Networking
50 articles
-
Why Data Locality Matters for AI Performance
Data locality for AI explained: moving data vs moving compute, transfer-time maths for 1GbE to 100GbE, model load times and a tiered storage layout.
Wei Fei
·
-
How to Connect DGX Station GB300 to Enterprise NAS and High-Speed Storage
How to connect a DGX Station GB300 to a NAS: which port, NFS vs SMB, NFS over RDMA, staging data on local NVMe, link speeds and GPUDirect Storage...
Wei Fei
·
-
Why Your AI Project Might Need More Networking Power Than GPU Power
When networking really limits AI and when GPU memory matters more: link speeds, model-loading maths and cluster advice for SME AI projects.
Wei Fei
·
-
The Hidden Cost of End-of-Life Hardware: When Vendors Stop Providing Firmware Updates
What happens when vendors stop firmware updates: Cisco EoL dates, Windows 10 ESU pricing, FBI warnings on old routers, and how to plan replacements.
Wei Fei
·
-
Your GPU Has 20 Petaflops of Power. So Why Is Your AI Model Still Slow?
Why a 20-petaflop GB300 can still run slow: memory bandwidth, CPU-GPU links, storage, software and network bottlenecks, with figures to diagnose each.
Wei Fei
·
-
NVLink vs InfiniBand: Why NVIDIA Uses Both in the Same AI Factory
NVLink scale-up at 900GB/s per direction vs 800Gb/s InfiniBand scale-out in GB300 NVL72 AI factories: roles, switches, traffic and Ethernet options.
Wei Fei
·
-
72 GPUs, One AI Supercomputer: Why NVIDIA GB300 NVL72 Needs NVLink, Not Just More GPUs
How 18 NVSwitch chips give each of 72 Blackwell Ultra GPUs 1.8TB/s in one hop, what 130TB/s aggregate means, and why NVL72 needs NVLink.
Wei Fei
·
-
InfiniBand vs Ethernet: Are AI Data Centres Paying More for Performance They Don't Need?
Quantum InfiniBand vs Spectrum-X Ethernet for AI training and inference: switch costs per Tb/s, SHARP, RoCE tuning, market share and how to choose.
Wei Fei
·
-
Why Your Next AI Data Centre Needs More Than Powerful GPUs
NVLink, NVSwitch, InfiniBand, storage, power and liquid cooling: what a GB300 NVL72 AI data centre needs beyond GPUs, with Singapore capacity and costs.
Wei Fei
·
-
The Hidden Cost of AI Networking: Why Cables and Switches Can Make or Break Your GPU Investment
What DACs, 800G optics and switches cost in an AI cluster, how many a 1,024-GPU fabric needs, and how cable quality affects GPU uptime.
Wei Fei
·
-
PCIe 6.0 vs NVIDIA NVLink: Is Traditional GPU Connectivity Reaching Its Limits?
PCIe 6.0, 7.0 and 8.0 vs NVLink 5 and 6 per direction and bidirectional, first Gen6 products, UALink and what it means for multi-GPU workstations.
Wei Fei
·
-
37TB of Memory and 130TB/s NVLink: Is NVIDIA GB300 NVL72 the Ultimate AI Factory?
GB300 NVL72's 37TB memory, 130TB/s NVLink and 1.44EF FP4: model capacity, training, inference, 132-142kW power and whether it is the ultimate AI factory.
Wei Fei
·
-
Is NVLink NVIDIA's Biggest Competitive Advantage, Not Its GPUs?
NVLink and NVSwitch vs AMD Helios, Google TPU, AWS Trainium and UALink: the evidence for and against NVLink as NVIDIA's competitive moat in 2026.
Wei Fei
·
-
NVIDIA GB300 NVL72 vs 72 Individual GPUs: Why Connecting Them Changes Everything
NVL72 NVLink at 900GB/s per direction vs 8-GPU servers on 800G networks and PCIe GPUs: bandwidth, all-reduce time, power, reliability and when each fits.
Wei Fei
·
-
800Gbps InfiniBand vs 130TB/s NVLink: Why These Numbers Don't Tell the Whole Story
Convert 800Gb/s and 130TB/s to GB/s per GPU per direction: bits vs bytes, aggregate vs per GPU, bidirectional, bisection, NCCL busbw and latency.
Wei Fei
·
-
NVLink vs InfiniBand vs PCIe: Is Your AI Supercomputer Only as Fast as Its Weakest Connection?
NVLink 900GB/s, InfiniBand 100GB/s and PCIe Gen6 121GB/s per direction compared. When the slowest link really limits AI training and inference.
Wei Fei
·
-
RTX PRO 6000 Blackwell Uses PCIe 5.0. So Why Does NVIDIA GB300 Need NVLink?
Why the RTX PRO 6000 is fine on PCIe 5.0, when PCIe becomes the bottleneck, and why GB300 uses NVLink-C2C and an NVL72 NVLink fabric. Figures compared.
Wei Fei
·
-
Why NVIDIA's Most Powerful AI Systems Don't Rely on PCIe Alone
NVLink 5, NVLink-C2C and NVSwitch vs PCIe: bandwidth, coherence, 72-GPU switching and power. Where PCIe still fits in GB300 and DGX Station.
Wei Fei
·
Wei Fei
