Blackwell 在 DeepSeek R1 測試每兆瓦吞吐量 Hopper 50 倍

• GB300 NVL72 Configuration: 72 Blackwell Ultra GPUs + 36 Grace CPUs per rack system with 288GB memory per GPU (2x H200) • NVLink Bandwidth: 130 TB/s interconnects enabling efficient distributed inference across 72 GPUs[3] • FP4 Quantization: Blackwell's high-density NVFP4 FLOPs accelerate MoE forward pass compared to Hopper's FP8, enabling 50x throughput gains on DeepSeek-R1[2] • DeepSeek-R1 Performance Metrics: Single-GPU throughput of 7360 TGS (tokens/GPU/second) in prefill-only (ISL=2k); 2x GB300 achieves 22476 TGS in prefill and 3072 TGS in mixed-context (ISL=2k, OSL=1k, batch=256)[2] • Memory Optimization: DeepSeek MODEL1 (V4) uses tiered KV cache storage reducing GPU memory consumption by 40%, enabling inference on consumer RTX 4090 (24GB) with batch=4 at ~550 tokens/second[6] • Sparse MLA Operations: MODEL1 achieves 350 TFLOPS on B200 Blackwell GPUs vs 660 TFLOPS on H800 (SM90a), demonstrating Blackwell-specific optimization[4] • CUDA 12.9 Requirement: DeepSeek MODEL1 requires CUDA 12.9 to leverage cutting-edge Blackwell instruction sets and SM100-specific interfaces[4] • Prefill vs Decode Performance: GB300 shows 14% improvement over B300 in prefill-only scenarios; significant decode phase acceleration from doubled memory bandwidth and NVFP4 support[2]

Blackwell Ultra's 50x efficiency gains over Hopper establish a new cost-performance baseline for large-scale inference, making agentic AI and long-context reasoning economically viable at scale. The 35x cost reduction per million tokens directly impacts AI service provider margins and accessibility. DeepSeek's rapid optimization (2x improvement in 2 months) demonstrates competitive pressure on software stacks, while NVIDIA's Rubin roadmap (10x further gains by H2 2026) signals continued hardware-software co-design acceleration. The shift toward FP4 quantization and sparse attention mechanisms (1M+ token context windows) indicates future models will prioritize memory efficiency and long-context understanding. Adoption by Microsoft and CoreWeave validates production readiness, positioning Blackwell as the dominant inference platform for 2026. However, the threat landscape remains cost-per-performance focused rather than technological obsolescence—competitors like AMD are narrowing the gap through software optimization, while consumer GPU inference (RTX 5090 with Blackwell architecture) may democratize access to frontier models.