Groq TSP AI Chip теперь питает Cloud Workloads через Nimbix

Groq tensor streaming processor (TSP) silicon добавлен в cloud, enabling customers accelerate AI workloads on demand. Nimbix теперь предлагает TSP-based machine-learning acceleration exclusively selected customers.

Конкурентный ландшафт

В competitive arena data-center AI silicon, Groq стоит alongside Graphcore как sole startups с commercially available accelerators integrated в cloud services. Graphcore chips уже offered через Microsoft Azure.

«Groq simplified processing architecture is unique, providing unprecedented deterministic performance для compute intensive workloads, exciting addition к our cloud-based AI и Deep Learning platform», — Steve Hebert, Nimbix CEO.

Groq только второй AI accelerator startup сделавший hardware available в cloud (Image: Groq)

Производительность TSP

Launched last fall, Groq TSP chip delivers staggering 1,000 TOPS (1 peta operations per second). В recent benchmarks chip achieved 21,700 ResNet-50 v2 inferences per second — more than twice speed contemporary GPU-based systems — underscoring status as one of fastest commercial neural-network processors.

«These ResNet-50 results validate that Groq unique architecture и approach to ML acceleration delivers substantially faster inference performance than competitors», — Jonathan Ross, Groq co-founder и CEO. «Real-world proof points based on industry-standard benchmarks, not simulations или hardware emulation, confirm measurable performance gains для ML и AI applications made possible by Groq technologies.»

Архитектурные преимущества

Groq design achieves massive parallelism needed для deep-learning acceleration без synchronization overhead typical CPUs и GPUs. Offloading control logic к compiler — hallmark Groq software-driven strategy — delivers fully deterministic, predictable performance quantifiable at compile time.

Standout advantage: Groq performance gains не depend on batching — common data-center tactic processing multiple samples simultaneously to boost throughput. TSP chip reaches peak performance даже at batch = 1, ideal для real-time inference streams. While chip offers moderate 2.5× latency advantage over GPUs at large batch sizes, advantage climbs to 17× at batch = 1.