I/O Design Challenges Grow In AI Data Centers And HPC Clusters


Key Takeaways: A designer’s choice of I/O connectors and interconnect protocols can be the difference between a massively profitable AI chip and a flop. I/O tradeoffs impact airflow, cooling, rack design, power coming into the rack, and other critical aspects of HPC chip design. Reliability is paramount, so standards must be followed, and I/Os need redundant pins. Other innovations... » read more

How To Build Billions of Bumps


Key Takeaways: Hybrid bonding can result in a package containing billions (and eventually trillions) of connections. Building that many connections successfully requires extreme process uniformity across a wafer. Inspection isn’t practical, and test benefits from internal test mechanisms. Hybrid bonding allows unprecedented signal pitch, but fully populating dies and inter... » read more

Swapping Out Chiplets: I/Os Vs. Compute


Key Takeaways: Companies can save time and money by swapping out a compute, memory, or I/O chiplet to gain technology improvements, while keeping the other dies stable. Chip architects may choose to keep their I/Os stable and swap out compute to move from a 5nm process node to 3nm to achieve performance and power improvements, or swap out memory from LPDDR5X to LPDDR6. Swapping out... » read more

Using SystemC TLM Modeling To Solve AI Data Movement Challenges


In AI silicon, the performance numbers tell only part of the story. Marketing claims often highlight headline metrics such as trillions of operations per second, tensor throughput, matrix dimensions, and accelerator density. But engineers building these systems understand the harder truth. Compute performance matters only when data arrives at the right rate, with the right latency, and without ... » read more

Large-scale, SRAM-based LLM Inference Deployment (Groq)


A new technical paper, "SHIP: SRAM-Based Huge Inference Pipelines for Fast LLM Serving," was published by researchers at Nvidia, with work done while at Groq. Abstract "The proliferation of large language models (LLMs) demands inference systems with both low latency and high efficiency at scale. GPU-based serving relies on HBM for model weights and KV caches, creating a memory bandwidth b... » read more

Confusion Grows With More Interconnect Options And Tradeoffs


Key Takeaways: Designers are frequently evaluating 5 or more different interconnects in a single system, each with a distinct purpose. While chip-to-chip (PCIe) and die-to-die (UCIe, BoW) technologies seem to be solving a similar problem, in practice they bring different challenges. PCIe, CXL, NVLink, and UALink are all active in the hyperscaler space, but Ethernet-based technologies... » read more

Gates Add Functionality, But Wires Create Problems


Key takeaways: While transistors see continuous improvement, wires keep getting worse because of the smaller geometries and larger chip sizes. There are limited ways to avoid such problems, but the biggest impact will come from floorplanning. Analysis today is not adequate. New developments, such as backside power and 3D integration, provide temporary relief but new materials are a d... » read more

Facilitating Complex SoC Design Through Automation And Integration


The design demands of today’s highly advanced system-on-chip (SoC) devices have long outgrown the capabilities of manual workflows to manage them effectively. As these chips become more complex, only sophisticated, high-performance, and scalable automation can ensure that every component of the SoC functions seamlessly. The SoC integration challenge A fundamental aspect of SoC design is the... » read more

TSV Complexity Leads To Manufacturing Bottleneck


Key Takeaways: Through-silicon vias are the biggest enabler of 3D chip stacking and chip-to-PCB connections through silicon interposers. The AI boom is causing HBM and advanced assembly shortages, straining the supply chain. Optimization around etch, fill and reveal help reduce TSV cost. Through-silicon vias (TSVs) provide essential interconnects between DRAM dies inside hig... » read more

Chiplet Standards Aim For Plug-n-Play


Key Takeaways Die-to-die chiplet standards are only the beginning. Many more standards are necessary for a chiplet marketplace. A number of such standards have either had initial versions released or are in progress. Existing work covers packaging, a system architecture, various design kits, a universal link layer, and updates to BoW. Today’s chiplets exist in silos. In a ... » read more

← Older posts