Intel was showing off its re-worked Larabee platform in Tapei this week. Renamed MIC ("multiple integrated cores"), the production samples were on a 32nm die with 32 cores, 8MB shared L2 cache and 4x hyperthreading, running at a low 1.2Ghz, meaning it can execute 128 threads at once. The first, single precision, tests from the Linpack suite look like giving the processor a performance of around 517GFlops, compared to the 360GFlops available through GPGPUs from nVidia and AMD/ATi.

The first production samples should start appearing towards the end of 2011 or in early 2012, on a 22nm die, with 50 cores, capable of 200 threads.

Source: c't Magazine 2010/13 page 20.

Wow that will be handy for my SETI data sets!

How many years (or decades) will it be before most software can take advantage of that many cores?

Software can already use that many cores - just look at the thousands of cores in some of the super computers that are used in Unis etc.

The real question is, where does consumer software lose the benefit of ever more cores? Things like normal OS tasks, word processing, internet, email and simple spreadsheets rarely need more than a couple of cores, because, with the exception of the OS itself, they are just sitting around waiting for user input, plus a few flurries of activity when loading and saving data, indexing etc. Even video encoding can't use an infinite number of cores, although it seems to scale well to 8 or 16 cores at present.

Massively parallel computers are either going to be for specific scientific tasks or for multiple users, with thin-client computing coming back into vogue - we are certainly looking at it, as individual computers on desks at remote locations just cause massive headaches and unsatisfied users, because they have to wait for the internet connection to become free again, after somebody has opened their email or a file on the central server...

I'm waiting for "Intel release decent graphics chip".

The software we have running our POS server at work has 256 cores.

It manages it really well by giving each active client connection it's own core. This normally lasts less than a second and is then freed up for use by someone else but it means that each client has the full speed of that particular core available to do its work.

It also means that when running an extensive report that it doesn't slow down the client connections for other users.

Not a very good example of a consumer use but we definitely benefit from a high number of cores.

I think that the main use apart from super computers, will be local processors for a number of thin clients, probably for more than 50 users off one processor. There are certainly many companies that could use that arrangement.

We have 35 users running off of 4 cores (2 physical processors) and 4GB memory. A 50 core processor could easily cope with a couple of hundred average users - the question is more, can the associated sub-systems cope with the IO (disk, memory, network etc.)?

I would have thought that memory would have coped but hard disks are generally the weak link in terms of speed. Though if using SSD that would not be a problem. Also having seen SSD transfer 1Gb in less than a second I would doubt that cabling would be an issue either. Though what might be is the network card having to deal with several hundred connections simultaneously. Though since humans are even slower even this could be eliminated with optical connections. It could make it a very cheap system for hundreds of users, and very green with only one processor, especially since most general business applications are not graphic or processor intensive.