AI Development at the Deskside.

Yes — two GX10 nodes linked via ConnectX‑7 can serve larger models like Llama 3.1 405B on‑prem. The ET900N (GB300) goes further with more compute and much larger unified memory.

A single GX10 delivers about ~1 PFLOP FP4 with 128 GB unified memory, plus built‑in ConnectX‑7 and 10G LAN — ideal for private prototyping, fine‑tuning and local inference.

Yes — it refers to FP4 AI performance on the GB300 Grace Blackwell Ultra platform used in ET900N, designed for high‑throughput LLM training and inference.

Inside each system, CPU↔GPU uses NVLink‑C2C for low‑latency shared memory access; between systems, GX10 uses ConnectX‑7 and ET900N uses ConnectX‑8 for high‑bandwidth clustering.

Local systems avoid unpredictable egress and cross‑AZ data charges and reduce iteration latency — you spend more time training and less time waiting (or paying for bandwidth).

ET900N ships with NVIDIA DGX OS and the NVIDIA AI software stack; GX10/DGX Spark comes ready for local model work with NVIDIA’s preinstalled stack.

Yes — on NVIDIA GPUs that support MIG (Multi‑Instance GPU) you can carve a physical GPU into multiple, hardware‑isolated GPU instances (each with dedicated memory, compute, and cache). This lets multiple users or services run in parallel without noisy‑neighbor effects, simplifies quota management, and improves utilization. Teams can target specific MIG instances via CUDA device IDs; in clusters, Kubernetes can schedule MIG profiles using the NVIDIA device plugin.