FugakuNEXT by RIKEN: status and timeline

FugakuNEXT

Give Japanese science and industry a machine about 100 times faster than Fugaku on real applications, within the same ~40 MW, by fusing simulation with AI.

ResearchStage 1 of 5

Basic design done Feb 2026 (report May 2026); detailed design under way, operation around 2030.

Updated 29 May 2026Checked 11 Oct0 updates this week

Milestones

Next · Detailed design phase

MEXT launches the FugakuNEXT projectJan 2025Complete.
Fujitsu wins basic design contract18 Jun 2025Complete.
NVIDIA joins; international initiative launched22 Aug 2025Complete.
Basic design completedFeb 2026Complete.
Detailed design phaseTarget 2026Current milestone.
System begins operation in KobeTarget 2030Not yet reached.

Most important updates

Upcoming

2026Detailed design phase (next)
2030System begins operation in Kobe

Current obstacles

Key choices still openGPU memory (due end-2026), GPU compute specs (mid-2027), scale-up size (4 vs 72 GPUs) and network (end-2027) are still open.
Porting Fugaku's software to GPUsLarge science codes tuned for Fugaku's CPUs must be adapted to GPUs to reach the speed-up target.
Fitting zetta-scale into 40 MWThe power cap is fixed, so every chip, link and cooling choice is judged on energy per useful result.

Physics limits

Moving data costs more than computingShifting a number between memory and processor uses far more energy than the arithmetic itself, so memory bandwidth, not raw FLOPS, caps speed for most science codes.
Power sets the ceilingA national site can supply only tens of megawatts. Peak speed therefore grows only as fast as chips improve energy per operation, which has slowed as transistor shrinking stalls.
Low precision trades accuracy for speedFP8 maths is many times faster than the 64-bit maths simulations need, but rounding errors grow; it only helps where algorithms can tolerate or correct them.

How it works

4 parts

CPU

MONAKA-X, a many-core Arm chip

Fujitsu's CPU is meant to run Fugaku's compiled programs as they are, and adds vector and matrix engines for AI-style maths.

GPU

NVIDIA accelerators for the heavy maths

GPUs do most of the arithmetic, especially low-precision matrix work for AI. It is the first Japanese flagship to use GPUs as accelerators.

Link

A fast CPU-GPU link

A memory-sharing link lets CPU and GPU work on the same data without slow copies. NVIDIA's NVLink Fusion is the lead option; decision due by end-2027.

Software

Algorithms supply most of the gain

Hardware adds about 5x; mixed-precision maths and AI-based solvers are meant to add a further 10-20x on real applications.

Spec sheet

vs Fugaku

Spec	FugakuNEXT	Fugaku	Change
AI peak (FP8, sparse)	600+ EFLOPS^R (reported)	—	—no comparable change
Effective AI performance target	50+ EFLOPS^R (reported)	—	—no comparable change
Power limit	40 MW^R (reported)	—	—no comparable change
Application speed-up vs Fugaku	Up to 100x^R (reported)	1x (Fugaku)	▲ 99 x
Hardware speed-up vs Fugaku	5x+^R (reported)	—	—no comparable change
MONAKA-X CPU process (planned)	1.4 nm, 3D chiplets^R (reported)	—	—no comparable change
Fugaku benchmark (predecessor)	442 PFLOPS (HPL)^R (reported)	—	—no comparable change

^R reported by the company

Papers & demos

1 item

29 May 2026paper
FugakuNEXT Basic Design Technical Report (opens r-ccs.riken.jp)Lays out the architecture, chips, network and performance targets agreed with Fujitsu and NVIDIA.

Update log

3 updates

About RIKEN

The team behind FugakuNEXT

RIKEN

Fugaku, FugakuNEXT supercomputers

RIKEN is Japan's national comprehensive research institute, founded in 1917. Its Center for Computational Science in Kobe runs the Fugaku supercomputer and leads the design of its successor, FugakuNEXT.

Operates Fugaku in Kobe (442 PFLOPS on the TOP500 benchmark), world number one in 2020 and 2021.
FugakuNEXT, with Fujitsu CPUs and NVIDIA GPUs, targets 600+ EFLOPS for AI in FP8 around 2030.
Published the FugakuNEXT basic design report on 29 May 2026; GPU, network and link choices come in detailed design.

riken.jp (opens riken.jp)Company profile →

Founded: 1917109 yrs
Headquarters: Japan
Status: Government
Valuation: government
Works in: ComputingAI factories & supercomputers
Coverage: 1 program · 3 updateslatest 29 May 2026checked 11 Oct
Partners: US Department of Energy (opens riken.jp)
Suppliers: Fujitsu (opens info.archives.global.fujitsu)NVIDIA (opens blogs.nvidia.com)
People: Makoto Gonokami (opens riken.jp)PresidentSatoshi Matsuoka (opens en.wikipedia.org)Director, RIKEN Center for Computational Science

Spec

FugakuNEXT

Fugaku

Change

AI peak (FP8, sparse)

600+ EFLOPS^R (reported)

—

—no comparable change

Effective AI performance target

50+ EFLOPS^R (reported)

—

—no comparable change

Power limit

40 MW^R (reported)

—

—no comparable change

Application speed-up vs Fugaku

Up to 100x^R (reported)

1x (Fugaku)

▲ 99 x

Hardware speed-up vs Fugaku

5x+^R (reported)

—

—no comparable change

MONAKA-X CPU process (planned)

1.4 nm, 3D chiplets^R (reported)

—

—no comparable change

Fugaku benchmark (predecessor)

442 PFLOPS (HPL)^R (reported)

—

—no comparable change

FugakuNEXT

Milestones

Most important updates

Upcoming

Current obstacles

Physics limits

How it works

MONAKA-X, a many-core Arm chip

NVIDIA accelerators for the heavy maths

A fast CPU-GPU link

Algorithms supply most of the gain

Spec sheet

Papers & demos

Update log

Fri 29 May

Fri 22 Aug 2025

Wed 18 Jun 2025

About RIKEN

Current obstacles

Physics limits

How it works

MONAKA-X, a many-core Arm chip

NVIDIA accelerators for the heavy maths

A fast CPU-GPU link

Algorithms supply most of the gain

Spec sheet

Papers & demos

Update log

Fri 29 May

Fri 22 Aug 2025

Wed 18 Jun 2025

About RIKEN