NVIDIA’s GPU computing ambitions got a major boost today with IBM’s announcement of the iDataPlex dx360 M3. The new HPC server pairs two Tesla GPUs with two CPUs inside the same server chassis. As such, IBM represents the first Tier 1 server vendor to bring CPU-GPU “hybrid” computing to the high performance computing market.

“This is the first time we’re in a mainstream server,” says NVIDIA’s Sumit Gupta, senior product manager for the Tesla GPU computing group. Last week, Appro, Supermicro, AMAX and Tyan announced integrated CPU-GPU server gear based on NVIDIA’s new Fermi architecture Tesla 20-series devices. What IBM provides is a broad global sales channel and unmatched brand recognition.

All these systems, including the new iDataPlex from IBM, make use of the latest Tesla M2050 computing modules that can be integrated into a CPU-based host system. Each M2050 delivers 515 gigaflops of raw double precision floating point performance (or 1,030 gigaflops single precision), and comes with 3 GB of GDDR5 memory. IBM customers can also opt for the M2070, which offers the same floating point performance, but with 6 GB of local GPU memory.

The base configuration on the new iDataPlex consists of a two-socket motherboard with the latest Intel Xeon CPUs. A riser card is used to hook in the Tesla modules. The configuration allows for relatively easy maintenance and replacement of the GPU components.

IBM’s move into the GPU computing space is a big win for NVIDIA and for GPU acceptance in HPC, in general. Over the past couple of years, the company had remained very quiet on the GPU computing front, and there were no indications it would be adding this capability to its HPC lineup. “I think what’s changed is that customers have been experimenting for a long time and now they’re getting ready to buy,” says Dave Turek, vice president of the deep computing group at IBM. “It’s as simple as that.”

According to Turek, IBM has been tracking customer demand for this capability for some time, and felt now was the time to jump onto the GPU computing train. From Turek’s point of view, this is less about the extra capabilities provided by NVIDIA’s new Fermi architecture (ECC memory, double precision, programmability) and more about the general increase in customer acceptance of the GPU computing paradigm. “If the marketplace hadn’t been ready at this time, we would have bypassed this for sure,” he admits. “It wasn’t the technology that drove us to do this. It was the maturation of the marketplace and the attitude toward using this technology.”

The company expects the new GPU-equipped iDataPlex to get the most traction in what have become the early adopter segments for GPU accelerated computing, namely the oil and gas industry, big science research at government labs and universities, and the biotech space (with perhaps some uptake by financial institutions). All of those segments have a few things in common that makes them an especially attractive target for GPU acceleration: a nearly endless need for more vector math capability, in-house programming expertise to push their apps over the GPU programming hurdle, and a limited dependency on ISVs who may or may not be interested in GPU support.

IBM’s decision to pursue the HPC market with a CPU-GPU offering is particularly relevant in another sense. Over the past couple of years, the company had pinned much of its hybrid supercomputing hopes on its own HPC variant of the Cell processor: the PowerXCell 8i. That processor was used to power the Roadrunner supercomputer, the first general-purpose computing system to break the Linpack petaflop barrier back in 2008. IBM still offers the Cell-based QS22 blades based on the PowerXCell 8i, but has halted plans to forge a successor to that chip design.

In fact, from IBM’s point of view, the GPU-equipped iDataPlex is just another entry in its rather large portfolio of HPC hardware. Between the new Power7-based 755 servers, the Blue Gene/P, and its x86-based iDataPlex gear, IBM has probably the broadest HPC offerings in the industry. The hybrid computing iDataPlex is another way the company thinks it can cover what has become a fairly diverse HPC market.

Turek says IBM will be careful not to overhype its new GPU-accelerated boxes. Although coprocessor acceleration seems to be in vogue right now, not every application is going to be able to take advantage of it. Certainly most matrix math-intensive apps will be able to realize a several-fold performance boost compared to a CPU-only implementation, but it really depends on how much of the code is engaged in these types operations and how much is just doing sequential threading.

If Linpack is a guide — and that’s really all it is — some apps will do very well indeed on the new Fermi GPUs. NVIDIA ran some benchmarks on its own CPU-GPU server, consisting of two Tesla C2050 cards (comparable to two M2050s) plus two Intel Xeon X5550 processors, with 48 GB memory. They found Linpack performance was eight times that of a comparable CPU-only server: 80.1 gigaflops for the CPU version versus 656.1 for the GPU-accelerated box. When they looked at price-performance and power usage, they found a five-fold advantage. So for $1 million worth of CPUs, you can get 10 teraflops of Linpack, while that same money spent on GPU-CPU gear will get you to 50 teraflops — and a certain spot on the TOP500 if you’re interested in HPC celebrity.

With IBM now in the GPU computing game, it’s almost a sure bet HP and Dell won’t be far behind. And with the tier 1 OEMs onboard, integrated CPU-GPU servers are likely to become standard operating equipment by most, if not all, HPC vendors over the next several months.