SGI has upgraded its HPC blade server lineup with the latest x86 silicon and a turbo-charged InfiniBand network. The Altix ICE 8400 is the successor to the company’s 8200 series and is designed as a premier solution for the HPC cluster market, scaling as high as 64,000 nodes.

The 8400 represents the fourth generation of the ICE product line, which was begun in 2007, although this iteration is probably the most significant upgrade to the system in its three-year run. Besides moving up to the latest Intel Westmere EP chips (Xeon 5600), for the first time SGI is adding an AMD Opteron option as well. Prior to the company’s merge with Rackable last year, SGI was an Intel-only vendor.

Both x86 blades come in a dual-socket setup and use standard chipsets. The Intel blades feature quad- or six-core Xeon 5600 processors, although customers can opt for the previous generation Xeon 5500 silicon as well. Using the six-core Xeon, up to 768 cores can be stuffed into one cabinet.

The most notable feature on the Xeon side is that SGI has designed the blades to handle the fastest and hottest SKUs in Intel’s arsenal, in other words, the 130 watt parts. To date, all other HPC server offerings have shied away from these top-end Xeon chips since the extra heat produced limits the density of the designs, especially for the closely-packed blade architectures. The standard high-end processors for x86 blades are the 95 watt parts.

“We realize there are some competitors with higher density,” admits Paul Kinyon, senior product marketing manager at SGI. “But what we’ve duly noted is that there is no free lunch.” From their perspective, the HPC market is very sensitive to application licensing costs, and just increasing the CPU server density at the expense of clock speed can end up costing more in software than might have been saved in hardware consolidation.

Speaking of density, the AMD option for the 8400 supports the new 8- and 12-core Opteron 6100 chips (Magny-Cours), which makes a 1,536-core cabinet doable. And since the new Opteron blade supports up to 16 DIMM slots (as opposed to 12 DIMMs on the Xeon blade), there’s more memory to go around as well.

Interestingly, SGI allows you to mix Xeon and Opteron blades in the same cabinet, and run them under the same system manager. A more likely configuration would be to keep the Xeons and Opterons confined to separate racks, using a job scheduler to push specific apps onto the different blades. The rationale is that Intel chips are more suitable to codes needing fewer faster cores, with the AMD chips offering the advantage in memory bandwidth and core count. According to Kinyon, they’ve seen “a fair amount of interest” from customers who are considering a mixed-vendor x86 cluster.

Customers who want to give this x86 odd couple scenario a whirl will have to wait until later in the year, though. While the Intel blades are available now, the AMD hardware won’t be shipping until Q3. From a pure blade perspective (sans CPUs), Kinyon says the AMD and Intel models are similarly priced. Once you add in the CPUs and memory, prices will almost certainly vary. While the Opteron CPUs tend to be less expensive than their Xeon counterparts, if additional memory is desired to support the extra Opteron cores, costs may even out.

Specialized service nodes, which appear as peers to the x86 nodes, can also be integrated into an 8400 cluster. These include shared memory UV10 and NVIDIA Tesla GPU nodes. The shared memory node option, in particular, seems to be gaining traction as an add-on for HPC distributed memory machines, and Kinyon says they’ve already bid this configuration on some recent RFPs.

CPUs and GPUs aside, the bigger story for the new 8400 is what SGI has done with the interconnect. Here they’ve decided to push InfiniBand about as far as it will go. The 8400EX version, in particular, is optimized for maximum interconnect performance. It uses a dual plane network and four integrated QDR InfiniBand switches per enclosure. For better price-performance, the 8400LX offers a single plane network and cuts the InfiniBand switches to two per enclosure.

SGI touts the 8400EX as tops in the industry for MPI performance, delivering a three-fold increase in bandwidth per node versus the competition. The company is claiming a world record result (51.3) for the 8400 on the SPECmpiL_2007 benchmark. Although more pricey, the dual plane design gives customers the option to either use the extra bandwidth as a single fat pipe, employ one of the planes for redundancy for MPI traffic, or dedicate one plane to MPI traffic and the other to I/O.

Multiple network topologies are offered, including hypercube, enhanced hypercube, all-to-all and fat tree topology. Except for the for all-to-all, the other topologies were available in the 8200 product, allowing customers to easily add to their legacy ICE systems by extending the same fabric.

SGI designed the new all-to-all topology to deliver maximum bandwidth (up to 12,000 MB/sec per node) and lowest latency, although this option only scales to 128 nodes. The enhanced hypercube — available in the 8200, but juiced up for the 8400 — is next in bandwidth performance and scales all the way up to tens of thousands of nodes. The fat tree topology is the highest in cost, requiring external switches, but enables all-to-all MPI communication at scale. The hypercube is the lowest cost, but the least performant of the bunch.

According to SGI, there are already some orders on the books for the new Altix. One early deployment is at NASA Ames, where the Pleiades supercomputer just added 32 racks of 8400 hardware, boosting its performance to just shy of a petaflop (973 teraflops).

As seems to be the current tradition in selling high-end servers, SGI is not talking pricing on the 8400. Kinyon says they were very careful when designing the new Altix to make sure that they didn’t price themselves out of the value end of the x86 cluster-based market. At the same time, they wanted to offer a solution for users “on the hairy edge of HPC.” The company believes they’ve struck the right balance of price and performance with the 8400. Says Kinyon: “We’re just jumping up and down and waiting to hear the competition whimper.”