HPC · GPU · AI/ML INFRASTRUCTURE
WE BUILDWHAT OTHERSCAN'T.
Full-stack HPC engineering for AI/ML labs and cloud companies. From PCB layout and firmware to cluster orchestration and ML runtime tuning.
Scope a ProjectCAPABILITIES
01 / CONSULTING
HPC Infrastructure Consulting
Cluster topology, interconnect selection, storage architecture, and capacity planning for AI/ML workloads. We scope around your actual requirements — not a service catalog.
Learn more →02 / HARDWARE
GPU Cluster Design
Multi-node GPU topology, high-speed interconnect, and rack-level power and cooling — purpose-built for training and inference at scale.
Learn more →03 / ENGINEERING
Full-Stack Engineering
PCB and FPGA design through firmware, drivers, and ML runtime tuning. We own the full stack.
Learn more →COMPATIBLE HARDWARE & ECOSYSTEMS
GPU Compute
- NVIDIA H100SXM5 / PCIe
- NVIDIA A100SXM4 / PCIe
- NVIDIA H200SXM5
- AMD MI300XOAM
Interconnect
- InfiniBand NDR400 Gb/s
- InfiniBand HDR200 Gb/s
- RoCE v2100–400 GbE
- NVLink 4.0900 GB/s
Storage
- NVMe-oFTCP / RDMA
- LustreParallel FS
- GPFS / Spectrum ScaleIBM
- WekaFSFlash-native
Software Stack
- CUDA / ROCmGPU runtimes
- Slurm / PBSJob scheduling
- PyTorch / JAXML frameworks
- KubernetesOrchestration
WHY BLACKSALT
NO GENERALISTS. NO HAND-OFFS. JUST ENGINEERS.
Production-hardened experience
Every engineer on our team has shipped production HPC infrastructure — inside hyperscalers, AI labs, or semiconductor companies.
Full-stack ownership
We own the engagement from hardware selection through firmware, drivers, and ML runtime. No finger-pointing across vendors.
Scoped to your workload
We design around your actual training and inference requirements — not a pre-packaged reference architecture.
Embedded, not advisory
We work alongside your team, not above it. Deliverables are running systems, not slide decks.
HAVE A HARD PROBLEM?
Tell us about your infrastructure challenge. We scope every engagement around your actual workload requirements.