What Happened: A New World Record
NVIDIA, running on a CoreWeave AI Cloud cluster, achieved the #1 spot on the 31st Graph500 leaderboard with a record-breaking score of 410 trillion traversed edges per second (TEPS). This was done using 8,192 NVIDIA H100 GPUs processing an enormous graph with 2.2 trillion vertices and 35 trillion edges
💡 Why It Matters
Graph500 is the industry benchmark for large-scale graph analytics, especially breadth-first search (BFS). BFS simulates real-world workloads like social networks, cybersecurity analysis, and interconnected systems. The higher the TEPS, the better a system can traverse complex data structures quickly. NVIDIA Blog
💪 The Tech Behind the Beast
This wasn’t just throwing GPUs at a problem — it was a full-stack engineering play:
GPU-only active messaging: NVIDIA engineered an approach where GPUs communicate directly with each other over the network using InfiniBand GPUDirect Async (IBGDA) and NVSHMEM, bypassing the CPU bottleneck. This is major because traditional systems rely on CPUs to shuttle messages between GPUs, which kills performance at scale.
Massive parallelism: The H100 architecture excels at handling thousands of threads and gigabytes of data moving simultaneously. Combined with smart network infrastructure, it drives unreal throughput.
Efficiency win: The cluster used just ~1,000 nodes to beat other setups that needed around 9,000 nodes for similar tasks — that’s about 3× better cost and performance efficiency.
📊 Real-World Scale Example
To picture the scale: if every human on Earth had 150 friends, the graph of all those connections would be ~1.2 trillion edges — and this system could traverse that entire social graph in about 3 ms. That’s insane speed.
🧠 Broader Impact
This isn’t just a “benchmark flex.” It shows that GPU-driven architectures can outpace traditional CPU supercomputers for sparse, irregular workloads, which are notoriously harder than dense matrix computations (like typical AI stuff). Fields such as weather forecasting, fluid dynamics, cybersecurity, and any domain with massive relational data stand to benefit from this shift.
Views: 7