AMD co-invented High Bandwidth Memory in 2011 and shipped the first product. Today Lisa Su flies to Seoul to beg for supply. Why patents and standards don't make chips — and what AMD actually gave away.
Read Full Analysis →AMD co-invented High Bandwidth Memory in 2011 and shipped the first product. Today Lisa Su flies to Seoul to beg for supply.…
AMD's Helios racks need 31TB of HBM4 per system, SK Hynix is locked to NVIDIA — and Samsung knows it. The memory…
Intel's Crescent Island is now the GPU D-Series: 1.3 PFLOPS of FP8, up to 480GB of LPDDR5X, and 350W — air-cooled. The…
NVIDIA's custom HBM moves the memory controller onto the base die — and now Micron and even Samsung are outsourcing that die…
Micron marked both of its 2GB GDDR7 parts End of Life — the memory used across NVIDIA RTX 50 desktop cards —…
Samsung Q3 guidance projects 107.4 trillion won (~$80B) operating profit, up 782% YoY with a 55% margin — beating NVIDIA's record quarter.…
Agent frameworks are competing on plugins and protocol support, not model quality. The Hermes Agent extension stack and MCP adoption show the…
TrendForce sees AI datacenter CPU-to-GPU ratios collapsing from 1:8 toward 1:1 as agents need orchestration, not just inference. NVIDIA and Arm launched…
ByteDance, Alibaba, and DeepSeek are all targeting 5-10 trillion parameter models. Mixture-of-Experts is why that's economically possible — and a new tier…
AMD's ~10% hike on Instinct, Radeon, and chipsets took effect October 1; Intel's reportedly lands October 5. Neither company confirmed either move,…
Synchronized with VLSI Symposia, ISSCC, & independent telemetry logs
