Supercomputers/Systems/ABCI 2.0
Japan · operational · ai training
ABCI 2.0
Also known as AIST ABCI 2.0 expansion
Operated by AIST at ABCI Data Center, Kashiwa .
Researched with AI assistance, checked by scripts. These rows were researched by AI agents reading public operator, laboratory and vendor sources, appended by scripts and checked by automated validation; no subject-matter expert has reviewed them (how this site is made). A mechanical second-reader pass has since fetched every cited source: 1350 of 1581 systems have a readable citation that names them, and 64 are genuinely weakly sourced. Every claim carries its source and a confidence tier. Treat anything below verified as a lead, not a citation.
Measured performance
- Rmax
- 22.21 PFlop/s
- Rpeak
- 54.34 PFlop/s
- Rmax ÷ Rpeak
- 40.9%
- Cores
- 504,000
- Power
- 1.60 MW
- Per watt
- 13.9 GF/W
Figures are as last publicly reported for the configuration described below, not a live measurement. Where a system was upgraded in place, the post-upgrade configuration is shown and the earlier one appears in the timeline.
How it compares
Ranked against the other machines in this dataset that report the same figure: those that first came into service within two years of ABCI 2.0, and those in Japan. Computed from our tables; a metric is ranked only when at least eight peers report it.
| Metric | Value | Against its era | Against its country |
|---|---|---|---|
| Rmax | 22.21 PFlop/s | #32 of 210 · top 15% | #12 of 131 · top 8% |
| Energy efficiency | 13.9 GF/W | #50 of 109 · top 45% | #19 of 74 · top 25% |
| Rmax ÷ Rpeak | 40.9 % | #202 of 207 · bottom 3% | #127 of 130 · bottom 3% |
| Power draw | 1,600 kW | #19 of 112 · top 16% | #10 of 74 · top 12% |
Every metric is ranked largest-first, so a high power-draw rank means a big power draw, not a good one. Rmax ÷ Rpeak uses HPL-measured results only. Rankings move as machines are added; they describe this dataset, not the world.
What this machine is made of
One row per supplier relationship. “Supplier at build” is the company that shipped the part at the time; where that company has since been acquired, the parent it rolls up to today is shown beside it. That distinction is what makes ticker-level aggregation possible across a thirty-year dataset.
| Role | Part | Supplier at build | Quantity | Confidence | Source |
|---|---|---|---|---|---|
| Integrator | - Fujitsu PRIMERGY GX2570 M6 (Compute Node A) and PRIMERGY CX2570 M4 (Compute Node V) servers | Fujitsu | - | Reported | abci.aitop500.org |
| Accelerator | NVIDIA A100 SXM4 40GB Accelerators CUDA (Ampere GA100) · 6,912C · 1.41 GHz · 7 nm · 400 W 120 Compute Node (A), each with eight NVIDIA A100 for NVLink 40GiB HBM2 | NVIDIA | 960 reported | Reported | abci.aitop500.org |
| Accelerator | NVIDIA Tesla V100 Accelerators CUDA (Volta GV100) · 5,120C · 1.53 GHz · 12 nm · 300 W 1,088 Compute Node (V), each with four NVIDIA Tesla V100 SXM2 16GiB | NVIDIA | 4,352 reported | Reported | abci.ai |
| CPU | - Two Intel Xeon Platinum 8360Y (36 cores, 2.4 GHz) per Compute Node (A) across 120 nodes | Intel | 240 estimated | Reported | abci.aitop500.org |
| CPU | Intel Xeon Gold 6148 CPUs x86-64 · 20C · 2.4 GHz · 14 nm · 150 W Two Intel Xeon Gold 6148 (20 cores, 2.4 GHz) per Compute Node (V) across 1,088 nodes | Intel | 2,176 estimated | Reported | abci.aitop500.org |
| Interconnect | NVIDIA InfiniBand HDR Interconnect Fat tree or dragonfly+ · 200 Gb/s per port · About 0.6 microseconds Four InfiniBand HDR (200 Gbps) links per Compute Node (A) | NVIDIA | - | Reported | abci.aitop500.org |
| Interconnect | Mellanox InfiniBand EDR Interconnect Fat tree · 100 Gb/s per 4x port · About 0.9 microseconds Two InfiniBand EDR (100 Gbps) links per Compute Node (V) | NVIDIA | - | Reported | abci.aitop500.org |
Engineering, funding and access
What ABCI 2.0 runs on, who paid for it and how people get time on it. Each line carries who said it and when.
Hardware
| Item | Detail | As of | Basis | Source |
|---|---|---|---|---|
| Compute nodes | 1,208 nodes120 Compute Node (A) with 960 A100 plus 1088 Compute Node (V) with 4352 V100 | 2021 | Operator stated | abci.ai |
| Storage capacity | 47 PBFive storage systems; includes 17 PB physical S3-compatible cloud storage | 2021 | Operator stated | abci.ai |
| Interconnect and topology | InfiniBand: full bisection among the A nodes, full bisection within a rack for the V nodes; SINET6 at 400 Gb/s external | 2021 | Operator stated | abci.ai |
Software
| Item | Detail | As of | Basis | Source |
|---|---|---|---|---|
| Parallel filesystem | Shared distributed file system (Lustre) with about 34 PB effective capacity plus S3-compatible ABCI Cloud StorageABCI's page spells the file system as Luster | 2021 | Operator stated | abci.ai |
The read
AIST's 2021 expansion of the ABCI AI facility, listed by TOP500 as ABCI 2.0. AIST's ABCI 2.0 page describes 120 Compute Node (A) Fujitsu PRIMERGY GX2570 M6 servers, each with two Xeon Platinum 8360Y CPUs and eight NVIDIA A100 40GB GPUs on four HDR InfiniBand links (960 A100 in total), plus 1,088 Compute Node (V) PRIMERGY CX2570 M4 servers with two Xeon Gold 6148 CPUs and four Tesla V100 SXM2 GPUs each on dual EDR links (4,352 V100); the V nodes are the 2018 ABCI hardware recorded in the abci row. Storage is a Lustre shared file system of about 34 PB and an S3-compatible object store with 17 PB physical capacity. TOP500 lists the entry as Fujitsu with 504,000 cores, Rmax 22.21 and Rpeak 54.34 PFlop/s, 1,600.32 kW (524.40 kW in an optimized run that reached 11.48 PFlop/s) and an installation year of 2021; its title names the A100 GX2570 M6 nodes while its processor field gives Xeon Gold 6148, so the entry appears to cover both node types, and AIST publishes no combined LINPACK figure to cross-check. It ranked 12th in June 2021, its best position, and was last listed in June 2024 at 39th, which is why status is held at that date. abci-3 (ABCI 3.0, H200) is the later generation and abci.ai says the service moved to version 3.0 in January 2025; AIST gives no separate end date for ABCI 2.0.
Timeline
Public list appearances
Pointers only: rank and edition, linking to the canonical entry. We do not reproduce list tables. See sourcing policy.
Site & facility
ABCI Data Center, Kashiwa
- Location
- Kashiwa, Japan
- Cooling
- Direct liquid cooling with hot-water loop
- PUE
- 1.1
- Commissioned
- 2018
Change history
- 2026-10-17 Added ABCI 2.0
Nearby in scale
The machines closest to ABCI 2.0 by Rmax, smaller to larger.
- KDDI GPU Cloud · Japan, operational · 22.21 PFlop/s
- Wisteria/BDEC-01 · Japan, operational · 22.12 PFlop/s
- DeepL Mercury · Sweden, operational · 21.85 PFlop/s
- Taipei-1 · Taiwan, operational · 22.3 PFlop/s
- Dammam-7 · Saudi Arabia, operational · 22.4 PFlop/s
- TELUS Sovereign AI Factory (Rimouski) · Canada, operational · 22.74 PFlop/s
Related systems
Same site
- ABCI · Japan, decommissioned, 2018
- ABCI 3.0 · Japan, operational, 2025
- ABCI-Q · Japan, operational, 2025
Same operator
- AIST AI Cloud · Japan, operational, 2017
- Blue Protein · Japan, operational, 2005
- AIST Super Cluster P-32 · Japan, operational, 2004
Same accelerator
- AlphaCentauri · Germany, operational, 2021
- Athena · Poland, operational, 2021
- Baskerville · United Kingdom, operational, 2021
- Berzelius · Sweden, operational, 2021
- Earth Simulator 4 · Japan, operational, 2021
- Hortense · Belgium, operational, 2021
Source check
We fetched this system's own citations and recorded whether each page actually mentions it. This is a corroboration signal, not a fact check, and it is published so you can see how well the sourcing holds up rather than take it on trust.
| Cited source | Result | Found on the page |
|---|---|---|
| abci.ai | names system, part and count | NVIDIA A100 SXM4 40GB, NVIDIA Tesla V100, Intel Xeon Gold 6148, NVIDIA InfiniBand HDR, Mellanox InfiniBand EDR · counts: 960, 4352 |
| top500.org | names system and part | NVIDIA A100 SXM4 40GB, Intel Xeon Gold 6148, NVIDIA InfiniBand HDR, Mellanox InfiniBand EDR |
Checked 2026-10-18 by pnpm verify. Re-run it and the table changes with the web.
Further reading
Operating a supercomputer
Acceptance testing, why "installed" and "in production" are six months apart, and the machine lifecycle.
Inside a node
Why accelerators exist, what the memory hierarchy costs, and how the CPU and accelerator merged onto one package.
The interconnect
Topologies, why latency and tail behaviour matter more than bandwidth, and how the fabric market consolidated.