MHPCC cut Makau's ribbon in August, months after the machine began carrying pioneer workloads. A system 13 times its size arrives at the same center this year.
On August 7, the Maui High Performance Computing Center cut the ribbon on Makau, an HPE Cray XD2000 rated at 2.64 peak petaflops. TOP500's June 2026 list had put the entry level at 2.66 Pflop/s on the Linpack benchmark.
The ceremony was not the machine's first day of work. MHPCC's own notices name Makau among the systems that "may be running reduced workload" during a storm on March 11, and describe it two weeks later as "back online and available for pioneer utilization" after its filesystem was wiped and rebuilt. Pioneer utilization is HPCMP's term for early access before a system enters general production. The machine had been carrying user jobs for roughly five months before anyone cut a ribbon in front of it.
Makau's 32 nodes divide into 4 login, 16 AI/ML, 10 mixed, and 2 transfer. Every node carries two AMD EPYC 9534 processors, 64 cores each at 2.45 GHz, for 128 cores per node. The 16 AI/ML nodes hold four NVIDIA H100 SXM5 GPUs apiece with 80 GB of HBM3; the 10 mixed nodes hold one H100 NVL each with 94 GB. Seventy-four GPUs, 3,328 compute cores, 4,096 counting login and transfer. Six cabinets, liquid-cooled rear-door heat exchangers, RHEL 9 and Slurm. HPCMP's systems summary publishes an end-of-life date of November 1, 2030.
Storage is worth stating carefully, because MHPCC's own documentation is inconsistent about it. Makau's Lustre filesystems are NVMe-backed per its user guide: 1 PB of scratch, 51 TB of home, 100 TB for projects, plus 3.84 TB of local NVMe on each of the 26 compute nodes. Beyond those, the system mounts a 753 TB center-wide NFS filesystem shared with MHPCC's other machines, and the program's tape archive. The Makau guide lists that shared filesystem as NVMe; MHPCC's Coral guide lists the same 753 TB volume as SAS. Coral's claim is more specific, since every row of the Makau guide's storage-type column reads identically.
The 2.64 figure is an Rpeak, the hardware's theoretical ceiling, while 2.66 is an Rmax, the score the 500th-place machine posted on a Linpack run. Rmax cannot exceed Rpeak, so Makau's peak sits below the entry threshold regardless of how well it would run the benchmark, and the question of what fraction of peak it could reach never arises.
No MHPCC system has appeared on the list since June 2017. June's list was topped by a Chinese all-CPU system that retook the number one position, a contest run at a scale with no bearing on what happens in Kihei. Machines sized to a mission instead of a ranking are no longer unusual: Japan scoped its next flagship around the work it has to do and said so publicly.
MHPCC has not published the basis for its peak figure. What follows is SCN arithmetic rather than confirmed fact, and it reproduces the announced number closely enough to be worth showing.
NVIDIA's published H100 specifications put FP64 vector throughput at 34 TFLOPS for the SXM part and 30 TFLOPS for the NVL. For the processors, assume 16 double-precision FLOPs per cycle per core, the standard figure for Zen 4. That assumption can be tested against HPCMP's own published numbers instead of argued from AMD documentation: Barfoot's 212,736 Genoa cores at 2.4 GHz work out to 8.169 peak petaflops on that basis, against 8.2 published, a gap of 0.4 percent. That does not prove the program used this arithmetic, but it does show the assumption is consistent with a number HPCMP publishes.
Applied to Makau: 2,176 TF from the 64 SXM cards, 300 TF from the ten NVL cards, 160.6 TF from the processors across all 32 nodes. Total 2.637 petaflops, against 2.64 announced. A gap of 0.13 percent.
Counting all 32 nodes rather than the 26 compute nodes is what makes it land, and the same convention shows up in the other headline figure: 19.2 TB of usable memory is 600 GB times 32, including login and transfer nodes. Compute-side memory is 15.6 TB. Two published numbers, one accounting.
The precision basis is the more revealing part. The same 74 GPUs, at NVIDIA's own published rates, deliver roughly 143 petaflops at dense FP8 and about 287 at sparse FP8. The announced FP64 figure is about 1.8 percent of the dense number and under one percent of the sparse one. That is not concealment: quoting FP64 peak is ordinary HPCMP practice, as Barfoot shows. It is a machine introduced in a release that leads with artificial intelligence and machine learning, then measured in the units its own community has always used.
MHPCC brands itself the "Vanguard Center for High Performance Computing," and its stated mission is to evaluate and optimize early production technology for the High Performance Computing Modernization Program. Its first listed focus area is emerging technology: reducing risk by transitioning research and benchmarking into HPCMP acquisition and operations.
The program describes this center differently from its peers. HPCMP's introductory guide says the supercomputing resource centers collectively deliver "a range of compute-intensive and data-intensive capabilities," then says of this one: "MHPCC is located in Maui, HI and provides research and development for the HPCMP." Research and development, in a sentence that had the word capability sitting right there.
It said the same thing about the previous machine, in writing, nine years ago. Announcing Hōkūleʻa in January 2017, HPCMP wrote that "as its name implies, it will be used as a test system to evaluate the performance of this novel architecture for DoW-specific software." Total life-cycle investment: $2 million.
Two of MHPCC's systems sit formally outside the production pool. Coral and Builder are both listed as non-allocated, apart from the allocation process that governs the program's production machines. Coral mixes x86 and ARM nodes with A100, MI100 and MI250 GPUs plus DPUs, described as "enabling architecture comparisons" and serving "as a testbed for ARM-based HPC evaluation." Builder is a single-node container build server. The public pages are not a full inventory, though: MHPCC's own March notice names Arcturus and Lava alongside Makau, and neither appears on any public systems list.
Queue policy is the sharpest documented difference. Makau's longest queue allows a 336-hour wall clock, two full weeks, with a ceiling of 1,536 cores. Barfoot at the Army's ERDC and Blueback at the Navy's both stop at 168 hours, and their ceilings are where the gap opens. Barfoot's 168-hour queues cap at 7,488 cores; its 100,032-core queues run 24 hours. ERDC makes that split explicit in the queue names themselves, _lw for long walltime and _lg for large jobs. Blueback allows 98,304 cores at 168 hours. Makau has no large-job queue at all, and its maximum job at any wall clock is 1,536 cores, about 1.5 percent of Barfoot's ceiling. What that configuration is meant to serve is not stated in the queue tables.
The fabric is a genuine outlier and an unexplained one. Makau runs HDR InfiniBand while Barfoot, Blueback, and all four systems in HPCMP's FY25 acquisition are EX4000 machines on Slingshot-11. HPE sells the XD2000 with Slingshot as well as InfiniBand, so the fabric was a separate decision rather than a consequence of the platform, and HDR is a generation behind the NDR InfiniBand HPE markets on that line. No public document explains the choice.
TOP500's site page for MHPCC lists twenty systems, and it is not a complete enumeration. At least one more MHPCC machine carries the same site ID without appearing in the table, and a separate site record covers the Maui Research and Technology Center, the Kihei technology park MHPCC occupies. What follows is a selection from roughly twenty records, most of them pre-2006 IBM SP systems.
The center's best placement was eighth, in November 1995. Jaws reached eleventh in November 2006. Mana peaked at 59th in 2009, Riptide at 150th in June 2013, and Hōkūleʻa at 468th in June 2017.
The shape is a sawtooth, not a slide. MHPCC placed 89th on its first list in June 1994 and eighth eighteen months later. It fell to 468th by November 2004 and returned to 11th when Jaws arrived two years later. It was 448th in November 2012 and 150th the following June. Across 43 list appearances between 1994 and 2017, there are at least six such reversals, because each new machine climbed back up the list and then decayed as the field moved past it.
What declines is the envelope. The peaks run 8, 11, 59, 150, 468, and the last of them was June 2017. Nothing has appeared since. The record establishes that much and no more; no public document says whether MHPCC has submitted results since.
HPCMP's systems summary, last updated June 26, 2026, lists an upcoming MHPCC installation for the third quarter of FY2026. Crux is described as "a coalition adaptable technology transition computing system": 240 standard nodes at 168 Genoa cores apiece, ten AI/ML nodes carrying four H100s each, ten visualization nodes, five login and landing-zone nodes, and a 5.2 PB all-NVMe WEKA filesystem. On HPCMP's published node counts, that works out to roughly 43,000 cores, about thirteen times Makau's compute core count. (SCN arithmetic.)
Crux is already in MHPCC's own notices: the March 11 weather advisory names Makau, Crux and Lava together. The two machines appear side by side in the same documents, and nothing public says one supersedes the other. The language HPCMP uses is worth holding onto, though. Technology transition, and coalition: the same register as the rest of the center's public description, and the same distance from capability delivery. Whatever else Makau is, it is one instrument in a growing estate rather than the whole of what this center runs.
Press coverage has called Makau a Space Force supercomputer, which is about two-thirds right. MHPCC is an Air Force Research Laboratory center under the Directed Energy Directorate at Kirtland AFB, managed by the University of Hawaiʻi under contract, directed by Dr. Scott Pierce, and one of five supercomputing resource centers in the HPCMP. It is also dual-organized as a flight within the 15th Space Surveillance Squadron, a US Space Force unit commanded by Lt. Col. Douglas Thornton. Since August 2025, the Army's Engineer Research and Development Center has issued the operating contract, which runs HPCMP on the department's behalf: award W912HZ25C0035, $9,388,067 obligated as of the record retrieved August 16, 2026.
Which of those organizations owns what is a subject in its own right, and it is the kind of question that increasingly determines what a supercomputing facility can do.
SCN did not identify a public award record naming Makau. USAspending keyword searches return nothing relevant, no HPE award with a Hawaii place of performance surfaced, and both the cost and the funding line are undocumented.
MHPCC has not confirmed whether 2.64 petaflops is an FP64 figure, though the reconstruction above makes it difficult to read any other way. Why the center chose HDR InfiniBand has no public explanation. Whether Makau has left pioneer status for full production has not been announced, and Arcturus and Lava are named in MHPCC's own notices without being described anywhere public.
What runs on Makau is not public either. The August release names three mission areas: space domain awareness, cyber operations, and digital engineering, modeling, simulation, and analysis. It names no workload and no code. MHPCC technical director John Marlia is quoted describing the system as "compressing data analysis timelines from weeks to seconds," which is his characterization of the improvement and not a published benchmark. The center sits near the telescopes of the Maui Space Surveillance Complex, and that proximity establishes nothing about what the machine processes.
Two of the program's own numbers need care as well. HPCMP publishes 9 peak petaflops for Coral without stating a precision basis, and the same column gives 4.6 petaflops for Builder, a single node with two V100s, which cannot be FP64; neither figure is comparable to Makau's 2.64 until that basis is pinned down.
AMOS convenes at Wailea on September 15, a few miles down the coast, and its technical program is one place more detail could appear. HPCMP's systems summary is another. It already carries Crux's configuration, an end-of-life date for Makau, and a peak figure for Coral whose basis it does not state.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
