Stephen Wheat, senior director of HPC at Intel, gave a presentation at last month’s HPCC Conference in Newport that had a lot of people talking. Wheat started with an overview of where the world’s largest chip manufacturer is headed and the challenges they’ll face getting there. But what sparked conversation were his closing comments on the state of the HPC industry and the risks for the future of HPC as a separate market.
The problem is one of identity. “Fewer and fewer HPC deployment activities are bearing any portion of the research and development costs of the market,” Wheat said in a recent conversation. If high performance computing wants to continue to be a distinguishable market space, it needs its own research and development activities.
There are two forces working against robust R&D in HPC. First, investment in research for high end computation has remained lackluster even while industry groups and a series of blue ribbon government panels have repeatedly identified the need for increased funding and coordination. Without external funding for computation-centric R&D, vendors in HPC have been left to fund innovation on business margins alone.
By itself, this change isn’t obviously bad. Lots of industries run their R&D off the business, right?
Except that HPC has also experienced a dramatic shift in the economics of its products and the dynamics of its customers. Over the past 10 years the ability to deploy a large number of FLOPS for an ever-decreasing price has created a FLOPS arms race. “In practice, the success of an installation is rarely measured by the performance of applications,” Wheat says, “but by the Linpack performance of the solution.”
Customers are often eager to be “on the list” or, lacking better tools and a vocabulary to talk about the effectiveness of a deployment in any other terms, simply focus on the one measure they can specify. Vendors who don’t respond to this customer focus lose procurements, so prices get pushed down and designs are increasingly based on commodity technologies. One industry insider I talked to said that margins have fallen by as much as 300 percent on HPC hardware as commodity technologies have pushed farther into HPC.
You might be inclined to observe that prices wouldn’t have fallen if companies hadn’t responded the way they did, and you’d be right. And the pure free marketeers among you would rightly observe that markets get what they are willing to pay for; if there isn’t funding for R&D, then just don’t do any. The problem is that supercomputing isn’t just any technology.
If consumers aren’t willing to pay for the R&D needed to put that elusive sixth blade on the new Gillette shavers, everyone will get on just fine with their five-bladed wonders. But, as has been pointed out many times, supercomputing is a key enabling technology for much of the basic science that we are counting on to advance our society, from drugs to implants to cleaner sources of energy. As Dan Reed, now director of Scalable and Multicore Computing Strategy at Microsoft, pointed out in his testimony before Congress several years ago, high performance computing is a “universal intellectual amplifier.” The problems that HPC faces — software issues with million of threads, multicore tools and compilers, memory and I/O bandwidth, etc. — need to be addressed in order for the other disciplines that depend upon computation to continue to thrive.
Dave Parry, senior vice president and product general manager at SGI, points out that one of the problems companies face in evaluating big procurements is in thinking of them as point events, and looking only at the costs of the specific systems offered. This excludes all the other costs a company needs to cover to keep the lights on and provide iced lattes to overworked engineers. A company might want to do this with one or two “marquee” procurements to claim a place at the top of the market. But when a large portion of a company’s customers start demanding the same pricing, the balance sheets start tipping in the wrong direction.
Perhaps a more troubling aspect of the very aggressive pricing and thin margin on these deals is that, while larger companies might experience some pain, smaller companies are completely priced out of the market. Joe Landman, founder and CEO of Michigan-based HPC company Scalable Informatics, says that “if a small company’s revenues fall off as they wait for payment from a major deal, [and] they couldn’t afford to spend time and effort working on smaller ones which could keep the revenue coming in, the small company will likely be forced into some very hard choices.” Or just stay out of the market altogether, limiting choice and, again, stifling innovation.
Why does all this matter now, and why is Intel talking about it? According to Stephen Wheat, Intel is embarking on a new push with its HPC businesses. The company now sees an opportunity for our community to partner with them to move chip technologies forward in a way that specifically benefits high-end and parallel computing. Intel sees itself as ready to talk about HPC R&D at a time when the R&D in our community is at a low.
Of course, many of the vendors are quick to point out that they are still innovating. Kevin Noreen of Dell, for example, points to the innovations that Dell is doing with the datacenter as a whole. But these innovations are farther up the value chain, and can be leveraged into more lucrative markets. While important, these steps don’t address the fundamental limitations HPC users are facing because of inadequacies inherent in our current crop of technology — both hardware and software.
So what about possible solutions? Wheat had several ideas to get the ball rolling.
One of his suggestions was to move away from “winner-take-all” procurements. In typical procurements, no matter how many vendors bid, only one wins. Without a chance for cost recovery, the losing parties then lose real money: money for the benchmarking, proposal writing, business case analysis, and pricing that goes into what can be a 12- or 18-month process to prepare a bid for a large procurement. Dave Parry of SGI says that vendor costs to submit a proposal for a larger procurement can top $1 million.
Another suggestion made by Wheat is to move away from long-range procurements, with very detailed technical specifications in the out years, and strict penalties for missing performance or milestones. The acquisition time in many HPC procurements now exceeds the time it takes the major chip vendors to manufacture a new generation of chips, and this means that vendors are proposing very specific performance parameters for gear that, in many cases, doesn’t exist when the bid is submitted. Several of the people I talked with recommended a series of more frequent, shorter acquisitions that would enable vendors to assess and manage risk more effectively, and also to be more responsive to changes in the marketplace.
Another approach? Explicitly fund proposal and non-recoverable engineering costs for participants in an acquisition. This could work with a lightweight proposal upfront, used to select a few strong candidates, and then proceed to a more in-depth proposal round. Vendors making it to this second round would be treated as consultants and reimbursed for 1) benchmarking, 2) non-recurring engineering costs required to put together a plan for a large system, and 3) the facilities planning at the customer site to support the proposed system. This would obviously only make sense for very large acquisitions — perhaps a top 15 system. But while new to us, the model is not that unusual in other sectors, such as the enterprise computing space.
Many I spoke to, including Dave Parry of SGI and Ed Turkel of HP, stressed the need for collaborative acquisitions that build teams with suppliers and customers, and reward performance more than they punish failure. HP’s Turkel identified the ASCI Q procurement as a solid model of collaboration. In this case, the DOE customers introduced HP to Quadrics, and all three proceeded together to develop a solution.
These suggestions all aim to increase rewards or decrease the risk of the acquisition process, but this tweaking could put customers, especially federal customers spending taxpayer money, in a difficult position. Purchasers have a fiduciary responsibility to their organizations to ensure that vendors deliver what was agreed and paid for. Relaxing terms and conditions, eliminating penalties, and paying vendors to submit bids so that customers and federal agencies can then pay more money for actual products can be a tough sell when everyone has already demonstrated willingness to participate without these incentives.
Dan Reed suggests that a different approach will be required to address the fundamental problem (lack of support for R&D in HPC) directly rather than by using the acquisition process as a surrogate. According to him, what we need are “integrated R&D programs with a long term strategy,” not better acquisition models.
Ed Turkel, manager of product & technology marketing for HP’s High Performance Computing Division, echoes that point of view and stresses that R&D funding for HPC can still result in technologies that are commercially viable. “There are still opportunities for innovation,” he says, “even though the industry has gained tremendously from the volume economics of commodity servers.” These opportunities arise through specializing general purpose technology, for example, or “tweaking” it to better suit the needs of HPC without degrading performance for everyone else.
There is ample supporting documentation already lying around to bolster the view that HPC needs more R&D. All that is needed is leadership to take action on the recommendations of the HEC RTF and PITAC reports, among many others, and coordinate HPC activities at the federal level. Of course it is key that this coordination not focus on acquisition activities, but on R&D. There is already quite a bit of activity at the federal level in which agencies participate in each others’ procurement processes and trade best practices for acquisition. This won’t broaden the vitality of the HPC market. As Reed observes, “You can’t build a national strategy over a set of point acquisitions.”
Those I spoke with were keen that any new federal support for R&D not focus on a very narrow portfolio of small or large projects. The consensus is that we need more than one or two other HPCS-style activities, and that a host of new very small activities won’t be sufficient either. A balanced portfolio, coordinated across the agencies, of small, medium, and a few large activities involving industry and academia is most likely to result in technologies that not only meet the needs of the HPC community, but also can fit into the dynamics of the commodity hardware market.
We are now poised for ‘good enough’ in parallel programming as well, as the millions of commodity software developers begin to fund research and cooperative efforts on multicore software development tools to satisfy their requirements. Like the recently announced Microsoft/Intel joint effort, these future efforts will likely focus on cores in a socket, resulting in substantial improvements for millions of desktop customers, but less so for high-end technical computing users.
Of course, most of those in the HPC business have indicated a willingness to make an HPC market without any incentives at all. And at least those that don’t go out of business can probably be counted upon to continue to do this for the foreseeable future.
The risk here is not that the HPC market will experience a catastrophic collapse. Indeed, if it did, there would likely be consensus on concrete steps we could take immediately to revive it, and the political will to act. The real danger is that the HPC market will simply cease to exist, having evaporated slowly, one company at a time. HPC consumers have already proven that they aren’t willing to pay for custom, highly differentiated hardware tailored specifically to their needs. As Joe Landman puts it, “We are in the era of ‘good enough’ where we look at the price knee which maximizes performance for a minimum price.” Until the nation’s innovation stewards chart a different path, maybe that’s all we can hope for.