Supermicro
Supermicro ARS-221GL-NR-OTO-70 2U rack server Grace Superchip 480GB LPDDR5X 2x H100 NVL 94GB 3.84TB+960GB SSD
2U rack server for AI and HPC workloads with dual Grace Superchips, 480GB LPDDR5X, two H100 NVL GPUs with 94GB HBM3 each, and 3.84TB plus 960GB SSD storage
- WorkloadAI/ML/HPC Server
- TypeRack
- Rack Height2U
- CPU TypeNVIDIA Grace Superchip
- Number of CPU Cores72-Core Processor
- CPU FeaturesSupports up to 500W TDP CPUs (Air Cooled)
- Dual 72-core Grace Superchip accelerates AI and HPC workloads
- 480GB LPDDR5X unified memory feeds GPU compute without bottlenecks
- Two H100 NVL GPUs deliver 94GB HBM3 each for large-model training
- 84TB E1.S plus 960GB M.2 storage supports hot-swap expansion
- 25GbE dual-port networking connects clusters at high throughput
Grace Superchip AI Server
The Supermicro ARS-221GL-NR-OTO-70 is a 2U rack system built for AI, ML and HPC workloads. It integrates a dual 72-core NVIDIA Grace Superchip with 480GB of LPDDR5X memory and two H100 NVL GPUs providing 94GB of HBM3 video memory each. The platform targets data-centre teams deploying large-scale training or inference clusters.
Installation and Common Issues
The unit weighs 67.5 lbs and measures 3.46 by 17.25 by 35.43 inches, so two-person handling or a lift is advised when sliding it into a cabinet. Eight front hot-swap E1.S bays allow drive changes without tools, but the 500W TDP air-cooled design demands unobstructed front-to-rear airflow. Operate only within 10°C to 35°C and 8% to 90% relative humidity to avoid thermal throttling.
Pre-Order Checks
Verify the rack supplies at least 25GbE connectivity via SFP28 to match the onboard dual-port controller. Confirm the facility can sustain the required temperature and humidity ranges continuously. Ensure the power infrastructure can deliver sufficient capacity for dual 500W TDP CPUs plus two H100 NVL GPUs. This system is not suited for environments lacking dedicated cooling or high-speed networking.
Highlights
- Dual 72-core Grace Superchip accelerates AI and HPC workloads
- 480GB LPDDR5X unified memory feeds GPU compute without bottlenecks
- Two H100 NVL GPUs deliver 94GB HBM3 each for large-model training
- 84TB E1.S plus 960GB M.2 storage supports hot-swap expansion
- 25GbE dual-port networking connects clusters at high throughput
Specifications
| Brand | SUPERMICRO |
|---|---|
| Model | ARS-221GL-NR-OTO-70 |
| Workload | AI/ML/HPC Server |
| Type | Rack |
| Rack Height | 2U |
| CPU Type | NVIDIA Grace Superchip |
| Number of CPU Cores | 72-Core Processor |
| CPU Features | Supports up to 500W TDP CPUs (Air Cooled) |
| Chipset | NVIDIA C2 |
| Installed Memory Size | 480GB |
| Memory Type | LPDDR5X |
| SSD (Installed) | 3.84TB + 960GB |
| GPU Type | 2 x NVIDIA H100 NVL |
| Video Memory | 94 GB HBM3 |
| Ethernet Controller | 1 x Standard NVIDIA 25GbE with 2 SFP28 ports |
| Expansion Bays | 8 front hot-swap E1.S NVMe drive bay(s) |
| Temperature | Operating Temperature: 10°C ~ 35°C (50°F ~ 95°F) Non-operating Temperature: -40°C to 60°C (-40°F to 140°F) |
| Humidity | Operating Relative Humidity: 8% to 90% (non-condensing) Non-operating Relative Humidity: 5% to 95% (non-condensing) |
| Dimensions | 3.46" x 17.25" x 35.43" |
| Weight | 67.5 lbs. |
| Shipping weight | 29.71 kg |
| Package size | 1219 × 711 × 241 mm |
Questions about this item
How can I verify the Grace Superchip and dual H100 NVL GPUs are detected correctly after installation?
Check the system BIOS or UEFI interface to confirm both 72-core CPUs and the two H100 NVL GPUs with 94 GB HBM3 each appear in the hardware inventory.
Which drives and networking components are included, and what must be purchased separately?
The system ships with a 3.84TB E1.S EDSFF SSD, a 960GB M.2 SSD, and a 25GbE controller with two SFP28 ports; SFP28 transceivers and any additional E1.S drives for the six empty front bays are not included.
Should I choose a different capacity if my workload needs more than 480GB LPDDR5X memory?
The 480GB LPDDR5X is soldered to the Grace Superchip and cannot be upgraded, so select this model only if that capacity meets your AI/ML/HPC requirements.
Also in this aisle
People compared these
Same shelf, same checkout — eight coins and a 30-minute rate lock.



