Chile’s NLHPC supercharged research performance with AMD
The National Laboratory for HPC in Chile doubled its scientific research performance per watt with AMD EPYC™ CPUs and AMD Instinct™ GPUs.
Advancing Data Center Sustainability
At AMD, energy efficiency has long been a guiding core design principle aligned to our roadmap and product strategy. For more than a decade, we have set public, time-bound goals to dramatically increase energy efficiency across the breadth of our portfolio and have consistently met and exceeded those targets.
As AI continues to scale, and as we move toward true end-to-end design of full AI systems, the need for innovative energy solutions is becoming increasingly important – perhaps nowhere more so than in the data center.
AMD 30x25 Energy Efficiency Goal
Our goal is to deliver a 30x increase in energy efficiency for AMD processors and accelerators powering servers for AI-training and HPC from 2020-2025.1 These important and growing computing segments have some of the most demanding workloads. This goal represents more than a 2.5x acceleration of the industry trends from 2015-2020 as measured by the worldwide energy consumption for these computing segments.2
Our 30x energy efficiency goal equates to a 97% reduction in energy use per computation from 2020-2025. If all AI and HPC server nodes globally were to make similar gains, billions of kilowatt-hours of electricity could be saved in 2025 relative to baseline trends.
As of mid-2025, we have achieved a 38x3 improvement over the base system using a configuration of four AMD Instinct™ MI355X GPUs and one AMD EPYC™ 5th Gen CPU. Our progress report utilizes methodology2 validated by renowned compute energy efficiency researcher and author, Dr. Jonathan Koomey.
AMD 20x Rack-Scale Energy Efficiency Goal for AI Systems
As workloads scale and demand continues to rise, node-level efficiency gains won't keep pace. The most significant efficiency impact can be realized at the system level. That’s why we have set our sights on a bold new target: a 20x improvement in rack-scale energy efficiency for AI training and inference by 2030, from a 2024 base year.4
We believe we can achieve 20x increase in rack-scale energy efficiency for AI training and inference from 2024 by 2030, which AMD estimates exceeds the industry improvement trend from 2018 to 2025 by almost 3x. This reflects performance-per-watt improvements across the entire rack, including CPUs, GPUs, memory, networking, storage and hardware-software co-design, based on our latest designs and roadmap projections. This shift from node to rack is made possible by our rapidly evolving end-to-end AI strategy and is key to scaling datacenter AI in a more sustainable way.
Our 20x goal reflects what we control directly: hardware and system-level design. But we know that even greater delivered AI model efficiency gains will be possible, up to 5x over the goal period, as software developers discover smarter algorithms and continue innovating with lower-precision approaches at current rates. When those factors are included, overall energy efficiency for training a typical AI model could improve by as much as 100x by 2030.5
Environmental Benefits
A 20x rack-scale efficiency improvement at nearly 3x the prior industry rate has major implications. Using training of a typical AI model in 2025 as a benchmark, the gains could enable:6
- Rack consolidation from more than 275 racks to <1 fully utilized rack
- More than a 95% reduction in operational electricity use
- Carbon emission reduction from approximately 3,000 to 100 metric tCO2 for model training
These projections are based on AMD silicon and system design roadmap and a measurement methodology validated by energy-efficiency expert Dr. Jonathan Koomey.
Industry Perspectives
Case Studies
Supporting Resources
- Learn more about AMD EPYC Energy Efficiency
- Learn more about the AMD Instinct accelerators
- Learn more about the AMD EPYC processor Total Cost of Ownership (TCO) calculator
- Learn more about AMD Corporate Responsibility
Footnotes
- Includes high-performance CPU and GPU accelerators used for AI training and High-Performance Computing in a 4-Accelerator, CPU hosted configuration. Goal calculations are based on performance scores as measured by standard performance metrics (HPC: Linpack DGEMM kernel FLOPS with 4k matrix size; AI training: lower precision training-focused floating-point math GEMM kernels operating on 4k matrices) divided by the rated power consumption of a representative accelerated compute node including the CPU host + memory, and 4 GPU accelerators.
- Based on 2015-2020 industry trends in energy efficiency gains and data center energy consumption in 2025.
- EPYC-030B: AMD takes compute node performance/watt measurements for AMD high performance CPU and GPU accelerators used for AI training and High-Performance Computing in a 4-Accelerator, CPU-hosted configuration.
- Performance for HPC workloads is based on Linpack DGEMM kernel FLOPS with 4k matrix size. Performance for AI training is based on lower precision training-focused floating-point math GEMM kernels operating on 4k matrices.
- Watts are based on the TDP of a representative accelerated compute node including the CPU host + memory, and 4 GPU accelerators
To make the goal particularly relevant to worldwide energy use, AMD worked with Koomey Analytics to assess available research and data that includes segment-specific datacenter power utilization effectiveness (PUE), including GPU HPC and machine learning (ML) installations. The AMD CPU socket and GPU node power consumptions incorporate segment-specific utilization (active vs. idle) percentages and are multiplied by PUE to determine actual total energy use for calculation of the performance per watt.
The energy consumption baseline uses the same industry energy per operation improvement rates as were observed from 2015-2020, with this rate of change extrapolated to 2025. The AMD goal trend line (Table 1) shows the exponential improvements needed to hit the goal of 30-fold efficiency improvements by 2025. The actual AMD products released (Table 2) are the source of the efficiency improvements shown for AMD goal status in Table 1.
The measure of energy per operation improvement in each segment from 2020-2025 is weighted by the projected worldwide volumes (as per IDC - Q1 2021 Tracker Hyperion - Q4 2020 Tracker, Hyperion HPC Market Analysis, April ’21). Translating these volumes to the ML training and HPC markets results in node volumes as per Table 3 below. These volumes are then multiplied by the Typical Energy Consumption (TEC) of the respective computing segment in 2025 (Table 4) to arrive at a meaningful aggregate metric of actual energy usage improvement worldwide.
Table 1: Summary efficiency data projected to 2025
2020
2021
2022
2023
2024
2025
Goal Trend Line
1.00
1.97
3.98
7.70
15.20
30.00
AMD Goal Status (energy-weighted performance / watt)
1.00
3.90
6.79
13.49
28.29
37.85
Table 2: AMD Products
2020
2021
2022
2023
2024
2025
EPYC Gen 1 CPU + M50 GPU
EPYC Gen 2 CPU + MI100 GPU
EPYC Gen 3 CPU + MI250 GPU
MI300A APU (4th Gen AMD EPYC™ CPU with AMD CDNA™ 3 Compute Units)
EPYC Gen 5 CPU + MI300X GPU
EPYC Gen 5 CPU + MI355X GPU
*AMD Products are supported by the latest software, including AMD ROCm.
Table 3: Volume Projections (millions/yr)
2020
2021
2022
2023
2024
2025
HPC GPU nodes sold
0.05
0.06
0.07
0.09
0.10
0.12
ML GPU nodes sold
0.09
0.10
0.12
0.14
0.17
0.20
Table 4: Base case 2025 electricity consumption of products sold in that year, for weighting efficiency indices (TWh/year)
2025
Base HPC
4.49
Base ML
29.79
Total Base
34.28
* Estimated values for 2025 on worldwide energy use are updated annually as HPC and ML compute node capabilities evolve from our original outlook, including with the growth of AI increasing the weighting for ML performance.
- AMD based advanced racks for AI training/inference in each year (2024 to 2030) based on AMD roadmaps, also examining historical trends to inform rack design choices and technology improvements to align projected goals and historical trends. The 2024 rack is based on the MI300X node, which is comparable to the Nvidia H100 and reflects current common practice in AI deployments in 2024/2025 timeframe. The 2030 rack is based on an AMD system and silicon design expectations for that time frame. In each case, AMD specified components like GPUs, CPUs, DRAM, storage, cooling, and communications, tracking component and defined rack characteristics for power and performance. Calculations do not include power used for cooling air or water supply outside the racks but do include power for fans and pumps internal to the racks.
FLOPS
HBM BW
Scale-up BW
Training
70.0%
10.0%
20.0%
Inference
45.0%
32.5%
22.5%
- Regression analysis of achieved accuracy/parameter across a selection of model benchmarks, such as MMLU, HellaSwag, and ARC Challenge, show that improving efficiency of ML model architectures through novel algorithmic techniques, such as Mixture of Experts and State Space Models for example, can improve their efficiency by roughly 5x during the goal period. Similar numbers are quoted in Patterson, D., J. Gonzalez, U. Hölzle, Q. Le, C. Liang, L. M. Munguia, D. Rothchild, D. R. So, M. Texier, and J. Dean. 2022. "The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink." Computer. vol. 55, no. 7. pp. 18-28.” Therefore, assuming innovation continues at the current pace, a 20x hardware and system design goal amplified by a 5x software and algorithm advancements can lead to a 100x total gain by 2030.
- AMD estimated the number of racks to train a typical notable AI model based on EPOCH AI data (https://epoch.ai). For this calculation we assume, based on these data, that a typical model takes 1025 floating point operations to train (based on the median of 2025 data), and that this training takes place over 1 month. FLOPs needed = 10^25 FLOPs/(seconds/month)/Model FLOPs utilization (MFU) = 10^25/(2.6298*10^6)/0.6. Racks = FLOPs needed/(FLOPS/rack in 2024 and 2030). The compute performance estimates from the AMD roadmap suggests that approximately 276 racks would be needed in 2025 to train a typical model over one month using the MI300X product (assuming 22.656 PFLOPS/rack with 60% MFU) and <1 fully utilized rack would be needed to train the same model in 2030 using a rack configuration based on an AMD roadmap projection. These calculations imply a >276-fold reduction in the number of racks to train the same model over this six-year period. Electricity use for a MI300X system to completely train a defined 2025 AI model using a 2024 rack is calculated at ~7GWh, whereas the future 2030 AMD system could train the same model using ~350 MWh, a 95% reduction. AMD then applied carbon intensities per kWh from the International Energy Agency World Energy Outlook 2024 [https://www.iea.org/reports/world-energy-outlook-2024]. IEA’s stated policy case gives carbon intensities for 2023 and 2030. We determined the average annual change in intensity from 2023 to 2030 and applied that to the 2023 intensity to get 2024 intensity (434 CO2 g/kWh) versus the 2030 intensity (312 CO2 g/kWh). Emissions for the 2024 baseline scenario of 7 GWh x 434 CO2 g/kWh equates to approximately 3000 metric tC02, versus the future 2030 scenario of 350 MWh x 312 CO2 g/kWh equates to around100 metric tCO2.
Footnotes
- Includes high-performance CPU and GPU accelerators used for AI training and High-Performance Computing in a 4-Accelerator, CPU hosted configuration. Goal calculations are based on performance scores as measured by standard performance metrics (HPC: Linpack DGEMM kernel FLOPS with 4k matrix size; AI training: lower precision training-focused floating-point math GEMM kernels operating on 4k matrices) divided by the rated power consumption of a representative accelerated compute node including the CPU host + memory, and 4 GPU accelerators.
- Based on 2015-2020 industry trends in energy efficiency gains and data center energy consumption in 2025.
- EPYC-030B: AMD takes compute node performance/watt measurements for AMD high performance CPU and GPU accelerators used for AI training and High-Performance Computing in a 4-Accelerator, CPU-hosted configuration.
- Performance for HPC workloads is based on Linpack DGEMM kernel FLOPS with 4k matrix size. Performance for AI training is based on lower precision training-focused floating-point math GEMM kernels operating on 4k matrices.
- Watts are based on the TDP of a representative accelerated compute node including the CPU host + memory, and 4 GPU accelerators
To make the goal particularly relevant to worldwide energy use, AMD worked with Koomey Analytics to assess available research and data that includes segment-specific datacenter power utilization effectiveness (PUE), including GPU HPC and machine learning (ML) installations. The AMD CPU socket and GPU node power consumptions incorporate segment-specific utilization (active vs. idle) percentages and are multiplied by PUE to determine actual total energy use for calculation of the performance per watt.
The energy consumption baseline uses the same industry energy per operation improvement rates as were observed from 2015-2020, with this rate of change extrapolated to 2025. The AMD goal trend line (Table 1) shows the exponential improvements needed to hit the goal of 30-fold efficiency improvements by 2025. The actual AMD products released (Table 2) are the source of the efficiency improvements shown for AMD goal status in Table 1.
The measure of energy per operation improvement in each segment from 2020-2025 is weighted by the projected worldwide volumes (as per IDC - Q1 2021 Tracker Hyperion - Q4 2020 Tracker, Hyperion HPC Market Analysis, April ’21). Translating these volumes to the ML training and HPC markets results in node volumes as per Table 3 below. These volumes are then multiplied by the Typical Energy Consumption (TEC) of the respective computing segment in 2025 (Table 4) to arrive at a meaningful aggregate metric of actual energy usage improvement worldwide.
Table 1: Summary efficiency data projected to 2025
2020
2021
2022
2023
2024
2025
Goal Trend Line
1.00
1.97
3.98
7.70
15.20
30.00
AMD Goal Status (energy-weighted performance / watt)
1.00
3.90
6.79
13.49
28.29
37.85
Table 2: AMD Products
2020
2021
2022
2023
2024
2025
EPYC Gen 1 CPU + M50 GPU
EPYC Gen 2 CPU + MI100 GPU
EPYC Gen 3 CPU + MI250 GPU
MI300A APU (4th Gen AMD EPYC™ CPU with AMD CDNA™ 3 Compute Units)
EPYC Gen 5 CPU + MI300X GPU
EPYC Gen 5 CPU + MI355X GPU
*AMD Products are supported by the latest software, including AMD ROCm.
Table 3: Volume Projections (millions/yr)
2020
2021
2022
2023
2024
2025
HPC GPU nodes sold
0.05
0.06
0.07
0.09
0.10
0.12
ML GPU nodes sold
0.09
0.10
0.12
0.14
0.17
0.20
Table 4: Base case 2025 electricity consumption of products sold in that year, for weighting efficiency indices (TWh/year)
2025
Base HPC
4.49
Base ML
29.79
Total Base
34.28
* Estimated values for 2025 on worldwide energy use are updated annually as HPC and ML compute node capabilities evolve from our original outlook, including with the growth of AI increasing the weighting for ML performance.
- AMD based advanced racks for AI training/inference in each year (2024 to 2030) based on AMD roadmaps, also examining historical trends to inform rack design choices and technology improvements to align projected goals and historical trends. The 2024 rack is based on the MI300X node, which is comparable to the Nvidia H100 and reflects current common practice in AI deployments in 2024/2025 timeframe. The 2030 rack is based on an AMD system and silicon design expectations for that time frame. In each case, AMD specified components like GPUs, CPUs, DRAM, storage, cooling, and communications, tracking component and defined rack characteristics for power and performance. Calculations do not include power used for cooling air or water supply outside the racks but do include power for fans and pumps internal to the racks.
FLOPS HBM BW Scale-up BW Training 70.0% 10.0% 20.0% Inference 45.0% 32.5% 22.5%
- Regression analysis of achieved accuracy/parameter across a selection of model benchmarks, such as MMLU, HellaSwag, and ARC Challenge, show that improving efficiency of ML model architectures through novel algorithmic techniques, such as Mixture of Experts and State Space Models for example, can improve their efficiency by roughly 5x during the goal period. Similar numbers are quoted in Patterson, D., J. Gonzalez, U. Hölzle, Q. Le, C. Liang, L. M. Munguia, D. Rothchild, D. R. So, M. Texier, and J. Dean. 2022. "The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink." Computer. vol. 55, no. 7. pp. 18-28.” Therefore, assuming innovation continues at the current pace, a 20x hardware and system design goal amplified by a 5x software and algorithm advancements can lead to a 100x total gain by 2030.
- AMD estimated the number of racks to train a typical notable AI model based on EPOCH AI data (https://epoch.ai). For this calculation we assume, based on these data, that a typical model takes 1025 floating point operations to train (based on the median of 2025 data), and that this training takes place over 1 month. FLOPs needed = 10^25 FLOPs/(seconds/month)/Model FLOPs utilization (MFU) = 10^25/(2.6298*10^6)/0.6. Racks = FLOPs needed/(FLOPS/rack in 2024 and 2030). The compute performance estimates from the AMD roadmap suggests that approximately 276 racks would be needed in 2025 to train a typical model over one month using the MI300X product (assuming 22.656 PFLOPS/rack with 60% MFU) and <1 fully utilized rack would be needed to train the same model in 2030 using a rack configuration based on an AMD roadmap projection. These calculations imply a >276-fold reduction in the number of racks to train the same model over this six-year period. Electricity use for a MI300X system to completely train a defined 2025 AI model using a 2024 rack is calculated at ~7GWh, whereas the future 2030 AMD system could train the same model using ~350 MWh, a 95% reduction. AMD then applied carbon intensities per kWh from the International Energy Agency World Energy Outlook 2024 [https://www.iea.org/reports/world-energy-outlook-2024]. IEA’s stated policy case gives carbon intensities for 2023 and 2030. We determined the average annual change in intensity from 2023 to 2030 and applied that to the 2023 intensity to get 2024 intensity (434 CO2 g/kWh) versus the 2030 intensity (312 CO2 g/kWh). Emissions for the 2024 baseline scenario of 7 GWh x 434 CO2 g/kWh equates to approximately 3000 metric tC02, versus the future 2030 scenario of 350 MWh x 312 CO2 g/kWh equates to around100 metric tCO2.