Search Articles

Search Results: NVLink

Coherent Fabrics Unpacked: Five Rules for CXL, NVLink, and InfinityFabric

Coherent Fabrics Unpacked: Five Rules for CXL, NVLink, and InfinityFabric

Memory‑hungry workloads are pushing the limits of traditional DRAM. A recent UC SDC study dissects how coherent fabrics—CXL, NVLink‑C2C, and AMD’s InfinityFabric—can extend capacity and bandwidth, and it distills five practical rules for developers and architects to harness these technologies without falling prey to latency traps.
AMD Achieves Milestone in Large-Scale MoE Pretraining with ZAYA1 on MI300X and Pollara

AMD Achieves Milestone in Large-Scale MoE Pretraining with ZAYA1 on MI300X and Pollara

A new study details the first large-scale mixture-of-experts (MoE) pretraining entirely on AMD hardware, introducing ZAYA1—a 760M active parameter model that rivals Qwen3-4B and Gemma3-12B. Packed with microbenchmarks for Pollara networking and MI300X optimizations, it offers actionable guidance for developers eyeing non-NVIDIA AI training stacks. This work underscores AMD's readiness for competitive foundation model development.
A Memcached for Attention: Inside the Cross-GPU KV Cache Marketplace for LLM Inference

A Memcached for Attention: Inside the Cross-GPU KV Cache Marketplace for LLM Inference

A new open-source experiment treats transformer KV caches as shareable, cross-GPU assets instead of per-request throwaways—turning redundant prefill into a distributed systems problem. For teams scaling chat, RAG, and multi-tenant LLM services, this is a glimpse of how inference infrastructure is about to get a lot more interesting.
AI Isn’t Hitting a Compute Wall. It’s Hitting a Network Wall.

AI Isn’t Hitting a Compute Wall. It’s Hitting a Network Wall.

As hyperscalers pour billions into AI, the real choke point isn’t just GPUs—it’s the fabric that connects them. From Nvidia and Broadcom to Lightmatter, Celestial AI, and PsiQuantum, a new arms race in photonic and custom interconnects is quietly deciding who will own the next decade of AI infrastructure.
Oracle's Zettascale Ambition: 18 ZettaFLOPS of Nvidia and AMD GPUs to Power AI Surge

Oracle's Zettascale Ambition: 18 ZettaFLOPS of Nvidia and AMD GPUs to Power AI Surge

Oracle announces an unprecedented 18 zettaFLOPS AI infrastructure rollout using 800,000 Nvidia Blackwell and 50,000 AMD MI450X GPUs, marking one of history's largest compute deployments. The clusters will leverage Nvidia's Spectrum-X networking and AMD's open UALink architecture, with OpenAI positioned as a key beneficiary. This massive scaling raises critical questions about practical FP4 utilization and the staggering energy demands of next-gen AI.