this kind of server is to build the fault tolerance at the rack, or at least 1U, level, not at the level of individual components.

If an individual server fails then your service shouldn't even blink.

The picture shows 8 tiny fans mounted inside (not counting the fan probably inside the PS)

I would have thought that part of the reason for going to ARM is the low power - hopefully no moving parts. It would have been cool (ha) to see them pull it off without fans.

When you have 608 of those in a rack I think the fans are very welcome :)
2GB per node - not enough IMHO (at least for the use I was thinking of ).
Cost?
At first I thought it was just a marketing page to see if there is interest, but on the spec sheet it says "Order now from Codethink for September delivery", but again, no price
What is the advantage of a half width server?

I mean with all the connectors on it it isn't like you can put two of em into a 1U slot.

Half depth, not width. Everything comes out the front, so you mount them back-to-back in a rack.
Been a while since I was heavily involved in high-density data centers... How does that work with traditional hot aisle/cold aisle layouts? Seems like you'd need to adjust the cooling to have an air channel up the middle of the rack?
You could do, or have one of them exhaust to its "front", and the other exhaust to its "back".
So the one at the "back" is getting pre-warmed air. Wouldn't this be a problem?

*Note I am used to normal data-centers but not high density stuff. I don't know of the special provisions a high-density setup would have.

In general, high-density or supercompute systems are installed with custom racks, cooling, power supplies, etc. This doesn't need anything custom, just something that isn't always the default in some DCs.

They're ARMs, they don't need much cooling anyway.

Rack then w/ ports out. And you can put two back to back.
From my understanding of ARM architecture, ARM cores cannot process heavy data loads in a power-efficient manner, making TDP under heavy server conditions prohibitively high. Am I right about this? I've never tried making an ARM server of my own.
What understanding is it that you have that suggests that? It can shovel data via DMA as well as the next architecture.
Well I'm not referring to reading/writing from a data store, but actually processing the data in registers and such. Arithmetic operations and that sort of thing.
High-density ARM servers are not ideal for every application. However, there are plenty of applications where you don't need massive amounts of computational grunt in each CPU, such as web serving, hadoop, mail and DNS servers, etc. Also, the density and lower power/cooling requirements can often offset the lack of performance. Each slab has a 260W PSU. The whole system can't pull more than 260W, ever. Some Intel CPUs need 200W just on their own.
Imagine a Beowulf of those!
I know you meant that as a joke, but I'd imagine one of these really would be an excellent candidate for being turned into a cluster.
I meant it as a joke, while also recognising that it would be ideal for being turned into a cluster. It's made me nostalgic about the days when I built Beowulfs at university!
Completely random, but I think it is funny the company is named "CodeThink Limited." That can't be good.