AI training
Multi-node HGX clusters with NVLink fabric and 400G/800G networking.
Enterprise hardware sourcing · ASTER RACKIT
Compute Workloads
Accelerated, high-performance and virtualized compute
8-GPUMax node density
800GFabric speed
HGX / PCIeGPU paths
4Server OEM families
Browse this section
Product families and buying paths.
01
AI Training
Browse the AI Training page for product families, configuration notes and quotation paths.
Open page 02AI Inference
Browse the AI Inference page for product families, configuration notes and quotation paths.
Open page 03HPC
Browse the HPC page for product families, configuration notes and quotation paths.
Open page 04Virtualization
Browse the Virtualization page for product families, configuration notes and quotation paths.
Open pageCapabilities
What this sourcing desk covers for compute workloads.
Inference at scale
Dense PCIe-GPU nodes tuned for throughput-per-watt serving.
HPC & simulation
High core-count EPYC/Xeon builds with fast NVMe tiers.
Data pipelines
Storage-dense ingest nodes with 25/100G connectivity.
Cluster fabric
Spine-leaf designs with optics and cabling quoted per link.
Rack integration
Power, cooling and cabling validated per rack before dispatch.
How it works
From requirement to delivery.
- 1
Model the workload
GPU-hours, memory footprint and storage throughput targets.
- 2
Design the node
GPU count, CPU ratio, NVMe tier and network per node.
- 3
Design the cluster
Node count, fabric topology and rack power.
- 4
Burn-in & deliver
Cluster-level validation with fabric tests before shipping.
From the facility
Real operations, photographed in-house.
Exact identity first
Model, MPN, generation, chassis layout and selected options are separated before a commercial offer.
Condition is explicit
Factory-new, OEM-program and inspected pre-owned supply paths are shown separately where relevant.
Project-ready quotation
Quantity, accessories, licenses, packing, lead time and delivery route are confirmed before shipment.
Common questions
Compute Workloads sourcing FAQ.
How many GPUs fit in one node?
Up to 8 in HGX/SXM form factors; PCIe designs commonly run 4–10 depending on chassis and TDP.
NVLink or PCIe for training?
Multi-GPU training with heavy inter-GPU traffic benefits from NVLink; inference farms often run fine on PCIe.
Can you quote power and cooling requirements?
Yes — per-rack wattage, cooling and PDU specs are included with cluster quotations.
Do you support inference-only builds?
Yes — L-series/RTX-class density nodes for cost-optimized serving are a standard configuration.
Explore the rest of this section
Need an exact model or PN?
Search the catalog or send the requirement list.
ASTER RACKIT can work from a BOM, model list, part number, configuration target or replacement requirement.