NVIDIA Blackwell + Together AI accelerates training for Zoom, Salesforce, InVideo
AI Impact Summary
NVIDIA Blackwell GPUs (HGX B200) deployed via Together AI significantly boost training and inference workloads for Zoom, Salesforce, and InVideo. Salesforce reports ~2x training speed for mistral-24B when upgrading from HGX H200 to B200, Zoom achieves 1.9x training speed gains versus Hopper GPUs, and InVideo sees an initial 25% uplift moving from H200 to B200 with further optimizations that more than double performance. The Together Training Stack, including a co-optimized Llama 3 70B model and optimizations like FlashAttention-3, FP8, and improved DDP/torch.compile workflows, demonstrates a clear path to higher MFU and tokens/sec per node, enabling faster model iterations and feature delivery. This indicates a viable migration target from H200 to B200 for large-model workloads and stronger price-performance for AI workloads across these customers.
Affected Systems
- Date
- Date not specified
- Change type
- capability
- Severity
- info