New Blackwell architecture with 20 PFLOPS FP4 and 192GB HBM3e, purpose-built for trillion-parameter training and high-throughput inference.
Several times the AI density of the previous generation at similar power.
20 PFLOPS FP4 accelerates pretraining and fine-tuning.
192GB HBM3e fits larger batches and longer context.
NVLink 1.8TB/s keeps multi-GPU overhead low.
Native liquid cooling for stable sustained loads.
Production-grade LLM workloads.
Needs large memory and high interconnect bandwidth.
Stable compute for consumer-scale concurrency.
High VRAM and compute for joint text/image/audio.
Blackwell lifts FP4/FP8 compute and memory bandwidth substantially for larger models and higher throughput.
Yes, via InfiniBand and NVLink for scalable training.
Pay-as-you-go and monthly/annual reservation.
Yes, GigaX offers heterogeneous scheduling per task.