LineShine hit 2.2 exaflops on the June Top500 HPL test — without a single GPU. The all-CPU supercomputer at NSC Shenzhen in China runs on in-house Arm silicon. That 52.1 gigaflops-per-watt number made Jack Dongarra pick up his keyboard again.

Dongarra has logged 36 years between the University of Tennessee and Oak Ridge National Laboratory. He's co-authoring a forthcoming paper with Torsten Hoefler of ETH Zurich and Satoshi Matsuoka of RIKEN and Tokyo Institute of Technology.

The title asks the question out loud: Do We Still Need GPUs? Rethinking AI and Scientific Computing on Matrix-Enhanced CPUs. It sits on arXiv now and heads to an Association for Computing Machinery journal soon.

What does LineShine have that GPUs don't?

The LX2 Arm server CPU at the heart of LineShine is almost certainly etched by SMIC on a 7 nanometer process. The chip runs at 1.55 GHz inside a 650 watt envelope. Pedestrian clocks next to Nvidia. Integrated matrix engines, SVE2 vector units, and SME do the heavy lifting instead of a co-processor.

Fugaku, the A64FX-powered machine at RIKEN that went live in March 2021, is the spiritual predecessor. Fujitsu's A64FX debuted Arm SVE vector extensions in 2016. LineShine's LX2 pushes further — SVE2, SME, and HBM-style stacked memory in one socket.

The combined machine delivers 2.74 exaflops of peak theoretical performance. It already runs trillion-parameter GenAI inference alongside traditional modeling and simulation in the same workload mix.

Why give up GPUs at all?

GPUs won HPC because CPUs could not deliver enough flops and could not feed them. Wider SIMD, matrix units, and stacked memory fixed that. Now those features live on the CPU die.

Dongarra, Hoefler, and Matsuoka frame it bluntly: "The strongest argument, then, is not that GPUs are useless. They are not. GPUs are extremely powerful and will remain important." A CPU, they add, "is no longer a conventional CPU" once it adds SVE, AMX, HBM, and matrix-multiply. It becomes "a general-purpose processor with accelerator-class numerical machinery."

The trio also see scientific computing and AI merging. "Future scientific applications will not simply run simulations or train neural networks separately. They will combine simulation, data assimilation, optimization, uncertainty quantification, and machine learning in tightly coupled workflows."

About half of the Top500 is still CPU-only, even after 15 years of GPU commercialization. Those buyers never wanted to pay Nvidia's CUDA tax — or shell out for HBM.

What about efficiency?

The real LineShine scores 2.2 exaflops at 42.2 megawatts. That works out to 52.1 gigaflops per watt. El Capitan at Lawrence Livermore, the number two machine on the June list, hits 60.9 gigaflops per watt. The most efficient Top500 entries pair Nvidia Grace CPUs with Hopper H100 or H200 GPUs and clear 70 gigaflops per watt.

Drop the LX2 onto TSMC's 3 nanometer process and the math changes. The author estimates fewer cores, lower clocks, smaller die, less power. LineShine could hit roughly 87 gigaflops per watt at 25 megawatts. That would halve the socket count for the same 2.74 exaflops theoretical peak.

Who else is putting matrix math on the CPU?

Intel added AMX to Sapphire Rapids Xeons in 2020. Those chips now power the Aurora supercomputer at Argonne. IBM ships matrix engines on Power10 (September 2021), Power11 (July 2025), and the z16 and z17 Telum mainframes. The company's Spyre accelerator puts bigger matrix math on PCIe cards. AMD has not yet enabled matrix units in Epyc cores.

The Arm roadmap runs in parallel. SVE shipped in 2016, SVE2 in 2019, SME with v9, and SME2 with v9.4 in 2022. Every core now carries the math.

So what comes next?

Watch for Nvidia to fold tensor cores into a future Arm server CPU. Same silicon as its GPU accelerators, the bet goes — so CUDA X keeps running.

LineShine is the proof point. The day a 3 nanometer LX2 ships, the GPU gap closes for good.