NVIDIA Advances Agentic AI with Gemma 4 Collaboration, NIM 2x Throughput Gains, and Neural Texture Compression

NVIDIA made multiple announcements this week: in collaboration with Google, it released the Gemma 4 model family optimized for NVIDIA hardware across devices from smartphones to IoT systems, including Gemma's first MoE model for agentic, on-device AI with local real-time data processing to reduce latency and cloud dependency. NVIDIA NIM (Inference Microservices) achieved 2x throughput improvements on H100 GPUs — benchmarked at 1,201 tokens/second versus 613 without NIM on Llama 3.1 8B — supporting DeepSeek, Llama, Mistral, and SDXL across cloud, data center, and PC environments. Additionally, NVIDIA demonstrated Neural Texture Compression (NTC) reducing VRAM usage from 6.5GB to 970MB (an ~85% reduction), positioning AI-driven compression as a complement to DLSS 5 focused on efficiency rather than image reconstruction.