Compute Devices

RTX 4090 × 8

Cost-efficient Ada GPU for experimentation, distillation, and high-FPS inference.

V100 32GB PCIe × 8

Proven Volta platform for classic DL workloads and lightweight inference.

L20 × 8

Ada-generation variant optimized for media & inference, with a larger memory footprint.

RTX 5090 × 8

Latest Blackwell consumer-class GPUs — best throughput per dollar for distillation and small-model training.

L40 × 8

Versatile Ada-generation inference GPU balancing generative AI, rendering, and virtualization.

A100 40GB PCIe × 8

Entry-tier Ampere for teams that need A100-class throughput without the 80GB premium.

A100 40GB PCIe × 4

Entry point into A100 inference — good for SLM serving and small fine-tuning runs.

A100 40GB PCIe × 2

Smallest A100 configuration — ideal for proof-of-concept and serving 7B models.

A800 80G PCIe × 8

PCIe form-factor A800 — flexible slot configuration, budget-friendly Ampere option.

A800 SXM 80G × 8

Ampere platform tuned for procurement-compliant deployments in the US market.

A800 80GB PCIe × 2

Compact A800 node for individual researchers and small teams.

A100 80GB SXM × 8

Ampere workhorse — still the best price-performance for fine-tuning 7B–70B models.

HGX H20 768 GB

Best-in-class inference GPU for large-model RAG and multi-tenant serving. 141 GB HBM3e per card.

H800 SXM × 8

Cost-optimized Hopper derivative for organizations with export-restricted procurement requirements.

H100 80GB SXM5 × 8

Flagship Hopper GPU for transformer training and distributed inference. Battle-tested for the widest range of AI workloads.

H200 141GB DELTA-NEXT HGX × 8

Hopper-refresh platform with 141GB HBM3e per GPU — single-node serving of 70B models and 8-way deployment of DeepSeek 671B.

B200 NVLink × 8

Top-of-the-line Blackwell platform for frontier-model training and large-scale fine-tuning. Non-blocking 400G IB fabric ready.