Darius Baruar
June 4, 2025 12:59
NVIDIA’s BLACKWELL Architecture shows the ability to revolutionize AI applications by achieving unprecedented performance on various AI workloads on the latest MLPERF training benchmark.
According to NVIDIA’s blogs, NVIDIA’s Blackwell Architecture showed excellent performance on a variety of AI workloads on the most recent MLPERF training benchmarks. The 12th repetition since its founding in 2018, the benchmark emphasized the function of the architecture that handles large language models (LLM) and recommended systems.
Record performance on the benchmark
The NVIDIA AI platform provided major performance in all benchmarks, including challenging LLAMA 3.1 405B free train tests. This achievement emphasizes the diversity and excellent features of the platform in large AI operations. This platform was the only platform that submitted the results from all MLPERF training V5.0 benchmarks and shows a wide range of applications.
Advanced infrastructure and collaboration
NVIDIA has achieved these results using two AI supercomputers, Tyche and NYX, driven by the BLACKWELL platform. Tyche is made of NVIDIA GB200 NVL72 rack scale system and NYX is based on the NVIDIA DGX B200 system. Cooperation with COREWEAVE and IBM improved performance using 2,496 Blackwell GPUs and 1,248 NVIDIA Grace CPUs.
Significant performance gain
LLAMA 3.1 405b free train benchmarks have seen Blackwell achieves 2.2 times larger performance than previous architectures. Similarly, the NVIDIA DGX B200 system, equipped with eight BLACKWELL GPUs, provides 2.5 times more performance in LLAMA 2 70b LORA micro -adjustment benchmark, indicating the main development of AI education efficiency.
Technology innovation
The improvement of the Blackwell Architecture includes high -density liquid cooling rack, 13.4TB of Coherent Memory and Advanced NVIDIA NVLINK and NVLINK InterConnect technology. Such development promotes the crucial scale up and scale -out operation for the next -generation multimodal LLM education and agent AI applications.
AI ecosystem expansion
NVIDIA’s data center platform integrates a wide range of software products, including GPUs, CPUs, high-speed networking and CUDA-X libraries, NEMO frameworks and NVIDIA TENSORT-LLM. This ecosystem accelerates AI model training and distribution to reduce the market and improve value creation.
The MLPERF Round has seen a wide range of participation from the NVIDIA’s partner ecosystem, with the submission of a company, including ASUS, CISCO, DELL TECHNOLOGIES, and Google Cloud. This collaboration efforts emphasizes that the importance of AI is increasing in various industries.
Image Source: Shutter Stock