During a talk for the Ken Kennedy Institute’s 2022 Energy High Performance Computing Conference, Dan Stanzione, executive director of the Texas Advanced Computing Center (TACC), gave a status update on TACC’s forthcoming Leadership Class Computing Facility (LCCF)—a massive NSF-funded expansion of its supercomputing capabilities that will launch with a new flagship supercomputer, “Horizon.”

“We have a lot of computers; they’re kinda large,” Stanzione explained, highlighting systems like Frontera (23.5 Linpack petaflops) and Stampede2 (10.6 Linpack petaflops) that rank highly on the Top500 and which support thousands of project teams. “Collectively, we have about 20,000 machines, we have over a million cores, we have a thousand GPUs, we deliver about seven billion core hours a year to our user community.” And it’s adding more all the time: Stanzione said that the graphics were just added to the new Lonestar-6 system the day before the talk.

All of this computing firepower supports projects across the NSF’s domains. “Generally, if it’s unclassified and it’s open science and it’s at an academic institution we probably support it at this point,” Stanzione said. And the demand for more remains: “[At] the beginning of the center, we had had requests for about five times the computation we produced. We have about 80,000 times that much computing available now… and we get about five times the requests for the amount of time that we can produce.”

“So apparently,” he reasoned, “demand for computing is invariant and not in any way dependent on the size of the computer you buy. So obviously, you should buy a bigger one.”

Taking a leadership(-class) role

Enter the LCCF and Horizon. Back in mid-2018, TACC won an NSF award to create Frontera (which, Stanzione said, is “technically phase one of the LCCF”), laying the groundwork for a longer-term leadership computing strategy. By the following year, word had gotten out that TACC was planning to follow up Frontera with a computer 10× as powerful around 2024. In 2020, TACC started talking more about these plans, calling the facility the Leadership-Class Computing Facility, shifting its target slightly to 2025 and releasing concept art for an expanded datacenter.

The core pitch for the LCCF, Stanzione said, is “about coming up with a more sustainable way to invest in computing than one-off system competitions every four years.” Several times, Stanzione made reference to the National Center for Atmospheric Research (NCAR), an NSF-funded computing center focused on weather and climate research. TACC, he said, wants “to be an anchor of the NSF computing environment in the way that NCAR is for atmospheric research.”

Stanzione explained that the plan for the LCCF is to “actually start construction—fitting out a datacenter— …two years from today” (March 1, 2024), pending budgetary approval from Congress. Then, they’ll aim to deploy the LCCF’s flagship system in the first half of 2025, with science users on the system by the back half of 2025. After that: ten years of support for the center, from 2026 to 2036, with “maybe an upgrade in there,” Stanzione said.

“We’ve never had more than four years of funding on an NSF system before as the initial commitment,” he stressed, citing NSF rules around the duration and renewal of its grants. “The reason I say we wanted to build an NCAR-like facility is because it’s been sitting there funded [by the NSF] in Colorado since 1957, so… perhaps my real challenge is figuring out, personally, how I was going to get around that limit that technically exists.”

The flagship system will, Stanzione said, be named Horizon, “if DOE doesn’t steal that name and I don’t have to change it again when they put out a system.” (Regarding the LCCF itself: “I’m gonna give it a better name and a better logo before it’s over, but that comes later in the funding cycle.”)

On the Horizon

“You probably all wanted me to say what machine [we are] picking,” Stanzione said, “and I’m not going to, because honestly, I haven’t made a final decision, because the best way to be wrong is to tell people four years in advance what your computer’s gonna look like and what it’s gonna do. … Two years before we deploy, ask me what it’s going to be.”

That said, Stanzione did go into some detail on the process of setting goals and attaining support for Horizon.

“[Frontera] is getting old very quickly, so we need a follow-on, and the only written instructions are ‘10× faster!’” he said. “And we’ve had I can’t tell you how many hours of conversations over what 10× actually means in that context—is that application throughput? Is it peak application performance? Does it have anything to do with flops?” (“I’ve argued no,” he said.)

He then broke down where they expect to get the 10× from.

“10× in that time frame is challenging for a couple reasons,” he said. “One: if we look at the Frontera baseline, if we just did nothing but rely on vendor performance improvement … at best, over this five-year timeframe, maybe we’ll get 3× out of that,” mostly from increases in memory bandwidth. “But three is not 10.”

Then, he said: “Buy more — that one almost always works … So we’re gonna double the budget over what Frontera was, and I got away with that, so.” That brings it up to 6×, leaving Horizon with a need to speed up by two-thirds. That remainder, Stanzione said, will be accomplished with improvements in software algorithms and methods.

Bringing things up to code

To that end, last spring, TACC solicited the research community to submit problems that they thought were core for future research problems and representative of problem designs in the supercomputing space. They received 140 proposals, whittling that down to 30, then selecting 21 to go forward for deep examination, looking to achieve significant speedups “and see where we can get.” The projects, which are funded to the tune of around $300,000 each, span mathematics, physical sciences, engineering, geosciences, life sciences and even the social sciences.

“If all the codes were well-engineered and really good, I’d worry a lot about our ability to make improvements and hit the numbers,” Stanzione said, continuing: “All codes are not really good, well-clustered, well-engineered codes. Most of them are what we’d call ‘software.’”

This code research, he added, would also help TACC sell the benefits of the LCCF and Horizon to decision-makers. “We know if we build a bigger machine, cool stuff’s gonna happen,” he said. “But I can’t sell hundreds of millions of dollars of investment on ‘cool stuff is gonna happen.’”

Now, back to the hardware

On the hardware front, Stanzione showed a laundry list of vendors under evaluation. “We’ve done a huge number of on-site evaluations, we’ve done others with partners,” he said. “We’ve looked at a wide variety of processor technologies … we’ve looked at various and sundry Arms, we’ve looked at NextSilicon … we’ve looked at a bunch of networking options … we’ve looked at a lot of the other, more exotic things with partner sites.”

A non-exhaustive list of hardware under evaluation by TACC for Horizon

• Processors: AMD; Fujitsu Arm; Intel; NEC; NextSilicon; Nvidia
• Networking: Cornelis; Nvidia; Rockport
• Filesystems: BeeGFS; DAOS; VAST
• Node disaggregation: GigaIO; Liqid
• AI/quantum: Quantum (via Stanford); Graphcore (via Argonne National Laboratory); Cerebras (also via Argonne); SambaNova (also via Argonne)

(“Argonne buys all these [exotic] chips,” Stanzione said, “so I just call Rick Stevens [associate laboratory director at Argonne] and say ‘how’s it going?’ rather than trying to emulate what they’re doing on that.”)

Stanzione also said that they plan to add about 10 percent capacity to the core system to ensure that there is room for smaller testbeds and research projects when the main system is occupied. “We will define the size of the system,” he said, “and then I will lie about it, because we’ll buy more than that. … We’re gonna have this piece of the system that does the 10× piece and then we’re gonna have a bunch of additional racks to deal with all these other use cases that don’t count towards the capabilities.”

Confounding these efforts, Stanzione said, an “unnamed program officer” had asked him: “what is the peak flops of the entire system if you include all those pieces?”

“That’s the question you weren’t supposed to ever ask!” Stanzione exclaimed. “And now I have to make up an answer, because no one knows what the peak flops in 2025 are of any of these processors, let alone the achievable ones. So I wrote an answer and sent it back.”

The hardware will be supported by an additional 15MW of power capacity, making TACC a 25MW facility. And, regarding that “maybe an upgrade in there” from earlier: “The idea is basically we’d buy a second system halfway through that [2026-2036] life,” Stanzione said, “so it’s two five-year lifespans.” He added that they would prefer to avoid a radical change in architecture between the two systems.

For now, though, Stanzione and TACC are embroiled in what he said was the “terrifying amount of project management” required to begin building the new facility in two years.