What runs on it

hpcnotes

2026.11.07

What runs on it

NSCS launched LineShine with four flagship applications. Two arXiv papers add the details of AI training at full scale.

Simulation#

Code	Field	Scale reported
PowerLLEL	turbulence, direct numerical simulation	16.5 trillion grid points at friction Reynolds number ≈ 21,000, heading to ≈ 43,000
CerebroSim	spiking brain model	86 billion neurons, 100 trillion synapses
GLARE-Exa	drug design	10^13 synthesizable compounds searched in 22.6 hours

16.5 trillion grid points of one velocity component in double precision is 132 TB, which fits in the machine's DDR about 88 times over (my arithmetic). DNS codes like PowerLLEL are FFT- and stencil-heavy, which makes them a good test of the HBM tier and of the network's all-to-all.

86 billion neurons and 100 trillion synapses are the commonly cited figures for a human brain. At even a few bytes per synapse, that is hundreds of terabytes of state, a natural fit for 11.6 PB of DDR.

AI training on CPUs#

Model	Task	Scale	Result
D2AR, 6.3 B parameters	Earth-observation compression	20,480 nodes	1.54 EFlop/s BF16 sustained, 2.16 peak, 15.7% MFU
MatRIS-MoE	universal interatomic potential	full machine	1.2 EFlop/s FP32 weak scaling at 90.3% efficiency

D2AR compresses satellite archives by 100 to 10,000 times. The team's optimisations, from SME kernels to SDMA pipelines, are on the software page.

Weak scaling of D2AR from 4 to 20,480 nodes: the 6 B model reaches 1,543 PFlop/s at 76% efficiency. Zhang et al., arXiv:2605.08633, CC BY-NC-SA 4.0, figure 7.

The MatRIS-MoE team reports the full set of scaling numbers:

Run	FP32 rate	Efficiency
Weak scaling, large model	1,200 PFlop/s	90.3%
Weak scaling, medium model	862 PFlop/s	91.3%
Strong scaling, large model	673 PFlop/s	50.6%
1,000 steps sustained, large model	1,033 PFlop/s	21% of FP32 peak

Strong scaling of MatRIS-MoE from 1/8 of the machine to all of it, on LineShine and on CNIS. LineShine reaches 672.5 PFlop/s with the large model. Zhou et al., arXiv:2604.15821, CC BY 4.0, figure 6.

They cut MoE all-to-all volume in half by compressing it to FP16 per atom type, and overlapped gradient synchronisation with the optimizer step. Step time fell from 40.7 s to 8.1 s on the large model.

The pattern#

Every one of these codes either uses dense matrix kernels (training, DNS with spectral methods) or holds an enormous state (brain, chemical space). That is the profile LineShine was designed for: a big SME peak for the dense parts, a lot of DDR for the state, and HBM in between.

#applications #ai

Simulation#

Code	Field	Scale reported
PowerLLEL	turbulence, direct numerical simulation	16.5 trillion grid points at friction Reynolds number ≈ 21,000, heading to ≈ 43,000
CerebroSim	spiking brain model	86 billion neurons, 100 trillion synapses
GLARE-Exa	drug design	10^13 synthesizable compounds searched in 22.6 hours

AI training on CPUs#

Model	Task	Scale	Result
D2AR, 6.3 B parameters	Earth-observation compression	20,480 nodes	1.54 EFlop/s BF16 sustained, 2.16 peak, 15.7% MFU
MatRIS-MoE	universal interatomic potential	full machine	1.2 EFlop/s FP32 weak scaling at 90.3% efficiency

D2AR compresses satellite archives by 100 to 10,000 times. The team's optimisations, from SME kernels to SDMA pipelines, are on the software page.

Weak scaling of D2AR from 4 to 20,480 nodes: the 6 B model reaches 1,543 PFlop/s at 76% efficiency. Zhang et al., arXiv:2605.08633, CC BY-NC-SA 4.0, figure 7.

The MatRIS-MoE team reports the full set of scaling numbers:

Run	FP32 rate	Efficiency
Weak scaling, large model	1,200 PFlop/s	90.3%
Weak scaling, medium model	862 PFlop/s	91.3%
Strong scaling, large model	673 PFlop/s	50.6%
1,000 steps sustained, large model	1,033 PFlop/s	21% of FP32 peak

Strong scaling of MatRIS-MoE from 1/8 of the machine to all of it, on LineShine and on CNIS. LineShine reaches 672.5 PFlop/s with the large model. Zhou et al., arXiv:2604.15821, CC BY 4.0, figure 6.