TickerTain — Make Analysis Fun Again

←

▶ 22:00 · Memory & Storage · New agentic memory tier (CMX for KV cache, STX for vector/object storage) delivers 5x inference throughput

now playing · Memory & Storage

AI Infrastructuretailwindscore 9/10kevin deierling

AI factories replace data centers as token revenue generators requiring full-stack co-design

Data centers are being re-architected from cost centers into 'AI factories' that generate tokens as revenue; this requires extreme co-design across energy, chips, infrastructure, models, an…

Grid & Power Infrastructuretailwindscore 8/10kevin deierling

DSX power smoothing enables 40% more GPU density per gigawatt by flattening training power spikes

AI training creates rainfall-like power spikes that stress grid transformers; Nvidia's DSX system stores excess energy during low-demand phases and releases it during peaks, flattening the…

AI Hardware & Chip Architecturetailwindscore 9/10kevin deierling

Seven co-designed chips (Rubin, Vera, NVLink, Spectrum, ConnectX, BlueField, CPO) act as one system to beat copper distance limits

Single-chip design is insufficient; Nvidia co-designs seven chips together with cache-coherent links (NVLink 72), scale-out fabric (Spectrum-6), and co-packaged optics (Spectrum CPO) to ove…

Networking & Optical Infrastructuretailwindscore 8/10kevin deierling

Co-packaged optics (Spectrum CPO) eliminates PCB signal degradation, saving tens of megawatts per AI factory

Running electrical signals across PCBs to front-panel optics degrades signal integrity; placing optics directly next to the ASIC inside the package preserves signal quality and cuts power d…

AI Agentstailwindscore 9/10kevin deierling

Agentic loops create exponential token demand, requiring purpose-built CPU (Vera) and storage architectures

Agents close the prompt loop autonomously, generating 100x more internal tokens per output token; this demands CPUs built for planning/tool use (Vera), accelerated libraries (CUDA-X), and n…

Memory & Storagetailwindscore 8/10kevin deierling

New agentic memory tier (CMX for KV cache, STX for vector/object storage) delivers 5x inference throughput

AI agents require short-term KV cache and long-term vector/object memory distinct from traditional DRAM; Nvidia's CMX/STX reference architecture accelerates prefill/decode and tool use, yie…

←

▶ 22:00 · Memory & Storage · New agentic memory tier (CMX for KV cache, STX for vector/object storage) delivers 5x inference throughput

now playing · Memory & Storage

AI Infrastructuretailwindscore 9/10kevin deierling

AI factories replace data centers as token revenue generators requiring full-stack co-design

Data centers are being re-architected from cost centers into 'AI factories' that generate tokens as revenue; this requires extreme co-design across energy, chips, infrastructure, models, an…

Grid & Power Infrastructuretailwindscore 8/10kevin deierling

DSX power smoothing enables 40% more GPU density per gigawatt by flattening training power spikes

AI training creates rainfall-like power spikes that stress grid transformers; Nvidia's DSX system stores excess energy during low-demand phases and releases it during peaks, flattening the…

AI Hardware & Chip Architecturetailwindscore 9/10kevin deierling

Seven co-designed chips (Rubin, Vera, NVLink, Spectrum, ConnectX, BlueField, CPO) act as one system to beat copper distance limits

Single-chip design is insufficient; Nvidia co-designs seven chips together with cache-coherent links (NVLink 72), scale-out fabric (Spectrum-6), and co-packaged optics (Spectrum CPO) to ove…

Networking & Optical Infrastructuretailwindscore 8/10kevin deierling

Co-packaged optics (Spectrum CPO) eliminates PCB signal degradation, saving tens of megawatts per AI factory

Running electrical signals across PCBs to front-panel optics degrades signal integrity; placing optics directly next to the ASIC inside the package preserves signal quality and cuts power d…

AI Agentstailwindscore 9/10kevin deierling

Agentic loops create exponential token demand, requiring purpose-built CPU (Vera) and storage architectures

Agents close the prompt loop autonomously, generating 100x more internal tokens per output token; this demands CPUs built for planning/tool use (Vera), accelerated libraries (CUDA-X), and n…

Memory & Storagetailwindscore 8/10kevin deierling

New agentic memory tier (CMX for KV cache, STX for vector/object storage) delivers 5x inference throughput

AI agents require short-term KV cache and long-term vector/object memory distinct from traditional DRAM; Nvidia's CMX/STX reference architecture accelerates prefill/decode and tool use, yie…