Hardware & Compute
Chips, data centres and the infrastructure behind AI.
-
Tenstorrent Blackhole and Wormhole architecture explained
Tenstorrent’s RISC-V based architecture utilizes Tensix cores and explicit DMA data movement to provide an alternative to NVIDIA. The Blackhole p150 developer…
-
SambaNova SN50 vs Cerebras WSE-3 for large scale inference
SambaNova SN50 delivers 5x the maximum speed and 3x the throughput for agentic inference compared to Blackwell B200 GPUs. Cerebras WSE-3 provides…
-
Blackwell supply bottlenecks and CoreWeave financial exposure
NVIDIA’s reservation of up to 850,000 CoWoS wafers for 2026 captures 60% of TSMC’s capacity, creating significant supply risks for CoreWeave’s $22.4…
-
Why poor rack planning kills H100 training uptime
High-density AI infrastructure requires 15 to 25 kilowatts per rack, often exceeding the 5 to 8 kilowatt capacity of traditional data centers.…
-
Why Cerebras Condor Galaxy 3 matters for national lab AI budgets
Cerebras and G42’s Condor Galaxy 3 provides 8 exaflops of AI compute using WSE-3 technology. This massive scaling capability arrives as the…
-
Thermal management failures in Azure AI clusters
Inadequate air cooling for high-density chips can cause power throttling exceeding 15 percent of thermal design power. This analysis examines liquid cooling…
-
AMD MI350X memory bandwidth and CDNA 4 architecture explained
The AMD MI350X features 288 GB of HBM3E memory and 8.0 TB/s bandwidth, allowing it to run 400B+ parameter models using only…
-
Qualcomm targets Tenstorrent in AI chip consolidation
Qualcomm is discussing an acquisition of Jim Keller-led Tenstorrent, valued at $2.7 billion, as the company uses RISC-V architecture to challenge NVIDIA.…
-
Cerebras’s $50B Valuation versus SambaNova’s Hardware Efficiency
Cerebras maintains a $50.5 billion valuation with significant revenue growth, while SambaNova’s SN40L delivers 10X better performance per area than Cerebras on…
-
The 120kW Wall in Equinix xScale Facilities
Deploying NVIDIA GB200 NVL72 racks presents massive challenges, including 120kW power draws and 3,000kg weights that exceed standard data center floor limits.…
-
Risks in SambaNova’s supply chain for Saudi Arabia’s AI cloud push
SambaNova faces significant supply chain hurdles and market consolidation threats as Saudi Arabia pursues a $40B sovereign AI cloud project. Key risks…
-
Groq LPU performance versus Llama 4 deployment realities
Groq claims a 10x speed advantage over H100 clusters for Llama 2 70B, but hardware costs for Groq runs are 40x higher…