NVIDIA Blackwell PUE Estimator
Calculates power usage effectiveness (PUE) for high-density liquid-cooled GPU racks and estimates annual energy costs and facility power draw.
NVIDIA Blackwell PUE Estimator
Assess thermal efficiency and operational costs for B200/GB200 cluster deployments.
B200 Thermal Management: NVIDIA's Blackwell architecture (GB200) can consume up to 1,200W per GPU. At $0.12/kWh, a single rack running at 80% utilization costs ~$60,000 annually just in electricity. Liquid cooling typically reduces PUE overhead from 1.6 to 1.15.
Configure your Blackwell rack specs to view thermal and cost projections.
Why Blackwell Changes the PUE Calculation Compared to Previous GPU Generations
PUE — Power Usage Effectiveness — has been the standard data centre efficiency metric since The Green Grid introduced it in 2007. For prior GPU generations including Hopper (H100) and Ampere (A100), the PUE calculation was straightforward: compute racks draw a predictable amount of power, and the facility overhead — cooling, power distribution losses, lighting — adds a multiplier on top of that IT load. A typical hyperscale facility running H100 clusters achieves PUE of 1.2 to 1.4 with precision air cooling. Blackwell changes this because the B200 GPU and GB200 NVL72 rack system have thermal design power figures that air cooling cannot handle at sustained load. The B200 SXM5 has a TDP of 1,000 watts per GPU. An NVL72 rack containing 72 B200 GPUs has a rack-level TDP of approximately 120 kilowatts — above the structural limit of standard raised-floor air cooling, which is typically rated at 15 to 30 kilowatts per rack. NVIDIA's NVL72 reference architecture requires direct liquid cooling to the rack manifold, with NVLink switch cooling also water-assisted. This means Blackwell deployments are not choosing between air and liquid cooling as equivalent options — they are working within a liquid-cooled architecture by default and choosing between different facility cooling plant configurations. PUE for Blackwell clusters is therefore a question of cooling plant efficiency: how efficiently does the facility's water cooling infrastructure — chillers, cooling towers, fluid distribution units, and heat rejection systems — convert the removed heat into facility overhead power draw.
What PUE, kWh, and CO2 the Tool Calculates and Where the Inputs Come From
The tool models three outputs from four inputs: rack count, IT load per rack in kilowatts, facility PUE, and regional electricity carbon intensity. IT load per rack is your measured or vendor-specified draw under sustained training workloads — the NVL72 reference is 120 kilowatts at full load, but throttled or inference-only deployments will be lower. PUE is your facility's actual or target PUE, which you can enter from your cooling plant vendor specifications or from measured facility PUE data if your data centre metering provides it. The tool calculates total facility power as IT load times rack count times PUE, annual kWh as total facility power in kilowatts times 8,760 hours, and annual CO2 emissions as annual kWh times the carbon intensity of your regional grid in kilograms per kWh. Regional presets use published grid average carbon intensity figures from IEA World Energy Outlook and EIA data: US average is approximately 0.386 kg CO2 per kWh, EU average is approximately 0.233 kg CO2 per kWh, and the tool includes region-specific presets for high-renewable grids (Nordic countries at approximately 0.027 kg CO2 per kWh) and carbon-intensive grids (Southeast Asia coal-heavy regions at approximately 0.55 to 0.70 kg CO2 per kWh). The CO2 output is relevant for TCFD climate disclosure, corporate net-zero commitments, and EU CSRD sustainability reporting, which increasingly require Scope 2 emissions from data centre operations to be disclosed at facility level.
Worked Example: 10-Rack NVL72 Cluster in Northern Virginia
Northern Virginia is the world's largest data centre market and the default location for many US hyperscale GPU deployments. Consider a 10-rack NVL72 Blackwell cluster. Each rack draws 120 kilowatts at sustained training load, giving a total IT load of 1,200 kilowatts. A modern liquid-cooled facility in the Northern Virginia market achieves PUE of 1.15 to 1.20 — this is achievable with a chilled water plant optimised for the local climate, which has moderate temperatures for approximately 7 months of the year allowing economiser operation. At PUE 1.18, total facility power is 1,200 times 1.18 equals 1,416 kilowatts. Annual energy consumption is 1,416 times 8,760 equals 12,404,160 kilowatt-hours, or approximately 12.4 million kWh per year. The US average grid carbon intensity for the PJM interconnect serving Northern Virginia is approximately 0.35 kg CO2 per kWh. Annual CO2 emissions are 12.4 million times 0.35 equals approximately 4,340 tonnes of CO2 equivalent per year for this cluster. For comparison, the same 10 racks in a Nordic facility using hydroelectric power at 0.027 kg CO2 per kWh would produce approximately 335 tonnes CO2 equivalent annually — a 13-fold reduction in emissions footprint from the same hardware running the same workloads, purely from grid carbon intensity difference.
Frequently Asked Questions
Commonly Asked Questions
- What is the thermal design power of the NVIDIA B200 and NVL72?
- The NVIDIA B200 SXM5 GPU has a thermal design power of approximately 1,000 watts per GPU under sustained compute load. The GB200 NVL72 rack unit — containing 36 Grace CPUs and 72 B200 GPUs connected by NVLink — has a rack-level TDP of approximately 120 kilowatts including the NVLink switch fabric and supporting infrastructure. This figure assumes sustained AI training workload at maximum utilisation. Inference workloads at lower batch sizes produce lower actual power draw, typically 60 to 80 percent of TDP. NVIDIA publishes thermal and power specifications in the NVL72 system datasheet and Data Center Design Guide, which should be used for facility planning rather than estimates.
- Can Blackwell GPU clusters be deployed with air cooling?
- Not at full rack density. The NVL72 form factor at 120 kilowatts per rack exceeds the structural cooling capacity of conventional raised-floor precision air cooling, which is rated at 15 to 30 kilowatts per rack in standard configurations and up to 40 to 50 kilowatts per rack in high-density air configurations with in-row cooling. Air cooling is used for the residual heat — approximately 10 to 15 percent of total rack power — that direct liquid cooling does not capture. The NVLink switch and some peripheral components in the NVL72 are air-cooled. Full deployment of Blackwell at reference density requires a facility with a chilled water plant and in-rack liquid cooling manifolds capable of handling the 120 kilowatt per rack primary load.
- What PUE is achievable for a Blackwell liquid-cooled data centre?
- State-of-the-art liquid-cooled AI data centres built for Blackwell-class density are reporting design PUE targets of 1.10 to 1.20. The lower bound of 1.10 requires near-ideal conditions: a cool climate allowing free cooling or economiser operation for most of the year, a high-efficiency chilled water plant with variable-speed drives, and minimal power distribution losses from high-voltage DC or 415V three-phase distribution. A more typical achievable PUE for a well-designed facility in a temperate climate is 1.15. Facilities in warmer climates without access to economiser cooling will typically achieve 1.20 to 1.30. PUE below 1.10 is possible only with waste heat reuse — for example, district heating connections — that converts removed heat into useful output counted against the facility overhead.
- Why does grid carbon intensity vary so much between regions and why does it matter?
- Grid carbon intensity reflects the mix of generation sources on the electricity grid — coal, gas, nuclear, hydro, wind, solar — weighted by their CO2 emissions per kWh generated. Nordic grids dominated by hydroelectric power have intensities below 0.03 kg CO2 per kWh. French grids dominated by nuclear power are approximately 0.05 kg CO2 per kWh. US grids average 0.35 to 0.45 kg CO2 per kWh depending on the regional interconnect, with coal-heavy southeastern grids above 0.45 and California below 0.25. For AI infrastructure sustainability reporting under TCFD and EU CSRD, the location of GPU clusters is a primary determinant of Scope 2 emissions — the same training workload produces 10 to 20 times more CO2 in a coal-heavy grid region than in a renewable-heavy one. This is why several major AI labs have announced GPU deployments in Nordic countries specifically to reduce the carbon footprint of training runs.
- What is the difference between market-based and location-based Scope 2 accounting for data centres?
- Location-based Scope 2 accounting uses the average grid carbon intensity for the region where electricity is consumed — the approach this tool uses by default. Market-based accounting allows organisations to subtract the carbon intensity of renewable energy certificates (RECs) or power purchase agreements (PPAs) purchased from specific renewable generators, potentially reducing reported Scope 2 emissions to near zero if 100 percent renewable PPAs are in place. Under GHG Protocol Corporate Standard and TCFD guidance, both methods must be disclosed. For data centre operators with large renewable PPA portfolios — Google, Microsoft, Amazon — market-based Scope 2 can be significantly lower than location-based. The tool calculates location-based emissions, which is the baseline figure required for disclosure before any PPA credits are applied.
- How do I use this tool's output for CSRD or TCFD sustainability reporting?
- The EU Corporate Sustainability Reporting Directive (CSRD) and TCFD recommendations both require disclosure of Scope 1 and Scope 2 greenhouse gas emissions. For data centre operators, Scope 2 emissions from purchased electricity are typically the dominant source. This tool produces the annual kWh consumption and annual CO2 equivalent figure needed to populate the Scope 2 location-based emissions line in a GHG Protocol-aligned disclosure. The inputs — rack count, IT load, PUE, and grid intensity — should be documented as methodology assumptions in the disclosure. For CSRD specifically, the European Sustainability Reporting Standard ESRS E1 requires disclosure of energy intensity metrics including kWh per unit of output and total energy consumption by source, making the kWh output of this tool directly relevant to the mandatory reporting template.