NVIDIA HGX B200 Reduces Embodied Carbon Emissions Intensity

NVIDIA HGX B200 relies on upgraded NVIDIA Blackwell B200 GPUs, that are designed to supply dramatically improved AI performance. B200 GPUs incorporate 180 GB of HBM3E memory—greater than double the memory of the NVIDIA HGX H100 and other performance enhancing features, including a second-generation Transformer Engine that introduces FP4 alongside FP8, fifth‑generation NVLink/NVSwitch (as much as 1.8 TB/s per‑GPU and 14.4 TB/s aggregate bandwidth), to spice up higher throughput at lower precision.

Throughput of the HGX B200 is 2.3x faster (FP16) than the HGX H100. This upgraded compute performance can be more energy efficient. For AI inference, the HGX B200 could be as much as 15x more energy efficient. That’s a 93% reduction in energy for a similar inference workload.

HGX B200 also generates less manufacturing-related emissions per FLOPS of compute. Specifically, the embodied carbon intensity for computing is reduced from 0.66 gCO2e per exaflop with HGX H100 to 0.50 gCO2e per exaflop with HGX B200 (estimated based on FP16 precision). That is an overall 24% decrease between the 2 generations. That is estimated based on PCF data and the whole variety of FLOPS at FP16 precision executed over the lifetime of the product.

Moreover, there’s a decrease within the materials and components—the very best emission producing category for every baseboard—between HGX B200 and HGX H100. Essentially the most notable reductions are for thermal components, ICs, and memory. 

The mixture of the HGX B200 energy-saving computational power and reduced embodied emissions intensity can result in a considerable lowering of emissions, especially during its use phase for giant workloads reminiscent of AI training and inference. 

Operational carbon impacts 

While this reduced embodied carbon intensity reflects a lower upstream carbon intensity for the HGX B200, the downstream carbon intensity improvement is much more pronounced. 

For instance, the HGX B200 is projected to deliver a 10x improvement in inference efficiency for the DeepSeek-R1 model, which translates to a 90% reduction in operational carbon emissions in comparison with the HGX H100 for operational carbon emissions gCO2e for processing 1 million inference tokens (100 TPS/user). 

Note that this data was calculated based on 2023 IEA emission aspects weighted by regional data center energy consumption. Emission aspects include upstream emissions and T&D losses related to electricity production. 

Methodology and data collection process 

Each of the externally published PCF summaries rely heavily on primary data from suppliers collected for over 90% of each products by weight, including material composition and production energy consumption. 

Secondary sources were integrated with this data, reminiscent of the imec.netzero tool for fabrication-related emissions, in addition to the ecoinvent 3.10 and Sphera LCA databases (Skilled Database 2024 and Extension Database XI: Electronics 2024) for modeling materials, transportation, and energy.

These PCF summaries are aligned with ISO Standard14040 and 14044 on life cycle assessments and were critically reviewed in conformance with ISO Standard 14067 on carbon footprints.

The longer term of sustainable computing

NVIDIA goals to diminish its product carbon footprint with each latest product it produces, while providing groundbreaking advancements in performance. This practice in transparency through detailed PCF summaries will improve understanding of the impacts of accelerated computing.

NVIDIA is committed to working toward publishing additional reliable data on the environmental impacts of NVIDIA products. NVIDIA will proceed to innovate towards a way forward for sustainable computing and AI development, without compromising on performance and scale. 

To learn more, read the Product Carbon Footprint Summary for NVIDIA HGX B200.