The National Science Foundation (NSF) in the USA has been handing out some grants this week. One of them is for the University of Tennessee, to the tune of $10 million, to deploy a new visualisation system:

The University of Tennessee (UT) will receive $10 million from the National Science Foundation over four years to establish a new, state-of-the-art visualization and data analysis center aimed at interpreting the massive amounts of data produced by today’s most powerful supercomputers.

The TeraGrid eXtreme Digital Resources for Science and Engineering (XD) award will be used to fund UT’s Center for Remote Data Analysis and Visualization (RDAV), a partnership between UT, Lawrence Berkeley National Laboratory, the University of Wisconsin, the National Center for Supercomputing Applications at the University of Illinois, and Oak Ridge National Laboratory (ORNL).

Showing that clusters, in fact, are not the solution to every HPC and visualisation problem out there, the new system will be a Single System Image (SSI) box from SGI. Details at the moment are pretty sparse, but the press release does say:

Much of RDAV will rely on a new machine named Nautilus that employs the SGI shared-memory processing architecture. The machine will feature 1,024 cores, 4,096 gigabytes of memory, and 16 graphics processing units. The new SGI system can independently scale processor count, memory, and I/O to very large levels in a single system running standard Linux.

Despite the large percentage of clusters in the Top500, they’re only really at home for jobs which can be properly parallelised. Shared memory systems are still much faster at certain types of compute jobs – especially visualisation:

Shared-memory processing can be even more useful than the world’s most powerful computers for certain tasks, especially those aimed at visualization and data analysis. The system will be complemented with a 1 petabyte file system and will be fully connected to the TeraGrid, the nation’s largest computational network for open scientific research.

1024 CPUs in an SSI system probably means that this will be an SGI Altix 4700 – unless SGI are indeed pushing a new x86 Altix out the door. 16 GPUs may not sound like much, however, if (as I suspect) they’re taking about NVidia Tesla S1070s, then you’re looking at 4 teraflops of performance for each one – that’s 64 teraflops of GPU performance in the system.

Hopefully we’ll be hearing some noise from SGI in the coming days, shedding some light on the exact configuration of the system.

You can read the full press release over at HPCWire.

Davin Chan, VP of the SGI User Group, has just dropped me an email with a reminder that their annual conference is coming up. Full details are:

We will be holding our 7th annual Conference in San Antonio, TX Oct 21-23.

Some of our featured keynotes will include:

SGI Directions
Mark Barrenechea, President & CEO, SGI

Needle Amongst the HayStacks: Analytics and Mining Sharded Datasets
S. Ryan Quick, Principal Architect, EBAY/PAYPAL

Technology Trends
Dr. Eng Lim Goh, Senior Vice President & CTO, SGI

The conference is a good opportunity to interact with SGI engineers, executives
and other SGI customers. You can find more information about the conference
including a preliminary program at www.sgiug.org.

The lineup of speakers sounds impressive – Dr. Eng Lim Goh is always a good speaker, and it would also be interesting to hear Mark Barrenechea’s take on recent events.

SGI have posted details about their new workstation, the Octane III. Although SGI are raiding the ghosts of workstations past for the name, sadly this isn’t heralding a return to MIPS and IRIX goodness.

In fact the Octane III seems to be going for the market created by the Cray CX-1, and the much missed machines from Orion – personal supercomputers.

As with all things from SGI, the first thing to look at is the case design.

Oh dear. Why drag out the good Octane name if you’re going to release a dull grey box? The Cray CX-1 looks like an extra from a sci-fi film – this just looks like a Dell. And – I’m sorry guys – but the new SGI logo makes the baby Jesus cry. It’s just terrible. (Brand New have an excellent deconstruction of the new logo, well worth a read)

Look at the awesome Cray CX-1:

Or SiCortex’s fantastic SC072:

Or the original Silicon Graphics Octane (mine’s still in use, they’re awesome bits of kit):

Inside the case, though, things start to sound a bit better. The machine is available in three different configurations, all of which can run Red Hat or SUSE Linux (which SGI’s excellent ProPack enhancements) or Windows HPC Server 2008:

• 19 Dual-Core Single Socket systems, giving 38 cores and 76GB RAM
• Ten dual-socket quad-core Xeons processor boards, giving 80 cores and 960GB RAM
• A single dual-socket Quad-Core Xeon processor board with 144GB Ram and 7 PCI-Express slots (2 x16, 4 x8, and a x4 for the RAID card)

The last option, the “graphics workstation”, is particularly underwhelming. The motherboard is mounted vertically, so the system essentially becomes the same as any stock, high end desktop.

Worse still, it seems all that NUMA goodness that made Origin 200s, 300s, and the rackmount Tezro such kick arse workstations is missing – sure, you can have 960GB RAM in an Octane III, but it’s not global shared memory. 10 processor boards, each with 96GB of local memory.

It’s a cluster in a box, using GigE or Infiniband (DDR or QDR) as the interconnect. And there’s nothing inherently wrong with that, except in a configuration like this it gives you zero flexibility. Unless you’ve got a workload that parallelises nicely, you’ll be running into some walls.

The prices start just under $8k for a single Nehalem node with 8 cores, and rises to $53k for 10 nodes, with 80 cores and 240GB of memory (24GB per node). Raid the piggy bank if you want to max out the RAM or have QDR Infiniband as the interconnect.

And here lies the big problem with this system – there’s too much overlap with the recently announced Cloudrack X2 systems. Sure, those require actual racks, and cooling, whereas the Octane III can run from a single 240V socket. (And it’s “whisper quiet”, apparently. Although to be fair, even large jet airliners are “whisper quiet” compared to the original Octane in fastfans mode).

Apart from being able to stick it on your desk, the Octane III just doesn’t seem to have much going for it, compared to the Cloudrack X2 or a decent high-end graphics workstation. The Cray CX-1 makes sense – Cray don’t really do ‘small’ supers, so having a deskside system is a good play – it’s a stepping stone to their bigger systems.

To make this work, SGI will have to sell in volume, and via it’s channel and partners. Can they duplicate Cray’s success with the CX-1? I’m not sure, especially as it’s not immediately obvious who this machine will appeal to. Really, I’d love to see SGI ship some sexy and powerful development and graphics workstations or baby supers, but the Octane III just seems too weak in too many areas.

You can grab full specs and datasheets over at SGI’s Octane III page.

SGI have just posted some pretty decent SPEC benchmarks, and it’s clear that they’re aiming straight for the datacentre, calling out IBM, HP and Sun’s big SSI (Single System Image) machines. The benchmarks posted are for the Altix 4700, and the cynic in me says this is a bit of “rah rah rah” flag waving exercise before we see a Xeon version rolled out.

From the press release:

* SGI is more than five times faster than its next closest SSI competitor in the SPECfp_rate_base2006 floating point performance benchmark. Altix 4700 proved 5.7 times faster than the Sun SPARC Enterprise 9000, its next closest SSI competitor; IBM and HP products trailed further behind.

* SGI performed over four times better in the SPECint_rate_base2006 integer performance benchmark. Altix 4700 proved 4.3 times faster than the Sun product, its next closest SSI competitor.

* SGI outperforms IBM by almost three times in Java performance with SPECjbb2005 benchmark. Altix 4700 is the leader in Java performance as measured by SPECjbb2005, outperforming IBM by 2.8 times.

* SGI has five times higher aggregate memory bandwidth than its next closest competitors. Altix 4700 has the highest aggregate memory bandwidth in the world, five times higher than its next closest competitors, NEC and IBM.

The arguments over how unrealistic and artificial the SPEC benchmarks are will rage forever, but these are some pretty impressive numbers and some pretty bold claims from SGI. You can read the full press release here.

What is SGI CEO Mark Barrenechea not telling us?