(合) ) )
105 日 日 日曜日 日曜日
(3) Type of the procurement : Rent
(4) requirements of the procurement:
A "Integrated Infrastructure System for
Simulation, In-
ference" must satisfy the following speci-
fications:
a This system consists and
Inference Infrastructure (Infrastructure
System for Data Exploitation Plat-
form)" part, "Simulation and Learning
Infrastructure" part, and Storage. Com-
pute nodes and Storage must be able to
be used transparently with each other.
B "Data and Inference Infrastructure"
part of "Integrated Infrastructure System
for Simulation, Data, Learning, and In-
ference" must satisfy the following speci-
fications:
a It must satisfy the following hard-
ware specifications:
①Compute nodes of the part consist
of the general-purpose CPU nodes,
the data analysis nodes, and the in-
ference nodes. The nodes must be
able to communicate outside the sys-
tem with a total bandwidth of
800Gbps or more.
② Total memory bandwidthe
general-purpose CPU nodes must be
350TByte/sec. or more, and total
memory capacity must be 100TiByte
or more
③ Total theoretical peak perfor-
mance of the general-purpose CPU
nodes (by double-precision floating
point arithmetics) must be
3.0PFLOPS or more. Compute ac-
celerators must not be emploved.
④Total memory capacity of co-
mpute accelerators in the data analy-
sis nodes must be 8.2TiByte or more.
⑤ Total perfor-
mance of compute accelerators in
the data analysis nodes (by floating
point arithmetics with FP4 or higher
precision, without considering spar-
sity) must be 260PFLOPS or more.
⑥ Total memory co-
mpute in the in the in the inference
nodes must be 8.2TiByte or more.
⑦ Total perfor-
mance of compute accelerators in
the inference nodes (by floating point
arithmetics with FP4 or higher preci-
sion, without considering sparsity)
must be 260PFLOPS or more
⑧Each compute node must be equip-
ped with NVMe-connected SSD(s)
with a physical capacity of 3.0TByte
or more.
⑨ The network interface of the in-
terconnect employed by compute
node must be 400Gbps or more per
node for the general-purpose CPU
node, and 400Gbps or more per ac-
celerator for the data analysis node
and the inference node. The data on
the main memory of the compute ac-
celerator device must be directly
transferred without accessing the
main memory of the general-purpose
CPU
⑩External network interfaces of
compute nodes must be 100Gbps or
more per node.
b It must satisfy the following soft-
ware specifications:
① The system must include container
management functionality based on
Kubernetes. The system must pro-
vide a web portal for its administra-
tion.
② The web
portal for project management.
C and Learning Infrast-
ructure" part of "Integrated Infrastruct-
ure System for Simulation, Data, Learn-
ing, and Inference" must satisfy the fol-
lowing specifications:
a It must satisfy the following hard-
ware specifications:
①Compute nodes of the part consist
of the simulation nodes and the lear-
ning nodes
② Total theoretical peak perfor-
mance (by double precision floating
point arithmetics) of the simulation
nodes must be 250PFLOPS or more.
③ Total memory bandwidth of the
simulation nodes must be 18PByte/
Sec or more.
④ Total theoretical peak perfor-
mance (by floating point arithmetics
with FP4 or higher precision, with-
out considering sparsity) of the learn-
ing nodes must be 2.6EFLOPS or
more
⑤ The network interface of the in-
terconnect employed by compute
node must be 400Gbps or higher per
accelerator. The data on the main
memory of the compute accelerator
device must be directly transferred
without accessing the main memory
of the general-purpose CPU.
⑥Each compute node must be equip-
ped with NVMe-connected SSD(s)
with a physical capacity of 3.0TByte
or more.
⑦ The simulation nodes and the lear-
ning nodes must be interconnected
by a network with a bandwidth of
5.0TByte/sec or more.
b It must satisfy the following soft-
ware specifications :
①A Linux operating system must
run
②Fortran2008,C11,and C++17lan-
guages must be supported including
automatic SIMD vectorization func-
tion and OpenMP API (Version 4.5
or higher). For the compute ac-
celerators, Fortran 2008, C11, and
C++17languages must be supported
with automatic parallelization func-
tion, OpenACC API (Version 2.7 or
higher), or OpenMP API (Version 5.0
or higher).
③ An MPI3.1 library must be pro-
vided
④ Python language must be sup
ported.
⑤Highly optimized math libraries
and learning libraries must be pro-
vided.
⑥A batch job system must be pro-
vided. A job that simultaneously uses
both the simulation nodes and the
learning nodes must be executed.
7 A container system must be pro-
vided.
⑧ The management servers for "Si-
mulation and Learning Infrastruc-
ture" part should be configured using
the general-purpose CPU node group
of the "Data and Inference Inf-
rastructure" part.
D Storage of "Integrated Infrastructure
System Learning, Learning,
and Inference" must satisfy the following
specifications:
a must satisfy hard-
ware specifications:
① Fast Storage System must be high
ly reliable with 20PByte or more ca-
pacity. Fast Storage System must ac-
hieve the access performance from
all the compute nodes of "Simulation
and Learning Infrastructure" part
with bandwidth of 1.2TB/s or more.
②Archive Storage System must be
highly reliable, with20PByte or
more capacity for a storage system
and5PByte or more capacity for a
tape archive system. Archive Stor-
age System must be available from
other Supercomputer Systems in
ITC/U.Tokyo.
b It must satisfy the following soft-
ware specifications:
①The area on Fast Storage System
must be mountable as parallel file
system from all the computing nodes
of this and POSIX access
must be available. It must have file
compression functions.