Argonne's Solstice points about 100,000 Blackwell GPUs at AI work in 4 to 16 bits

In the write-up, the GPU count and the FP64 emulation belong to different platforms. The 100,000 Blackwell GPUs belong to Solstice ^[4]. The FP64 emulation, built on the Ozaki scheme, is tied to the Vera Rubin platform ^[5]. So two separate claims sit side by side: Solstice is a low-precision machine, and FP64 survives through emulation on a different platform.

The reason for the design target is silicon budget. According to The Register's report, as relayed in the dev.to write-up, NVIDIA has traded much of its accelerators' FP64 throughput for higher throughput at AI precisions ^[9]. The write-up spells out the cost: a system sized for AI training spends its transistors on low-precision matrix throughput, and those are the transistors that would otherwise deliver double precision for classical simulation ^[10]. NVIDIA's last flagship Department of Energy machines, Summit and Sierra, were commissioned in 2018 ^[12]. The gap is eight years ^[15].

The write-up does not include emulated FP64 throughput or accuracy figures. For emulation to transfer to a given lab code, it would have to hold that code's accuracy tolerance at a rate that makes up for the FP64 hardware NVIDIA gave up. We'd expect each lab to run its own validation against native FP64 before moving production codes.

The money is a separate matter. Averaged evenly over five years, $1 billion is $200 million a year ^[14]. The write-up says the pledge spreads across three delivery models: grants, laboratory programs and rented capacity ^[13]. NVIDIA also collaborates on several phase 2 Genesis awards, in fields from quantum computing and fusion to accelerator design and microelectronics ^[3].

NVIDIA's chief executive said: "With a $1 billion investment, NVIDIA is putting advanced Super Intelligence in the hands of America's scientists to accelerate breakthroughs in medicine, energy and materials." ^[11]