HPC & SmartNIC
Crypto offload with zero compute CPU impact
FPGA as SmartNIC for streaming encryption, CPU for MPI jobs, GPU for scientific kernels.
FPGA
Encrypted RDMA NIC offload
Workloads
- ▸ RDMA RoCE v2 AES-GCM
- ▸ MPI over encrypted IB
- ▸ Checksum offload
Performance
400 Gbps / 0 % CPU
CPU
Scheduler + control plane
Workloads
- ▸ SLURM / PBS Pro
- ▸ MPI rank orchestration
- ▸ Encrypted checkpointing
Performance
< 1 % scheduling overhead
GPU
Scientific kernels
Workloads
- ▸ CFD / FEM / DFT
- ▸ Weather forecast kernels
- ▸ Molecular dynamics
Performance
> 1 PFLOPS FP64
Multi-agent scenario
An MPI job starts: CPU distributes ranks, FPGA encrypts each RDMA message at 400 Gbps, GPU runs the CFD kernel, final checkpoint is encrypted before writing to shared storage.