Resources

Computing hardware owned and operated by CAAC — 5 servers and 15 GPUs as of 2025.

1 Servers

Servers

504+

CPU cores

1,280+ GB

System memory

GPUs installed

203.7+ TB

Storage

Name	Qty	Year	CPU	Memory	GPU
ITMAYA Tensor-R42-4000A	×1	2025	Intel Xeon 4410Y x2	128 GB	NVIDIA RTX 4000 Ada (20GB) x4
ITMAYA Tensor-R42-5000A	×1	2025	Intel Xeon 6426Y x2	128 GB	NVIDIA RTX 5000 Ada (32GB) x4
Supermicro ARS-221GL-NHIR	×1	2025	—	Up to 960GB ECC LPDDR5X onboard + up to 288GB ECC HBM3e (GPU)	NVIDIA H100 (on GH200 Grace Hopper Superchip, air-cooled) (96GB) x2
Supermicro GPU A+ Server AS-4125GS-TNRT2	×1	2025	AMD EPYC 9354 x2	256 GB	—
Supermicro Hyper A+ Server AS-2126HS-TN	×1	2025	AMD EPYC 9965 x2	768 GB	—

ITMAYA Tensor-R42-4000A

ITMAYA (ASUS E12 platform)

GPU server ×1 Purchased 2025 2U rack

CPU: Intel Xeon 4410Y (2 GHz, 12 core) x2

Memory: 128 GB — 32GB DDR5-4800 ECC RDIMM x4

GPU: NVIDIA RTX 4000 Ada (20GB) x4

Storage: 2TB NVMe M.2 x1 (OS + data)

Network: 2x 1GbE (RJ-45)

Power: 2600W Platinum PSU x2

OS: Ubuntu, CUDA Toolkit, cuDNN, NVIDIA Docker

Memory slots: 16 DIMM (max 3072GB)

Onboard graphics: AST2600 64MB

Warranty: 1 year (on-site limited)

Dimensions (WxDxH): 439.5 x 800 x 88.9 mm

ITMAYA Tensor-R42-5000A

ITMAYA (ASUS E12 platform)

GPU server ×1 Purchased 2025 2U rack

CPU: Intel Xeon 6426Y (2.5 GHz, 16 core) x2

Memory: 128 GB — 32GB DDR5-4800 ECC RDIMM x4

GPU: NVIDIA RTX 5000 Ada (32GB) x4

Storage: 2TB NVMe M.2 x1 (OS + data)

Network: 2x 1GbE (RJ-45)

Power: 2600W Platinum PSU x2

OS: Ubuntu, CUDA Toolkit, cuDNN, NVIDIA Docker

Memory slots: 16 DIMM (max 3072GB)

Onboard graphics: AST2600 64MB

Warranty: 1 year (on-site limited)

Dimensions (WxDxH): 439.5 x 800 x 88.9 mm

Supermicro ARS-221GL-NHIR

Supermicro

Superchip node ×1 Purchased 2025 2U rack

Memory: Up to 960GB ECC LPDDR5X onboard + up to 288GB ECC HBM3e (GPU)

GPU: NVIDIA H100 (on GH200 Grace Hopper Superchip, air-cooled) (96GB) x2

Storage: Micron 7450 PRO 960GB M.2 NVMe PCIe 4.0 x1

Network: NVIDIA ConnectX-7 NDR/400GbE (MCX75310AAS-NEAT) x1

Power: 2000W (2+2) Titanium redundant PSU x4

Processor: 2x NVIDIA GH200 Grace Hopper Superchip (Dual Socket 2x Mirror Mezz)

CPU-GPU interconnect: NVLink-C2C

GPU-GPU interconnect: NVIDIA NVLink

Expansion slots: 4x PCIe 5.0 x16 FHFL

Drive bays: 3x front-fixed E1.S NVMe, 2x M.2 NVMe

Warranty: 3 years standard

The two onboard Hopper GPUs are integrated into the Grace Hopper Superchip modules and are not listed as discrete units in the GPU inventory sheet.

Supermicro GPU A+ Server AS-4125GS-TNRT2

Supermicro

GPU server ×1 Purchased 2025 4U rack

CPU: AMD EPYC 9354 (3.25 GHz, 32 core) x2

Memory: 256 GB — 128GB DDR5-5600 ECC RDIMM x2

Storage: Intel P5520 15.36TB 2.5in PCIe 4.0 U.2 NVMe x1

Network: 2x 10GBASE-T

Power: 2000W (2+2) Titanium redundant PSU x4

GPU capacity: Up to 10x double-width GPUs (not populated at purchase)

CPU-GPU interconnect: PCIe 5.0 x16 switch, dual-root

GPU-GPU interconnect: NVIDIA NVLink Bridge (optional), AMD Infinity Fabric Link (optional)

Expansion slots: 11x PCIe 5.0 x16 FHFL (10x GPU, 1x AOC), 1x AIOM (OCP 3.0)

Memory slots: 24 DIMM (max ~9.2TB)

Warranty: 2 years standard + 1 year paid extension

Purchased as a bare GPU-server chassis; no GPUs listed on the order. Cross-check against the GPU inventory before assuming this host is empty.

Supermicro Hyper A+ Server AS-2126HS-TN

Supermicro

CPU / storage node ×1 Purchased 2025 2U rack

CPU: AMD EPYC 9965 (2.25 GHz, 192 core) x2

Memory: 768 GB — 32GB DDR5-6400 ECC RDIMM x24

Storage: Solidigm D5-P5336 30.72TB 2.5in PCIe 4.0 QLC NVMe x6; 400GB M.2 NVMe PCIe 4.0 x2

Network: NVIDIA ConnectX-7 400GbE/NDR single-port OSFP OCP 3.0 x1

Power: 2000W (1+1) Titanium redundant PSU x2

Drive bays: 24x front hot-swap 2.5in NVMe

Expansion slots: 1x PCIe 5.0 x16, 1x PCIe 5.0 x16 AIOM (OCP 3.0)

Memory slots: 24 DIMM (max ~9TB)

Warranty: 2 years standard + 1 year paid extension

CPU/storage node — no GPUs installed. Requires ambient temperature at or below 25C for the 500W EPYC 9965 processors.

2 GPUs

Total GPUs

728 GB

Total VRAM

Distinct models

Datacenter-class

Purchase years

Model	Memory	Count	Architecture	Host
NVIDIA H100 NVL	94 GB	4	Hopper	Not recorded
NVIDIA RTX 5000 Ada Generation	32 GB	4	Ada Lovelace	ITMAYA Tensor-R42-5000A
NVIDIA RTX PRO 6000 Blackwell Max-Q	96 GB	1	Blackwell	Not recorded
NVIDIA RTX 4000 Ada Generation	20 GB	4	Ada Lovelace	ITMAYA Tensor-R42-4000A
NVIDIA GeForce RTX 3090	24 GB	2	Ampere	Not recorded

NVIDIA H100 NVL

Datacenter

94 GB · ×4 · purchased 2025

Architecture: Hopper

Memory type: HBM3

Bandwidth: 3900 GB/s

TDP: 400 W

FP16 (Tensor): 1671 TFLOPS

Host: Host not recorded

Specs: vendor datasheet

NVIDIA RTX 5000 Ada Generation

Professional

32 GB · ×4 · purchased 2025

Architecture: Ada Lovelace

Memory type: GDDR6

Bandwidth: 576 GB/s

TDP: 250 W

Host: ITMAYA Tensor-R42-5000A

Specs: vendor datasheet

NVIDIA RTX PRO 6000 Blackwell Max-Q

Professional

96 GB · ×1 · purchased 2025

Architecture: Blackwell

Memory type: GDDR7

Bandwidth: 1800 GB/s

TDP: 300 W

Host: Host not recorded

Specs: vendor datasheet

NVIDIA RTX 4000 Ada Generation

Professional

20 GB · ×4 · purchased 2025

Architecture: Ada Lovelace

Memory type: GDDR6

Bandwidth: 360 GB/s

TDP: 130 W

Host: ITMAYA Tensor-R42-4000A

Specs: vendor datasheet

NVIDIA GeForce RTX 3090

Consumer

24 GB · ×2 · purchased 2025

Architecture: Ampere

Memory type: GDDR6X

Bandwidth: 936 GB/s

TDP: 350 W

Host: Host not recorded

Specs: vendor datasheet

3 LLM API

Self-hosted vLLM server

CAAC runs its own vLLM inference server, exposed as an OpenAI-compatible API. Point any OpenAI-SDK-compatible client at the base URL below with your API key.

Base URL: https://data.kasi.re.kr/vllm/v1

Models currently served

Model	Context length	Owned by
Qwen/Qwen3.6-35B-A3B-FP8	131,072	vllm

Synced from the live server via scripts/gen_llm_models.py — see src/content/llmModels.yaml.

BizRouter

CAAC also uses BizRouter, an external LLM routing service, for workloads that call out to hosted frontier models rather than our self-hosted server.

Need an API key for either service? Contact the CAAC team — see People.