NCSA's Nov. 16 announcement of Garnet, a University of Illinois research supercomputer funded by a $334 million Schmidt Sciences grant, says the initial deployment will use the Blackwell generation of Nvidia GPUs but names neither the chip variant nor the GPU count. That choice decides how fast Garnet can run the double-precision (FP64) math behind climate, fluid-flow and astrophysics simulation. Nvidia's technical brief, tabulated in a 2026 research paper, lists about 37 trillion floating-point operations per second (teraflops) of FP64 for the B200 chip and 1.2 teraflops for the B300 chip, also called Blackwell Ultra, a difference of about 31 times.

The National Center for Supercomputing Applications, which R. Srikant has directed since Jan. 1, says the system is expected to launch in mid-2027. For researchers deciding whether to port simulation codes to Garnet, and for Illinois faculty who will compete for its reserved time, the unpublished numbers are the story: the announcement states the money, not the capability.

NCSA Has Not Published Garnet's GPU Count or Blackwell Variant

NCSA says it will acquire and operate Garnet, and its announcement describes Nvidia Blackwell GPUs without giving a GPU count, a chip variant or any performance, storage or networking figures. The center says the system is designed to evolve, incorporating newer hardware, a phrase that leaves open both the launch hardware and what replaces it later.

Tim Boerner, NCSA's associate director of computing infrastructure, will lead the project and calls Garnet the cornerstone computing resource of the Schmidt Science Computing Network (SSCN). NCSA also marked its 40th anniversary this year, and the announcement points to that record of running national-scale computing infrastructure.

The dollar figure does not reveal the GPU count. The University of Florida paid $24 million for 504 Blackwell GPUs in 63 Nvidia DGX B200 systems, announced Jan. 27, 2025. Garnet's $334 million is about 13.9 times that sum, but multiplying Florida's GPU count by that ratio would be a guess, because the announcement names neither the variant nor the count to scale from.

Does Blackwell B300 Cut FP64 Performance Compared With B200?

On Nvidia's published figures, yes, by a wide margin. The technical brief lists the B200 at about 37 teraflops of FP64 and the B300 at 1.2, according to the same research paper. Nvidia's HGX product page tells the story at board level, listing 296 teraflops of FP64 for an eight-GPU HGX B200 baseboard and 10 for an HGX B300 board, which works out to roughly 37 and 1.2 per GPU. Other Blackwell parts differ: The Register reported that the GB200 superchip delivers about 45 teraflops per GPU, so the exact figure varies by product.

For context, Nvidia's Hopper-generation H100 and H200 deliver 34 teraflops of vector FP64 and 67 of matrix FP64, and the B100 sits at 30, according to HPCwire. The Register's 2024 analysis found Blackwell's vector FP64 about 32 percent faster than Hopper's while its matrix FP64 ran about 32 percent slower. A buyer who assumes "Blackwell" means "faster at everything" can therefore end up with less matrix FP64 than the prior generation or, with the B300, far less.

The cause is a design priority, not a defect. Jack Dongarra, presenting the TOP500 list at SC25, emphasized that Nvidia's FP64 performance showed no meaningful improvement from Hopper to Blackwell, HPCwire reported. Nvidia's Dion Harris told the publication the company recognizes that FP64 is "certainly core" to its HPC business.

Nvidia's FP64 Emulation Helps Dense Matrix Codes but Not Vector-Heavy Simulation

Nvidia's answer to the FP64 squeeze is software emulation: its CUDA libraries use the Ozaki method to reproduce FP64 matrix multiplication on low-precision tensor cores, and The Register reported Nvidia's claim that emulation can reach up to 200 teraflops of matrix FP64 on its newer Rubin chips. The emulation does not replace hardware FP64, which stays near 33 teraflops on Rubin.

Many production simulations do not look like matrix multiplication. TechTimes reported that vector-dominated codes such as fluid dynamics see no gain from emulation, and that AMD's Nicholas Malaya said fewer than 10 percent of real-world applications have made the changes needed to use it. NCSA lists AI, astrophysics and the biological and climate sciences among Garnet's target fields, so the share of that work that can use emulated FP64 would decide how much of a B300's missing FP64 can be recovered.

Peak FP64 is also not the only yardstick. At SC25, TOP500 co-maintainer Erich Strohmaier noted that Blackwell systems rank higher on the memory-bandwidth-bound HPCG benchmark than on the TOP500 list, which is built on FP64 math. A system's usefulness to science therefore depends on the variant, the memory configuration and the code, none of which NCSA has specified for Garnet.

Garnet Reserves Time for Illinois Campus, but Allocation Rules Are Unpublished

NCSA's most recent GPU system, DeltaAI, was built with nearly $30 million in NSF awards and runs 608 Nvidia H100 GPUs, packaged as 152 nodes of four Grace Hopper superchips each. University of Illinois researchers receive 10 percent of allocations from the NSF award. Garnet's grant is about 11 times the DeltaAI award, and it comes from a private foundation instead of a federal agency.

For Garnet, NCSA says allocations will be reserved for campus and that Illinois researchers will access it much like NCSA systems funded by the National Science Foundation, but the announcement gives no allocation sizes or application procedures. Schmidt Sciences says access runs through allocation processes managed by participating institutions and selected Schmidt Sciences programs.

Blue Waters Drew $358 Million in NSF Funding and Skipped TOP500 Listing

NCSA's previous flagship offers a scale comparison. NSF awarded $208 million for Blue Waters and provided $150 million over five years for operations, a combined $358 million, and the state of Illinois contributed $65 million toward the facility that houses it. Blue Waters launched in 2013 and was decommissioned by the end of 2021. Those are nominal dollars for a differently scoped project, so the totals are only a rough guide.

NCSA decided not to submit Blue Waters to the TOP500 list; deputy project director William Kramer argued the ranking was misleading as a measure of sustained science performance. The Garnet announcement does not mention benchmark submissions, and without a TOP500 or comparable entry, outsiders would have fewer independent ways to size the machine.

NCSA Lists No Power or Cooling Figures for Facility Built in 2010

NCSA's announcement places Garnet in the National Petascale Computing Facility (NPCF) on the Urbana-Champaign campus. NCSA's facility page describes an an 88,000-square-foot building built in 2010, with a 20,000-square-foot machine room, campus chilled water, three cooling towers and 480-volt AC power, but it states no megawatt capacity and no liquid-cooling provision.

Rack design raises the stakes. Nvidia says its GB200 NVL72 rack needs 120 kilowatts of cooling capacity delivered by direct liquid cooling, and Schneider Electric's infrastructure supports up to 132 kilowatts per rack for such systems. NCSA has not said that Garnet will use that rack design, so the question for the project is open, not a finding.

Argonne National Laboratory sets the national-lab reference point: its Solstice system would use 100,000 Blackwell GPUs, and its Equinox system would use 10,000. TechTimes reported Solstice at about 200 megawatts and slated for 2027. Without a Garnet GPU count, it is not yet clear where Garnet will rank among Blackwell machines.

Schmidt Sciences Funds the Network but Its Page Lists No Dollar Amounts

Schmidt Sciences describes SSCN as a distributed network of advanced computing systems owned and operated by participating universities and research institutions. The $334 million figure appears in NCSA's announcement; the Schmidt Sciences network page lists no dollar amounts.

Schmidt Sciences President Stuart Feldman created the Make build tool, served as an IBM Research vice president and was ACM president from 2006 to 2008. Schmidt Sciences was founded in 2024 by Eric and Wendy Schmidt.

Eric Schmidt has a long technology-policy record: he was Google's chief executive from 2001 to 2011, chaired the Defense Innovation Advisory Board until 2020, chaired the National Security Commission on Artificial Intelligence, and founded the Special Competitive Studies Project in 2021. Wendy Schmidt, who co-founded the Schmidt Ocean Institute with Eric Schmidt, leads the couple's philanthropy.

For anyone weighing what Garnet can do, the decision comes down to four open items. NCSA has not named the Blackwell variant, and Nvidia's own figures put B200 and B300 FP64 about 31 times apart. It has not published a GPU count, so the $334 million cannot be converted into capacity. Nvidia's FP64 emulation covers dense matrix work, not vector-heavy simulation. Access beyond Illinois's reserved share is unannounced, and Garnet's mid-2027 launch leaves time for NCSA to answer all four before researchers have to commit their codes.

Frequently Asked Questions

What is double-precision (FP64) math, and why do scientists need it?

FP64 stores each number with 64 bits, which keeps rounding errors small across the millions of calculation steps in a climate, fluid-dynamics or orbital simulation. AI training usually tolerates lower-precision formats, which is why chip makers can trade FP64 capability for AI speed.

Can software emulation make up for a low FP64 rate on Garnet?

Only partly. Nvidia's emulation targets dense matrix multiplication, so codes dominated by vector operations see little or no benefit, according to The Register's emulation report and TechTimes reporting on emulation.

Will Garnet appear on the TOP500 list?

NCSA's announcement does not say. NCSA declined to list Blue Waters on the ranking, HPCwire reported in 2012, and an unlisted machine is harder for outsiders to measure.

Who will be able to use Garnet?