0:04
Welcome back to Altitude Addicts, where
0:06
we dive into the engineering behind the
0:08
world's most complex machines and
0:10
scientific infrastructure.
0:12
Today we are looking inside a brand new
0:15
liquid cooled supercomputing cluster
0:17
just installed at Russia's Siberian ring
0:19
photon source. We will break down the
0:22
raw hardware numbers, how it handles
0:24
massive torancets of experimental data,
0:26
and how it stacks up against global
0:28
exoscale giants. So stick around until
0:33
The Siberian ring photon source known as
0:35
SKF in Russia has been equipped with a
0:38
390 teraflop computing cluster. This
0:42
setup serves as dedicated
0:44
high-performance computing
0:45
infrastructure for one of the nation's
0:48
major scientific efforts.
0:50
The system processes experimental data
0:52
generated by Synretron Research. And
0:55
this system developed by the Russian
0:56
technology company RSC combines liquid
0:59
cooled servers, GPU accelerators, and
1:02
highcapacity storage.
1:04
As Russia expands its scientific
1:06
computing infrastructure, officials have
1:09
finalized the installation and testing
1:11
of a dedicated data center at skiff in
1:13
Kulsovo, Novasabursk region.
1:16
The computing cluster reportedly
1:18
provides a performance of 390 teraflops
1:22
in double precision FP64 calculations
1:25
which is equivalent to 0.39 paflops.
1:30
The facility is intended to facilitate
1:32
the continuous collection, processing
1:34
and storage of experimental data
1:36
produced by skiff's research stations.
1:40
The Russian RSC group, which specializes
1:42
in data center infrastructure and high
1:45
performance computing systems, devised
1:47
and executed the project. The
1:50
installation has successfully completed
1:52
testing and is prepared for operation as
1:55
indicated in the skiff project
1:56
announcement on November 13th, 2026.
2:01
This computing infrastructure, an
2:03
essential component of the broader
2:04
scientific system, is currently under
2:07
development as a fourth generation
2:09
synretron radiation facility.
2:12
Think about why this actually matters
2:14
for a second. Modern scientific
2:16
facilities simply drown in data if they
2:19
do not have dedicated computing muscle
2:21
right next door. Advanced detectors can
2:24
spit out an absurd amount of raw
2:26
information in seconds, which means you
2:29
need computing resources that can
2:31
swallow those massive data sets without
2:33
choking or disrupting ongoing
2:37
Stepping inside Russia's 390 teraflop
2:40
computing cluster, we find a very
2:43
specific hardware layout. The SCF
2:46
computing infrastructure comprises 24
2:48
RSC Tornado servers which collectively
2:51
generate 182 teraflops of performance
2:55
alongside two additional servers that
2:56
are outfitted with GPU accelerators.
3:00
The combined FP64 performance is
3:03
reported to be 390 teraflops with the
3:06
GPUs contributing an additional 28
3:11
The systems liquid cooling architecture
3:14
is a critical component of this setup.
3:17
The 24 RSC Tornado servers employ a
3:20
custom cooling solution that RSC has
3:22
developed to manage the thermal load
3:24
generated by continuous computing
3:26
activity and support reliable operation.
3:30
Server and networking equipment, data
3:32
storage infrastructure, environmental
3:34
control, and an uninterruptible power
3:36
supply with a capacity of 300 kilowatts
3:40
are also included in the installation.
3:43
Storage is another critical component of
3:45
the system and it is split into two
3:50
The data center is equipped with 900 tab
3:53
of storage for warm data and an
3:55
additional 368 tab of distributed hot
3:59
data storage. Meanwhile, its network
4:02
infrastructure is capable of
4:04
transmitting experimental data from
4:06
detector systems to computing servers at
4:09
a rate of up to 200 Gbits per second.
4:13
This combination of storage, networking,
4:16
and computing is tailored specifically
4:18
to the needs of scientific
4:20
investigations as opposed to traditional
4:22
office or enterprise workloads.
4:25
According to a technical report
4:27
transmitted via Interfax, every piece of
4:30
this architecture is built around raw
4:32
data throughput. Here on Altitude
4:35
Addicts, we always look at how
4:36
specialized engineering solves extreme
4:39
bottlenecks. And coming up next, we are
4:42
going to look at why a synretron
4:43
particle accelerator needs a dedicated
4:46
supercomputer in the first place.
4:50
So why does a synretron need a
4:53
The computing cluster is being developed
4:56
to provide support for skiff, an
4:58
important scientific installation that
5:00
is intended to generate synretron
5:03
radiation of high intensity. The
5:06
facility generates radiation that
5:08
researchers can utilize to investigate
5:10
the internal structure and behavior of
5:12
various materials by accelerating
5:14
electrons to an energy of 3
5:19
Synretron radiation allows scientists to
5:22
examine materials at extremely small
5:24
scales. It is used in the study of
5:27
biological molecules, the analysis of
5:29
chemical reactions, the development of
5:32
new functional materials, and the study
5:34
of crystal atomic structure.
5:37
Pharmaceutical development, micro
5:39
electronics, energy technologies, and
5:41
industrial manufacturing can also
5:43
benefit directly from this research.
5:47
When fully developed, SCF is intended to
5:49
house 30 experimental stations. The
5:52
computing cluster is anticipated to
5:55
manage the data produced by the initial
5:57
group of stations as the first stage
5:59
infrastructure is introduced gradually.
6:03
Let's put those data numbers into plain
6:05
English so your brain doesn't glaze
6:07
over. Just one single piece of equipment
6:11
currently installed there can crank out
6:13
roughly 5 terabytes of raw experimental
6:16
data in a single day using phase
6:18
contrast tomography. According to this
6:21
project announcement, imagine trying to
6:23
move that over a normal home internet
6:28
That brings us directly to how Russia
6:30
engineered a flexible computing
6:32
architecture for different experiments.
6:35
The resource management architecture of
6:37
the skiff computational cluster is one
6:39
of its most critical features.
6:42
Rather than designating fixed computing
6:44
resources per group, the system can
6:46
dynamically allocate processor cores,
6:49
GPU accelerators, memory, and high-speed
6:52
storage to individual research groups in
6:54
accordance with the needs of each
6:58
This enables researchers to customize
7:00
virtual computing environments to
7:02
accommodate particular scientific
7:06
The allocated resources can then be
7:08
released and reassigned to other users
7:10
after an experiment has been completed
7:13
while experimental data is transferred
7:15
from high-speed storage to the primary
7:18
storage infrastructure for longerterm
7:22
This arrangement is especially important
7:24
due to the fact that the computing
7:26
requirements of various experiments vary
7:30
While some research teams generate
7:32
relatively small amounts of data that
7:35
necessitate a significant amount of
7:37
processing capacity, others generate
7:40
massive data sets that demand far
7:42
greater storage and network
7:45
The project also includes connecting
7:47
skiff's computing cluster with
7:49
supercomputers supporting other Russian
7:52
mega science facilities through the
7:54
national research computer network or
7:58
This integration could enable scientific
8:00
organizations to coordinate data
8:02
processing activities and share
8:04
computing resources across multiple
8:06
research installations, tying together
8:08
the technical capabilities of RSC and
8:11
skiff. Before we wrap things up, stick
8:14
with me because we are about to compare
8:16
this cluster against the world's fastest
8:19
exascale supercomputers and look at how
8:22
much of this hardware is actually built
8:27
Now let's examine how this system
8:29
compares with Russia's and the world's
8:33
At 390 teraflops, skiff is not designed
8:37
to compete with the world's largest
8:41
El Capitan and Frontier which are
8:43
American systems operate at the exoscale
8:46
level achieving performance measured in
8:48
thousands of pedlops and the reported
8:50
performance of Russia's largest publicly
8:52
benchmarked commercial computing systems
8:55
such as Chervanis from Yandex and
8:57
Christophari Neo from Spur is also
9:00
significantly higher.
9:03
Nevertheless, it is necessary to qualify
9:07
The SKF figure is a representation of
9:09
the reported FP64 performance. Whereas
9:13
international supercomputer rankings
9:15
typically utilize measured
9:16
high-performance LINAC results and those
9:19
two measurements are not directly
9:23
Furthermore, the skiff installation sits
9:26
in a completely different category
9:28
within Russia compared to large
9:30
commercial AI and generalpurpose
9:34
The architecture and resource allocation
9:36
of the system are specifically tailored
9:39
to meet the realtime needs of synretron
9:43
Consequently, its success will be
9:45
contingent upon the efficiency with
9:47
which it manages experimental data,
9:49
provides assistance to scientific users
9:52
and ensures continuous operation.
9:55
At the same time, this project is all
9:57
about building domestic capabilities in
10:00
high performance computing.
10:02
The initiative illustrates the growing
10:05
role of Russian companies in the
10:07
development of specialized computing
10:09
infrastructure for national research
10:13
RSC has been responsible for the design
10:16
and implementation of the cluster which
10:18
includes the integration of computing
10:20
storage and networking resources as well
10:23
as its custom liquid cooling solution.
10:26
Let's be real here for a second. Just
10:29
because a domestic company bolts the
10:31
racks together and writes the management
10:33
software doesn't magically mean every
10:35
silicon chip inside was etched in a
10:38
local factory. The published project
10:41
specifications do not guarantee that
10:43
each processor, GPU, storage component,
10:46
and networking device is manufactured in
10:48
Russia. Even so, acquiring the
10:51
engineering skills to design, integrate,
10:54
chill, and operate specialized
10:56
high-performance computing systems
10:58
remains a critical step in building
11:00
national technological capabilities.
11:03
Ultimately, the scientific
11:04
infrastructure of Russia is set to be
11:06
strengthened by the broader skiff
11:08
initiative. Its experimental facilities
11:12
are expected to facilitate breakthroughs
11:14
in physics, chemistry, biology, geology,
11:17
and material science while giving
11:19
industrial organizations access to
11:21
advanced synretron techniques. Thanks
11:24
for watching Altitude Addicts, and let
11:26
us know in the comments what piece of
11:28
mega science engineering you want us to
11:34
We thank the people who have shared,
11:36
subscribed, and hyped the videos.