NVLink

Event Core In a landmark performance benchmark on the NVIDIA GB300 NVL72 system, the 2.4-trillion-parameter Qwen 3.8 model achieved a staggering total throughput of 288,000 tokens per second in FP8 precision. The system demonstrated a per-GPU rate exceeding 4k tokens/s and a per-user latency profile of over 350 tokens/s without requiring additional fine-tuning. ▶ Interconnect Revolution: The GB300 NVL72 architecture, featuring 72 fully interconnected Blackwell GPUs, effectively eliminates the inter-node communication bottlenecks previously inherent in trillion-parameter model inference. ▶ FP8 Production Standard: High-performance inference at FP8 precision "out-of-the-box" signals that ultra-LLMs have transitioned from experimental feasibility to industrial-scale high-concurrency deployment. Bagua Insight The leak of these performance metrics effectively declares the end of the "inference wall" for trillion-parameter models. Previously, serving a 2.4T model (comparable to rumored GPT-4 scales) was plagued by prohibitive latency and memory fragmentation. However, the fifth-generation NVLink on the GB300 NVL72 treats 72 GPUs as a single, massive compute entity. Crucially, a per-user speed of 350 tokens/s far exceeds human reading capabilities (approx. 5-10 tokens/s). This excess compute will inevitably be channeled into more complex Chain-of-Thought (CoT) reasoning or real-time Multi-agent orchestration. Qwen 3.8’s performance on this stack proves that top-tier Chinese models are achieving world-class optimization within the Blackwell ecosystem. The center of gravity in the AI arms race is shifting from raw VRAM capacity to the "interconnect bandwidth" paradigm. Actionable Advice Infrastructure Strategy: Enterprises targeting real-time responsiveness for 1T+ MoE or dense models should prioritize the TCO advantages of the GB300 NVL72 over fragmented H100 clusters. Optimization Focus: With per-GPU throughput hitting 4k+ tokens/s, developers must shift their focus from kernel-level acceleration to sophisticated KV Cache management and high-concurrency scheduling to fully saturate Blackwell’s pipeline. Model Roadmap: Given the maturity of lossless FP8 inference, pre-training regimes should integrate FP8-native compatibility to ensure a seamless transition from training to production-grade serving.