AI workstation
When the model is the constraint, not the budget.
141 GB of HBM3e per GPU at 4.8 TB/s, four petaFLOPS of FP8 and up to seven MIG instances per card. This is the configuration behind sovereign inference services and in-house fine-tuning programmes.
- 141 GB HBM3e
- GPU memory
- 4.8 TB/s
- Bandwidth
- 4 petaFLOPS
- FP8
- 120
- CPU cores
Key features
Data-centre class HBM3e for serious inference and training.
- 141 GB of HBM3e GPU memory
- 4.8 TB/s of memory bandwidth
- 4 petaFLOPS of FP8 performance
- Up to 7 MIG instances at 18 GB each
- Confidential computing supported
Full specification
- Model
- SOIKASMH200
- Device type
- Dual-socket workstation, 120 CPU cores
- Graphics
- 2–4× NVIDIA H200 (SXM form factor)
- Processor
- 2× Intel® Xeon® w9-3595X — 120 cores / 240 threads
- Memory
- 8× 64GB DDR5-4800 2Rx4 ECC RDIMM (512GB)
- Storage
- 4× 8TB NVMe
- GPU memory
- 141 GB HBM3e
- Memory bandwidth
- 4.8 TB/s
- Multi-Instance GPU
- Up to 7 MIGs @ 18 GB each
- Thermal design power
- Up to 700 W, configurable
- Confidential computing
- Supported
- Network
- 2× 10GbE RJ45 + 1× management LAN
Included with every system
Software, licensed and pre-installed
Soika Stack
GPU clustering, LLM deployment and inference management, pre-configured and licensed.
Mockingjay Network
No-code AI agent design and agent-fleet management with the Enterprise licence.
Ubuntu LTS
Hardened Linux base image with the full CUDA and container toolchain already in place.
- Parts
- 3 years
- Live service
- 3 years, engineer-led
- Support tier
- Standard 3-year; on-site service and SLA available through partners
Related
Other configurations
Sizing
Configure the SM H200
Tell us your workload and concurrency. We will confirm the configuration, lead time and the partner who delivers it in your region.