Market analysts are predicting big things for HPC. As growth in the rest of the server industry flattens out, they see HPC continuing to expand. Most of this growth is at the bottom, and the vision is one of many new customers stepping up to buy clusters of 16, 32, or 64 nodes costing less than $50,000.

It's a believable vision, but one with serious obstacles to overcome. HPC is hard -- inexcusably so. In a blog post several weeks ago, Tabor Research analyst Chris Willard passed on one of the many definitions of the term "supercomputer" he has heard over the years: "Any computer where the compute time is worth more than the programmer time." This is a pervasive definition that has worked its way deep into our thinking about HPC. It is also totally wrong-headed.

Another obstacle to broad entry-level adoption of HPC is change: the technologies of HPC evolve faster than most of the vendors of the software tools that are driving low-end HPC adoption can keep up with. Cost is another problem; although costs of hardware are falling and have fallen dramatically, they are still out of the reach of much of the small enterprises that could make use of HPC, and the entry path often isn't a smooth slope. One is either using a $5,000 workstation or a $100,000 cluster. Until recently, there hasn't been much in the middle, and the offerings are still pretty thin.

Finally, the very people who need to buy the computers to fulfill the visions of dramatic HPC growth do not identify with HPC, or supercomputing, as solutions to their problems. They are a "dark market." So HPC has to find them, in small corners of very diverse markets, and find a way to sell to them that will cause the vast new market at the bottom of the HPC customer pyramid to finally take shape and start making actual purchases.

Acceleware's strategy is to to bring HPC to a diverse set of customers, one vertical at a time. I recently spoke with Ryan Schneider, Acceleware's CTO, and Robert Miller, the vice president of marketing and product management, about the company, its technology, and its approach to the market.

Acceleware's technology is software deployed on top of NVIDIA's GPUs and HP's servers. Although the company doesn't manufacture the hardware, it does intense integration with the result that it owns the entire solution delivered to a customer, from silicon to application software. The company works with ISVs in a vertical to modify the key software tools in that domain in order to take advantage of GPU acceleration using Acceleware's middleware, the "Acceleware Technology Platform." Right now Acceleware has offerings in virtual prototyping, imaging, and the oil and gas industries.

There may be a lot of code work done under the hood, but the users continue to work with their favorite software using their familiar UIs and workflows without disruption. The only change users will see is something akin to a checkbox asking whether they want to run their simulation with acceleration, or not. That's it.

What's the payoff? For users it can be tremendous. They buy hardware in a familiar workstation form factor or as one of the smaller C30 clusters, running the same OS and applications they've already run. The unit comes ready to plug in, totally configured with their domain application software licensed and ready to run. And they get performance advantages that can be significant. For example, cell phone manufacturers using Acceleware's kit have seen turnaround times decrease from 10 hours per simulation run -- in an analysis suite that typically requires 400 runs -- to 15 minutes. Medical providers using Acceleware's imaging kit for reconstruction of medical image data have seen their run times shrink from a month to a few days, and a large drug company saw their imaging workflow shrink from a week to a few hours.

What's the motivation for ISVs to work with Acceleware? Ryan Schneider, Acceleware's CTO, described it as a build-versus-buy value proposition for these companies. For the most part, the software companies Acceleware works with have limited resources that are focused very tightly in their application domain writing the software their users want. Porting to Acceleware's middleware allows companies to insulate themselves from technology change risk. As new technologies replace or augment GPU acceleration as viable alternatives in technical computing, Acceleware simply adds support for the new hardware underneath the middleware layer.

The relationship with the software vendors gives Acceleware direct access to what would otherwise be a dark market at the low end. Acceleware sells directly into the verticals they are targeting, augmenting channel sales driven by the ISVs.

What's the motivation for customers? That one is easy, and I've already touched on it. For a nominal expense -- anywhere from $10,000 for a single acceleration card to $250,000 for a four node, 16 GPU cluster, including hardware and application licensing costs -- users get access to what is, in some cases, orders of magnitude better performance without having to change their workflow. This puts Acceleware's products square in the middle of production design and operational computing for some very serious companies. Firms like Boston Scientific, Philips, Nokia, Samsung, Boeing, Eli Lilly, and Hitachi are using Acceleware's equipment.

Speaking of equipment, what does Acceleware offer? All of the company's hardware supports the same performance middleware, so any application ported to one of the company's products runs on them all. At the very low end, users can buy a single GPU card, the A30, with the application and licensing for around $10,000. The D30 is a dual card configuration installed in a workstation with software for roughly $25,000, and there is a quad card configuration with workstation for roughly $65,000.

The company also recently announced a clustered solution. The C30 series of hardware clusters 4, 8, 12, or 16 nodes together to form a GPU cluster. Each node has two dual-core Intel processors connected to four GPUs; the nodes are interconnected with InfiniBand. MPI is used for communication between the nodes on the coarse-grained problem decomposition, and each node then parcels out work to its GPUs as appropriate for the problem. The company claims that the C30-16, with 64 GPUs, is capable of hitting 32 TFLOPS peak single-precision performance. I asked about the single-precision part; evidently in the verticals the company is currently operating in, that's A-OK. Seismic acquisition data, for example, starts life as less than 32-bit precision anyway. The C30-4 node configuration is priced around $250,000.

Inside the Acceleware's middelware platform are a host of common and industry-specific algorithms that have been tuned for optimal performance on NVIDIA's GPUs. In the library you'll find optimized FDTD algorithms, Feldkamp image reconstruction, Kirchhoff algorithms for the oil and gas industry, and more general purpose matrix math libraries.

Of course there is competition. Companies like RapidMind are building software akin to Acceleware's middleware for abstracting the hardware and allowing software to be retargeted at FPGAs, Cell processors, GPUS, and so on, and this will be a good solution for many users outside the verticals that Acceleware is operating in. But for sheer ease of use, Acceleware seems to offer a product that is uniquely focused on making the transition from ordinary computing to supercomputing a seamless one.