Hardware & Compute
Chips, data centres and the infrastructure behind AI.
-
Power delivery failures in SambaNova SN40L rack deployments
Deploying SambaNova SN40L racks requires precise power and thermal validation to avoid instability. Common mistakes include insufficient PDU margin testing and overlooking…
-
Risks in Intel’s Falcon Shores delays for Argonne’s Aurora 2 upgrade
Intel’s shifting GPU roadmap and the transition to Falcon Shores create hardware gaps that threaten exascale deployments like Argonne National Laboratory’s Aurora…
-
Samsung Foundry and Tenstorrent build next-generation AI chiplets
Samsung Foundry manufactures Tenstorrent’s RISC-V Wormhole AI chiplets using the SF4X 4nm process. Led by Jim Keller, Tenstorrent aims to provide an…
-
Groq LPU inference pricing edge over NVIDIA H100s
Groq provides a cost advantage for Llama 3.3 70B workloads compared to NVIDIA H100s when using highly optimized hardware. While Cerebras offers…
-
SambaNova SN40L claims vs reality in enterprise LLM deployments
The SambaNova SN40L RDU excels at low-concurrency MoE workloads like Llama 4 Maverick, delivering 2,800 tokens per second. However, NVIDIA H200 and…
-
Physical constraints of AMD MI400 and OpenAI Stargate timeline
AMD faces significant supply chain risks for its MI400 series chips due to TSMC CoWoS packaging bottlenecks and HBM4 memory production limits.…
-
The economics of NVIDIA Blackwell Ultra pricing squeeze
Microsoft and Meta face massive capital requirements as NVIDIA ramps Blackwell Ultra infrastructure. Microsoft plans to spend $120 billion on AI infrastructure…
-
Thermal failures in the age of high-density AI
A voltage transient in the europe-west4-a data hall caused a chiller controller to fail, leading to a 14-hour disruption for 24 private…
-
Cerebras WSE-3 dominance in defense AI procurement
Cerebras WSE-3 offers a massive memory bandwidth advantage of 21 petabytes per second, outperforming Nvidia H100 by 7,000 times. This wafer-scale architecture…
-
Memory Bandwidth, Not FLOPs, Is What Limits LLM Inference
Accelerators are sold on peak FLOPs, but token generation is bound by how fast weights move out of memory. What to actually…
Trending Now
- How to use Harvey’s $715M AI legal assistant for Big Law workflows
- The economics of Devin and the shift from replacement to task delegation
- Perplexity valuation hits $3 billion as enterprise demand scales
- The Google DeepMind evolution and the $500B infrastructure race
- How Wayve’s embodied AI model scales autonomous vehicle partnerships