The American Journal of Applied Sciences
111
https://www.theamericanjournals.com/index.php/tajas
TYPE
Original Research
PAGE NO.
111-127
10.37547/tajas/Volume07Issue07-12
OPEN ACCESS
SUBMITED
11 June 2025
ACCEPTED
28 June 2025
PUBLISHED
26 July 2025
VOLUME
Vol.07 Issue 07 2025
CITATION
Lulla, K. L., Chandra, R. C., & Sirigiri, K. S. (2025). Proxy-Based Thermal
and Acoustic Evaluation of Cloud GPUs for AI Training Workloads. The
American Journal of Applied Sciences, 7(07), 111
–
127.
https://doi.org/10.37547/tajas/Volume07Issue07-12
COPYRIGHT
© 2025 Original content from this work may be used under the terms
of the creative commons attributes 4.0 License.
Proxy-Based Thermal and
Acoustic Evaluation of
Cloud GPUs for AI Training
Workloads
Senior Board Test Engineer, NVIDIA, CA, USA.
Tools and Automation Engineer, Amazon, CA, USA
.
Software Developer, Redmane Technology, IL, USA
Abstract:
The use of cloud-based Graphics Processing
Units (GPUs) to train and deploy Deep Learning models
has grown rapidly in importance, with the demand to
learn more about their thermal and acoustic behavior
under real-world workloads. A normal cloud cannot
make direct telemetry like temperature, fan speed, or
acoustic emissions. To overcome such shortcomings, this
study quantifies GPU workloads' thermal and acoustic
output with a proxy-based model derived from available
metrics such as GPU utilization, memory provisioning,
power consumption, and empirical Thermal Design
Power (TDP) values. They compare the two typical AI
tasks, BERT on natural language processing and YOLOv5
on real-time object detection, on Colab-based NVIDIA
GPUs (T4, V100, P100). The nvidia-smi was used to gather
runtime logs, and the specifications of the GPUs have
been obtained in the form of public Kaggle datasets.
Proxy statistics, including TDP-per-MHz and thermal
load (Power * Duration), were calculated to model heat
loss due to workload. To measure the degree of acoustic
impact, a threshold of TDP was applied to approximate
the level of fan-driven acoustics. The visual analytics,
such as boxplot, scatterplot, and bubble plot,
demonstrated certain considerable distinctions in the
stress patterns of GPUs: the BERT jobs demanded
extremely high cumulative thermal load and medium
acoustic effect, whereas the YOLOv5 demonstrated
bursty power footprint and substantial acoustic imprint
on high-TDP GPUs. The findings reveal that proxy
The American Journal of Applied Sciences
112
https://www.theamericanjournals.com/index.php/tajas
estimation is reproducible, interpretable, and a
lightweight substitute for determining the GPU thermal
and acoustic behavior of a machine used in the cloud
setting. Such a solution facilitates making thermal-aware
schedules, optimizing the infrastructure, and deploying
AI models with reduced energy consumption in multi-
tenant GPU environments.
Keywords:
Cloud GPUs; Thermal Load Estimation;
Acoustic Classification; Proxy Metrics; AI Workloads;
Energy-Aware Computing.
1.
INTRODUCTION
Artificial intelligence (AI) is evolving as quickly as it has;
therefore, the mounting computational pressure,
particularly demanding the training and deployment of
deep learning- based artificial intelligence models, is
exponentially increasing the demand [1]. High-
performance computing is fundamental to high-
performance applications, such as natural language
processing (NLP) and computer vision [2], where
Graphics Processing Units (GPUs) have become the
standard computing resource of choice due to their low
cost and scalable model training. High-end GPUs, such
as the NVIDIA T4, P100, and V100, are available through
services like Google Colab, Amazon Web Services (AWS),
and Google Cloud Platform (GCP), and this is making AI
workloads available to more researchers and developers
around the world [3].
However, the thermoacoustic engineering issues are
raised by the rising density of such workloads on shared
GPU servers [4]. Heat generation and heat dissipation in
data centers may adversely affect hardware life cycle
and power efficiency, and there is a great potential to
amplify energy consumption [5]. Analogously, a high
acoustic output, mainly caused by breakneck fan speeds
provoked when the GPU is running intensive tasks, can
result in unwanted noise pollution within data centers
and institutions in laboratories [6]. Although GPU
supercomputers with enterprise-scale GPU memory use
may include active thermal management solutions and
have rack-level noise suppression abilities, a low-level
thermal performance and audio response, when applied
to individual users in public clouds, may not be visible
and manageable. The mismatch between the required
work and the system-level thermal awareness results in
a crucial gap in the long-term and ethically responsible
functioning of AI systems in the cloud arena [5].
In applied thermal engineering terms, thermal profiling
is needed in predictive maintenance, effective design of
the cooling system, and to make intelligent work
schedules in thermally limited facilities [7]. However,
empirical studies on quantifying the impact of various AI
workloads upon GPU-related heat dissipation and
acoustics, especially on platforms where telemetry data
[8], temperature sensors, or the work of fans are not
exposed to the end-user, are hard to find. This limitation
should be addressed, mainly due to the increasing
popularity of cloud providers running multi-tenant
systems in which multiple jobs running in parallel
reinforce total thermal stress [9]. Although the cost of
high-performance AI training to the environment and
operation chains has been recognized more frequently,
cloud-based systems like Google Colab fail to deliver
customers with sensitive telemetry in terms of heat or
acoustics [10]. Namely, heating of a GPU, fan speed, and
power consumption in real-time during a workload
execution are not directly visible [11]. Such a lack of
sensor-level visibility hinders the creation of thermally
sensitive AI programs. It restricts the capacity of users to
maximize model settings to enable the sustainable
consumption of resources.
This means that the researchers are ascertaining the
thermoacoustic footprint of workloads through proxy
metrics, like the number of active GPUs, the extent of the
memory consumption, and published Thermal Design
Power (TDP) as labeling. Yet, no standardized
procedures or repeatable experiments have been
devised to exploit the available indirect indicators to
measure model-specific thermal and acoustic behaviors.
This paper will fill that gap by suggesting a proxy-based
estimation system incorporating information on publicly
known GPU specifications and logging runtime behavior
on actual AI workloads.
The current research aims to develop a lightweight,
reproducible approach to assessing the thermal and
acoustic behavior of AI model training workloads
deployed on cloud networks with GPUs based
exclusively on indirect, accessible feedback indicators.
Specifically, it is concentrating on two deep learning
models which are popular: BERT (Bidirectional Encoder
Representations from Transformers), an example of the
training of large-scale NLP with persistent GPU
utilization and a long period of execution, and YOLOv5
(You Only Look Once, version 5), which is a symbol of a
The American Journal of Applied Sciences
113
https://www.theamericanjournals.com/index.php/tajas
real-time object detection task that causes short but
powerful GPU utilization limitations.
To estimate thermal behavior, the framework uses the
product of GPU utilization, the approximate power draw
of GPUs according to TDP, and model training duration
as a proxy of cumulative thermal load. The estimate of
acoustic impact is based on the published GPU noise
benchmarks and classification, with higher TDP numbers
having higher de facto fan noise (in decibels A-weighted,
dBA). Each simulation is conducted on Google Colab Pro
instances and compared to publicly available datasets
on GPU hardware specifications on Kaggle, containing
their properties, like memory type, clock speeds, and
bus interfaces.
This paper has the following four contributions: (1) the
establishment of a proxy-based model to estimate
thermal and acoustic behavior of cloud-based AI
workload, (2) comparison of thermal profiles and noise
classification of BERT and YOLOv5 on a variety of GPUs
(T4, V100, P100), (3) unification of GPU utilization
request logs with public GPU specifications datasets to
facilitate transparency and reproducibility, and (4)
recommendations on shaping effective thermal-aware
scheduling and acoustic profiling of remote multi-tenant
GPU workloads.
The subsequent part of the paper is constructed in the
following way. In section 2, the terminologies that are
applied throughout the paper are presented. The third
section surveys the div of literature regarding both the
thermal behavior of GPUs and the acoustic
representation of GPUs and the nature of gaps in
benchmarking standards about modeling AI work.
Section 4 explains the experimental process, such as
selection of the workloads, the datasets used to specify
GPUs, the proxy computation reasoning, and the
visualisation process. Section 5 shows the outcome of our
simulations, including charts displaying GPU utilization,
GPU thermal load, and GPU acoustic classification by
specific workloads and GPU variants. Section 6
addresses implications of these findings as far as
sustainability and system design are concerned. Section
7 summarizes the acquired knowledge, and Section 8
plans future research, including integration with real-
time telemetry and physical sensor-based acoustic
validation.
2.
Nomenclature
Table 1. Nomenclature
ABBREVIATION
DESCRIPTION
TDP
Thermal Design Power - the identification of the highest quantity of heat that a GPU
should be able to dissipate with maximum possible working loads
dBA
A-weighted decibel - the unit of measurement of sound magnitude of intensity that is
weighted to coincide with the perception of a human being
GPU_util
GPU use - proportion of time the GPU is busy carrying out work
BERT
Bidirectional Encoder Representations from Transformers, or the LARGE NLP model
YOLO
You Only Look Once - a family of object detection models in real-time
COCO128
Subset of 128 images of the MS COCO dataset used by rapid-ini
SQuAD
Stanford Question Answering Dataset, an evaluation dataset of NLP models
3. LITERATURE REVIEW
3.1 Thermal Behaviour in GPUs
The way the thermals in GPUs respond to architecture,
power consumed, memory bandwidth, and workload
profile affects thermal behavior. Under manufacturer
specifications, Thermal Design Power (TDP) is
considered a conservative upper mark regarding the
level of heat that a GPU will produce when operated
The American Journal of Applied Sciences
114
https://www.theamericanjournals.com/index.php/tajas
under ideal circumstances [12]. For example, NVIDIA T4
uses a TDP of roughly 70 W, whereas V100 and P100
coordinate GPUs have much bigger TDP rates of 250 W
and 300 W, respectively. Unlike other cooling
specifications, these are vital to designing cooling
systems and form a helpful proxy where direct
temperature telemetry is unavailable [13].
Contemporary GPUs feature dynamic power and
thermals (Dynamic Power and Thermals include several
time-varying mechanisms that dynamically control the
clock rate and the fan speed, e.g., adaptive clock
throttling) [14]. The temperature in the GPU increases as
the number of usages grows and sparks a rise in the
number of rotations per minute (RPM). These fan speed
curves are generally non-linear and dependent on the
manufacturer; controlled by internal firmware or system
BIOS, and may be unavailable in virtualized or cloud-
based (Google Colab, AWS SageMaker) environments
[15]. This failure to read these real-time parameters
constrains the end-user control and observability of
thermally significant behaviors when training an AI or
using inference workloads [16].
Also, GPU power consumption is directly connected to
the workload. Transformer models such as BERT have a
high memory occupancy and a consistent use of
compute, making them sustain moderate levels of heat
production [17]. Conversely, vision-related models (e.g.,
YOLOv5) can be quite bursty in usage, causing
temporary thermal spikes. Such changes may be more
challenging to control thermally, perhaps in a data
center application where thermal inertia complicates
the overall scale of cooling response [18].
Real-life GPU thermal tests have not been easy to
perform, due to a lack of temperature sensors or access
to temperature sensors within the hardware
benchmarking area or inside a probed gas laboratory.
Nonetheless, there has been limited peer-reviewed
research on the thermal behavior of restricted-access
cloud environments under the real-life AI workloads
[19]. This leaves a methodological vacuum that can be
filled by applied thermal engineers interested in
designing
or
optimizing
energy-efficient
AI
infrastructure at scale.
3.2
Acoustic Analysis in Cloud Data Centers
Although directly linked to heat output, acoustic
emissions are a significant secondary aspect that should
be considered in the thermal management of high-
performance computing facilities [20]. Acoustic noise,
usually quoted in dBA, is caused mainly by cooling tools,
e.g., liquid cooler pumps or high-RPM fans [21]. In
addition to being a comfort and safety concern to
human operators, the acoustic footprint of a GPU-
intensive system provides a proxy measure of system
stress and thermal load.
Rack-based thermal management solutions such as
redundant fans, cold aisle containment, and adaptive
airflow control are essential in cloud data centers [22].
GPU work offers higher thermal dynamics, which causes
system firmware to push up fan rpm to ensure safe
operating temperatures are met [23]. This, in fact,
results in increased acoustic output, usually beyond 45-
50 dBA in racks under full load [24]. Where the
hyperscale facilities are concerned, the metrics of the
acoustics can be part of the overall energy management
approach. However, these metrics will usually not be
revealed at the user level.
Despite its applicability, little work has been done in
integrating acoustic analysis into AI workload
benchmarking. Most published benchmarks (e.g.,
MLPerf or TensorFlow Model Garden) only consider
latency, throughput, and energy efficiency and leave the
noise level aside. However, noise may play a serious role
in hybrid edge-cloud scenarios or the academic lab
environment with local GPU clusters in general offices.
This under-researched aspect bears relevance to
sustainable computing, especially where the design has
to reduce not only thermal but also acoustic emissions.
3.3
Benchmark Gaps
Existing dynamic markets in benchmarking, e.g., MLPerf
(TensorFlow/Pytorch), HuggingFace Transformers, and
ONNX Model Zoo, focus on model accuracy, throughput,
and computation latency. Although these are essential
to performance assessment, they fail to acknowledge
the thermodynamic or acoustic consequences of the
runtime of AI models. Consequently, the issues in terms
of infrastructure level, like cooling system stress, fan
power consumption, and acoustic pollution, stay
beyond the boundaries of traditional benchmarking
methods.
Additionally, in public cloud, most users do not get access
to low-level telemetry like real-time GPU temperature,
voltage, or rpm of fans [25]. This restriction prohibits
The American Journal of Applied Sciences
115
https://www.theamericanjournals.com/index.php/tajas
granular thermal profiling and hard-to-enforce
workload-aware scheduling policies [26]. Not all
academic research to simulate thermal behavior has
used synthetic workloads, but they do not necessarily
reflect the time-dependence of the accurately detailed
deep-learning models' state-of-the-art.
The literature also does not provide a common approach
to estimate the acoustic impact based on available
routine
metrics.
GPU
reviews
include
dBA
measurements under a stress load; however, these are
not normalized between models or loads. Without a
proxy-based estimation framework, users cannot
predict an AI model's thermal or acoustic energy
expenditure [27], especially when working in shared or
energy-restricted settings.
The proposed research will fill these gaps by proposing
and proving an AI proxy-based, lightweight, and
reproducible profiling of thermal and acoustic
performance of AI workloads in cloud GPUs.
4.
METHODOLOGY
4.1.
Environment and Workloads
The paper was run on the Google Colab Pro+ GPU
machine learning cloud environment, where you can
gain temporary access to powerful GPUs, including the
NVIDIA Tesla T4, P100, and V100. The GPUs are popular
AI compute used in academic and business workloads,
representing realistic thermal conditions in a cloud
computing data center environment. Google Colab was
chosen because of its availability, consistent time limit,
and serial distribution of workload by repeat and scale
deployments without dedicated hardware.
The evaluation of thermal and acoustic characteristics
was served by two benchmark loads, including BERT and
YOLOv5.
The
model
(Bidirectional
Encoder
Representations from Transformers), BERT, was
improved on the SQuAD v2 benchmark, a commonly
used NLP benchmark covering more than 150,000 pairs
of questions and answers. BERT workloads were selected
based on long-lasting GPU usage and memory usage
without taking over the space of the machine, e.g.,
because of a long training process with not-so-dynamic
hardware load distribution. By contrast, YOLOv5 (You
Only Look Once, version 5) was unleashed to detect
objects in real-time through the 128-image subset
(COCO128) of the COCO database, which was very small.
YOLOv5 has a bursty computation pattern, where the
processing time within a short time is high, and GPU
utilization is random. These two types of workloads will
be used to compare computational and inference-
intensive model behavior in similar run-time situations.
4.2
Logging and Data Collection
Google Colab lacks direct telemetry of GPU temperature
and GPU fan speed, so, using such programmable
attributes, the study had to base its ideas on too indirect
measures according to the results gathered with the
help of a command-line tool that interacts with the
NVIDIA Management Library (NVML) called nvidia-smi.
The logging process was facilitated to take
measurements of 10 seconds when the model was
running. In particular, the script captured the percentage
of GPUs utilized (gpu_util), the amount of memory in
MB, and the immediate power consumption (in watts)
used as the starting point of proxy-based thermal
analysis. These logs were saved and time-stamped so
that they could be synchronized with training/inference
steps.
In addition to GPU metrics, logs were gathered about the
system-level use of the CPU and the RAM. They do not
cause any changes to GPU thermal profiles, but can offer
some background context to resource consumption, and
may affect model scheduling or performance variation.
The plot material used in implementation addresses the
usage of GPU variably, over time, on both BERT and
YOLOv5 workloads. Such plots of time series indicated
that during the fine- tuning process, BERT had a constant
load on the GPU between 70-85% the entire time,
whereas YOLOv5 had highs above 90% and a subsequent
dramatic low, as can be expected of an inference-
oriented program.
4.3
Dataset Integration
To strengthen the strength of proxy estimations,
publicly accessible Kaggle datasets were included in the
analysis. The primary dataset was tpu_gpus.csv; it had
the specifics on more than 150 GPU models. Main
specifications were TDP values, GPU and memory clock
frequencies, memory type (GDDR, HBM, etc), and bus
interface (e.g., PCIe 3.0, 4.0). This dataset has enabled
the advent of this study to align a GPU used in every
Colab session with identified hardware characteristics to
derive proxy thermal and acoustic scores.
The American Journal of Applied Sciences
116
https://www.theamericanjournals.com/index.php/tajas
Simultaneously, we compared some data related to
thermal trends in CPUs based on the tpu_cpus.csv file,
and cross-validated it. Even though this study is GPU-
based, CPU thermal characteristics provided a basis for
building clock-speed-to-TDP relationships to formulate
derived values such as TDP-per-MHz. In addition, CPU
datasets described the overall historical trend in
processor design and heat production, and all these were
presented in relative plots. These data sets enabled a
predictable and augmenting framework that did not
need genuine sensor statistics.
4.4
Proxy Formulas
The necessary direct thermal and acoustic telemetry
were not there, and several proxy formulas have been
established to determine the respective performance
parameters. The TDP-per-MHz calculation revealed that
the GPU, which was equipped with a Thermal Design
Power, was divided by its minimum clock speed in MHz.
This is a normalized measure of thermal efficiency in that
higher numbers reflect thermally inefficient hardware.
An example would be a GPU with a TDP of 250 W and
with a clock frequency of 1250 MHz; the TDP-per-MHz
would be 0.2 W/MHz.
The measure of acoustic impact was approximated to a
binary proxy classification. Limits of TDP > 150 W were
based on industry data and manufacturer reports used
to determine that such conditions correspond to the
notion of high acoustic load when the sound pressure of
fans became higher than 45 dB A. Any GPU with TDP <
or = 150 W was considered to be in a moderate
acoustical range. This is a very simple way, but it matches
published acoustic profiles on GPUs at the server-class
under load and serves as a fairly reasonable proxy of the
system- level fan response.
Finally, total Thermal Load was set as a product of power
draw (in watts) and execution time (in minutes). This
estimation is the accumulated energy consumed in heat
to execute the model. For example, a YOLOv5 model
powered by 150 W for 20 min will have a total thermal
load of 3000 Wmin. This measure was used in both
workloads to compare the thermal footprints with the
varied run-time attributes.
4.5
Data Preprocessing
The GPU and CPU datasets were preprocessed before
being analyzed to clean and transform essential
attributes. GPU_clock and Memory_clock were cleaned
of adjectival suffixes (MHz) and directed into numerical
conversion of the GPU data set. Unavailable values, such
as in- memory details, were given a NaN value and not
included in the ratio type of calculations. A regular
expression pattern was used to extract the type of GPU
memory used to sort the most common technology type
(e.g., GDDR5, HBM2) and to filter out all the possible
values so that the research could test the same
parameter and check whether the memory setup may
affect the acoustic or the thermal performance.
The TDP value was sifted out in the CPU dataset to
eliminate the incomplete or incorrect values. Other
models used ranges (e.g., 2.4-3.8 GHz), which were
content parsed to retrieve the minimum and maximum
values to be used in normalization. This allowed the
derivation of TDP-per-MHz of CPUs as a benchmark
reference point compared to GPUs. Some plot
references produced in the preprocessing stage were a
histogram of CPU TDP distribution, a bell-shaped curve
with a peak around 80120 W, and a scatterplot of TDP
and clock speed combined by socket type. Such
visualizations confirmed the usefulness of normalized
thermal measures in per-processor architecture.
The distribution of the GPU types of memory was also
visualized on a countplot that demonstrated the number
of memory standards prevailing. Different versions of
GDDR prevailed in the dataset, and HBM and DDR
variants were presented in smaller proportions. Such
distribution played a vital role in externalizing thermal
variations, where various types of memories have
varying power and heat dissipation behavior, especially
in a high throughput mode, e.g. typical of BERT and
YOLOv5.
The American Journal of Applied Sciences
117
https://www.theamericanjournals.com/index.php/tajas
Figure 1.
System Architecture
Figure 1 represents a proxy-based system architecture
to test the GPU performance under thermal and
acoustic conditions in clouds. It starts with AI workload
execution (BERT, YOLOv5) on Google Colab that makes a
log of the run-time metrics through nvidia-smi. These
logs are combined with Kaggle GPU spec files (e.g., TDP,
clock speed, memory type). The self-predicting
infrastructure system formulates combined data that is
input into a preprocessing path where proxy measures,
TDP-per-MHz, Thermal Load, and Acoustic Level are
calculated. Visualization modules/Analytical models
then produce outputs such as the bar chart, bubble plot
which would lead to thermal classification, acoustic
estimation, and scheduling information of the energy-
efficient deployment of GPU workload.
5.
RESULTS
5.1. Model Behaviour on GPUs
The observed patterns of GPU use between BERT and
YOLOv5 workloads showed a core behavior difference in
ways compatible with the respective network
architectures. As graphically pointed out in Figure 2:
GPU Utilization Patterns for BERT and YOLOv5, the BERT
fine-tuning task of a 90-minute duration activity had
maintained a high average GPU usage (~85%) with a
small degree of variation. This constant g men
(continuously using transformer-based models that
necessitate constant matrix operations and attention-
weighting) is a pointer towards the constant thermal
output and memory consumption (~6000 MB).
The American Journal of Applied Sciences
118
https://www.theamericanjournals.com/index.php/tajas
Figure 2.
GPU Utilisation Patterns for BERT and YOLOv5
In contrast, the YOLOv5 workload, which was trained on
the COCO128 subset, displayed a bursty utilization
behavior. There was a high GPU usage that would often
reach above 95 percent during training epochs but would
stall drastically during intermediate evaluation, batch
loading, or checkpoints. The RTT amounted to ~20 min,
and the average memory load was~4000 MB. The steep
curves of the GPU load imply temporary high-power
consumption and fan turbo-ups, which result in local
thermal spikes without overall shortened span.
This contrast between the continuous workload
experienced within BERT and the burst- informed
inference manner of operation in YOLOv5 formed the
basis for interpreting downstream thermal and
acoustical attributes.
5.2
Thermal Proxy Comparisons
Each GPU model was computed with the normalized
value TDP-per-MHz to measure thermal efficiency using
available Kaggle specifications. This ratio with a base
GPU clock was observed by converting it graphically into
a boxplot, as shown in Figure 3: Normalized Thermal
Output (TDP/MHz). The distribution showed that most
modern GPUs have performance clustering below 0.05
W/MHz. However, some high-performing cards, such as
P100 and V100, had values beyond 0.25 W/MHz, which
indicates increased heat generation per frequency.
The American Journal of Applied Sciences
119
https://www.theamericanjournals.com/index.php/tajas
Figure 3.
Normalized Thermal Output (TDP/MHz)
Figure 4.
Thermal Load Estimate: BERT vs YOLOv5
Figure 4, BERT vs YOLOv5 directly compares the
cumulative thermal footprint (Power x Duration). The
total approximate load in terms of Watt-minute
generated by BERT was estimated to be 6300 W min (70
W vertical multiplied by 90 min), as compared to the
approximated ~3000 W·min generated by YOLOv5 with a
shorter but power-intensive session (150 W vertical
multiplied by 20 min). Although the peak draw of
YOLOv5 is greater, the cumulative heat dissipation was
more than twice as long due to the long time BERT takes.
Hence, BERT workloads impose moderate but consistent
thermal pressure that can be addressed with
temperature-predictable cooling methods, whereas
The American Journal of Applied Sciences
120
https://www.theamericanjournals.com/index.php/tajas
YOLOv5 creates sudden and brief spikes in thermal loads
that may push data centers to the limit and go back in
quick succession.
5.3
Acoustic Classification
An acoustic prediction procedure was carried out based
on the two-category characterization: GPUs with TDP
greater than 150 W received the label HN (>45 dBA), and
those that were less than or equal to 150 W became M
(<45 dBA). Figure 5: Estimated Acoustic Level Based on
TDP reveals that most of the GPUs landed in the
moderate group, but above that, a significant number,
roughly equal to V100, P100, etc., registered more than
a high noise level.
Figure 5.
Estimated Acoustic Levels Based on TDP
This is vital in deploying workloads. The fact is that BERT
was largely implemented on the NVIDIA T4 (TDP 70 W),
so it was always linked to a minimal acoustic footprint.
YOLOv5, however, using large-TDP GPUs, measured in
the high-noise category, demonstrated that inference
runs may cause rather unreasonably high acoustic stress
with the ratio. Such results indicate that high-draw and
bursty workloads such as YOLOv5 must be directed to
racks with better acoustic isolation or redundancy
cooling.
5.4
Cross-Hardware Performance Patterns
The various scatter and box plots investigated the overall
architectural characteristics of GPUs. Figure 6: GPU Clock
Speed vs Memory Size uses the graph to illustrate the
poor relationship between the two variables, GPU
frequency and onboard memory, using raw data from
Kaggle.
Outliers,
long-clock
speed-low-memory-
footprint GPUs were also present, which signaled
specialised or outdated designs. Most new AI GPUs
operated at a range of 1200 1800 MHz with memory
capacity varying between 16 and 32 GB.
The American Journal of Applied Sciences
121
https://www.theamericanjournals.com/index.php/tajas
Figure 6.
GPU Clock Speed vs Memory Size
Figure 7.
GPU Clock vs Memory Clock by Bus Type
Figure 7 indicates the bus interface's significance in
memory performance. PCIe 4.0 and NVLink-based cards
went further with memory clocks centered even higher
than 1500 MHz, while older buses (AGP, PCI) plateaued
at much lower frequencies. As the memory bandwidth
not only influences the throughput but also results in
thermal accumulation, this further supports the idea
that the current generation of GPUs is simply better
prepared to handle the thermal spikes of huge AI
models.
The American Journal of Applied Sciences
122
https://www.theamericanjournals.com/index.php/tajas
Figure 8.
CPU TDP Distribution by Release Year
There has been a generational change to thermal design
philosophy, as illustrated in Figure 8. The TDP figures
scarcely ranged above 100 W between 1998 and 2010.
Since 2015, there has been a sharp rise in the median
TDP, with several 2020-2023 GPUs going above 250 W.
This indicates the development of architecture and the
increased need for AI-optimized silicon.
6.
DISCUSSION
6.1 Workload-Specific Thermal Dynamics
Thermal patterns in BERT and YOLOv5 workloads exhibit
opposite trends, which suggests significant details of
workload-specific stresses on GPUs. BERT is a
transformer-based NLP model with a long, stable
pattern of use with large memory occupation and a
steady pattern of GPU usage [28]. Previous studies
showed that transformer-based NLP models have
enduring energy demands during fine-tuning operations,
particularly on big datasets such as the SQuAD v2 [17]. In
our experiments, this resulted in a very high cumulative
thermal load given a relatively moderate power draw,
translating into the bar plot array comparing Thermal
Load.
Conversely, a convolutional object detection model
(YOLOv5) tuned to real-time usage had a hostile usage
pattern with large power surges and occasional idle
periods on the GPU [29]. These spikes in the behavior of
our GPU utilization logs confirm previous findings that
computer vision models are characterized by short
electro-thermal spikes of immense demand, which
overtax their steady-state cooling capabilities unless
handled appropriately.
These behavioral differences are essential to
infrastructure management regarding their thermal
implications. BERT's thermal footprint is predictable,
encouraging it to fit into a consistent cooling range,
whereas a burst pattern common to YOLOv5 could cause
spikes of overheating or fan speed if the burst has not
been scheduled with such thermal padding. Such
dynamics call for mitigating the significance of matching
workload types with suitable equipment and cooling
approaches, especially in mixed-use settings.
6.2
Acoustic Engineering Insights
The acoustic classification through proxy performed in
this study gives the initial information about the trends
of GPU-dependent noise in cloud and institutional data
centers. We could classify workload based on indirect
power-based heuristics by defining a conservative value
of TDP of 150 W as a boundary between moderate and
high levels of acoustics. The findings indicated that the
BERT, usually performed on T4 GPUs, with
The American Journal of Applied Sciences
123
https://www.theamericanjournals.com/index.php/tajas
comparatively low TDP (70 W), was clearly inside the
zone labeled Moderate Acoustic Load. YOLOv5 sessions,
in turn, often use V100 or P100 GPUs with a TDP of
250~300 W, which fall in the “High Noise”
category.
These measurements are confirmed by system-level
measurements in which the authors point to high-
performance workloads on GPUs greater than 200 W
TDP persistently causing fan speeds in excess of 5000
RPM and resulting in acoustic emissions reaching more
than 45 dBA [29]. Although it was impossible to measure
absolute values of dBA because of the Colab limitation
in this study, our classification gives a convenient rough
estimate of the acoustic disturbances caused by fans.
As a practical implication, YOLOv5 and all other short-
duration but high-performance scalers are best suited to
be slotted on thermally decoupled GPU servers or racks
with better noise suppression. In some university labs
and cloud-native server clusters, inference-heavy GPU
workloads are already co-located in acoustically shielded
areas. Our proxy analysis affirms this design philosophy,
which means that the inherent need to map the
behavior of AI models to physical infrastructure
attributes should be reinforced.
6.3
Sustainability in Shared Environments
Sustainability-wise, such trade-offs open prospects for
deploying long-duration batch jobs (e.g., BERT) and
bursty inference jobs (e.g., YOLOv5) in multi-tenant
cloud GPU infrastructures. Long-term jobs can be
thermally friendly because their maximum load is small.
Still, their total energy demand is significant, well
beyond 6000 Wmin in our experiments, which casts
doubt on cooling expenses and total power demand.
In contrast, bursty jobs might only take a short time to
complete but can cause rapid swings in system
temperature, promoting more intense fan cycling and
immediate energy spikes [30]. This corresponds to
previous findings, which proved that uneven workloads
within heterogeneous systems contribute to oscillating
power consumption and disrupt dynamic cooling control
[18].
According to our study, schedulers should concentrate
on sustainability-related constraints to account for the
peak and total thermal loads when scheduling
workloads to common GPUs. For example, high-burst
jobs can be coupled to thermal buffer periods, and long
jobs can be alternated with low-load background tasks to
more equally share the thermal stress. In addition, the
metrics that we use to achieve our method are
interpretable and could be implemented into cloud
orchestration systems to embrace green AI frameworks.
6.4
Accuracy and Limitations of Proxy Approach
Interpretability is one of the suggested framework's
strengths. Compared to more complicated simulations
or closed-loop monitoring systems, our proxy-based
approach only uses moderately available runtime logs
and already published GPU specifications. This makes it
compatible with reproducibility between platforms and
scalable where telemetry APIs are absent or constrained
[31]. The method helps monitor the lightweight
infrastructure with minimum required equipment by
translating
the
observed
utilization,
power
consumption, and TDP values into normalized thermal
and acoustic metrics.
However, it has some significant limitations. No direct
measure of temperature or fan RPM values available,
which restricts checks of proxy estimates against actual
ground truth measurements. Although our thresholds
and formulas were oriented on vendor documentation
and available benchmarks, we are generalizing these
values to all the conditions in the data center, and it is
an approximate process. For example, selecting the rack
airflow design, ambient room temperature, or type of
material used in the heat sink can influence the thermal
behavior independent of any aspect related to the
workload.
Moreover, the binary definition of acoustic impact (only
TDP thresholds are considered) can be too simplistic in
practice. The fan speed curves are not linear and are
usually controlled by vendor-specific firmware
algorithms, which could differ between vendors and
across different releases of the BIOS. Therefore, future
work can be expected to consider integrating controlled
laboratory measurements of GPU acoustics under
benchmark workloads to refine proxy calibration.
6.5
Comparison with Known GPU Specs
The final phase of our work was dedicated to analyzing
the correspondence between noticed model behavior
and official GPU specifications. The bar plot shows that
BERT workloads consumed more total thermal loads
overall despite using lower-TDP hardware, owing to the
The American Journal of Applied Sciences
124
https://www.theamericanjournals.com/index.php/tajas
longer runtime. Conversely, YOLOv5 also produced a
greater instantaneous load on greater- TDP GPUs, but
completed faster, leading to less dissipated energy. The
bubble plot also confirms this finding, as in the lower-
right quadrant (low power, high duration), we have
BERT, and in the upper-left quadrant (high power, low
duration), we have YOLOv5. Another factor that
supports the distinction in resource allocation patterns
between the two models is the size of the bubbles,
which would equate to the usage of GPU memory.
These correlations are consistent with observed GPU
performance tables published by NVIDIA and
performance evaluations of Anzt et al. (2021), indicating
that V100 and P100 GPUs are optimized to run a high-
throughput burst workload. In contrast, the T4s are
better suited to a sustained, latency-tolerant workload.
Such differences in design are empirically confirmed
with our results in workload testing in the real world of
Colab GPUs.
Therefore, this work has effectively shown that even
open-sourced GPU specifications can be used to forecast
the behavior of AI models via straightforward, albeit
effective, proxy techniques. The techniques help gain
knowledge on the suitability of the workloads, thermal
profiling, and acoustic prediction in restricted conditions
where access to hardware telemetry is not available
directly.
7.
CONCLUSION
This paper presented and confirmed a proxy model for
assessing GPU clouds' thermal and acoustic efficiency in
training AI tasks. Even in platforms such as Google Colab
Pro, where the direct telemetry of such variables as GPU
temperature or fan speed is difficult to acquire, the
framework proved capable of high-quality derivation
of interpretable and actionable insights based on GPU
usage, power consumption, memory use, and known
characteristics (such as TDP).
We identified the opposite tendencies of GPU behavior
based on similar benchmarking studies of two sample AI
workloads, BERT (NLP) and YOLOv5 (computer vision).
BERT showed an inferred longer duration, sustained
feature use with a high cumulative thermal loading but
low peak power, whereas YOLOv5 had blistering,
multiplicative features that used the GPU with a high-
power draw but lower cumulative thermal load. The
range of these distinctions was measured based on
thermal loading estimation and acoustic classification
with the help of TDP- based averages and standardized
comparisons such as TDP-per-MHz.
The results were backed up by visualizations using
bubble plots, boxplots, scatterplots, and bar charts,
providing evidence of workload-specific GPU strain and
efficiency profiles. The acoustic proxy showed that
workloads in YOLOv5 regularly caused GPUs to enter
high-noise states, but with BERT, execution on low-TDP
GPUs such as the T4 kept it in moderate sounds. Hence,
the results highlight that a lightweight, reproducible
thermal and acoustic evaluation in the limited cloud
areas is viable. This allows more efficient scheduling of
workload,
energy-
sensitive
computation,
and
infrastructure design without the necessity of invasive
sensors or access to proprietary telemetry. The fact that
the framework concurred with the publicly accessible
datasets guarantees its broad applicability in research
programs in universities and industries relevant to
sustainable implementations of AI systems.
8.
RECOMMENDATIONS AND FUTURE WORK
The findings of the current work present a few significant
suggestions to system designers, data scientists, and
infrastructure
engineers
involved
in
the
AI
implementation:
Thermal-Aware Scheduling
: Workloads must also be
scheduled based on the profile of thermal load rather
than the GPU availability. The steady heat-producing
BERT-like models are better applied in conditions of
constant cooling capacity. Burst compatible models,
such as YOLOv5 and others, can benefit thermally
insulated nodes by avoiding overheating or fan surge.
Acoustic Zoning in Data Centers
: Depending on the
correlation between TDP and noise levels identified,
high-TDP GPUs must be assigned to acoustically isolated
racks, particularly in academic labs or server rooms
located at points of consumption where background
noise is
essential. Stress benchmarking on acoustic output
should be considered in future AI-capable hardware.
Integrated Proxy Monitoring Tools
: The methodology
can be automated to be used in dashboarding schemes
on GPU usage, which calculate and present utilization
trends and approximate thermal/acoustic load in real
time. It allows active policies to be actively tuned to cool
The American Journal of Applied Sciences
125
https://www.theamericanjournals.com/index.php/tajas
and gives real-time alerting without any changes to the
hardware.
Reproducibility via Public Datasets
: This study's
reproducibility is proven by the fact that the Kaggle
GPU/CPU datasets and the runtime logs (acquired via
nvidia-smi) can be integrated into the qualitative
analysis. These proxies should be standardized in
developing AI benchmarking tools and academic
pipelines.
Future research will involve identifying the limitations of
the present-day approach. Firstly, value addition to the
telemetry, such as Google Colab Pro+ or AWS EC2
telemetry (in case API access is available), would allow
correlating proxy values with real temperature/fan
signals, resulting in higher calibration accuracy. Second,
using physical acoustic sensors in an experiment should
lead to a ground truth measure to optimize the dBA
classification. Third, the framework's application can be
generalized to multi-GPU workloads and hybrid CPU-GPU
systems (TPUs).
Finally, combining such thermal-acoustic profiling with a
model of energy costs and environmental quantities
(e.g., carbon intensity of power consumption) would aid
in green AI efforts. This would enable developers to
make intelligent choices regarding performance,
accuracy, and the sustainability of their AI compute
workflow.
A
CKNOWLEDGMENTS
The authors would like to acknowledge the help of
Google Colab Pro, which allowed us to run GPU-based
workloads in a real-world cloud environment with
constraints on most dimensions. And we are also
grateful to the Kaggle open hardware datasets
contributors who not only gathered and updated free,
publicly available, comprehensive GPU and CPU
specifications, but also because of which the
reproducibility of this work was made possible. Their
work made a large proxy-based thermal and acoustic
study possible without physical access to hardware.
Furthermore, we would like to embrace the general
open-source community (Matplotlib, Seaborn, and
pandas developers), whose visualization and analysis
tools played a significant role in understanding and
presenting the findings.
References
1.
Artificial Intelligence, Machine Learning, and Deep
Learning for Advanced Business Strategies: A Review
| Partners Universal International Innovation Journal
[Internet]. [cited 2025 Jun 28]. Available from:
https://puiij.com/index.php/research/article/view/
143
2.
High Performance Computing for Understanding
Natural Language: Computer Science & IT Book
Chapter | IGI Global Scientific Publishing [Internet].
[cited 2025 Jun 28]. Available from:
https://www.igi-
global.com/chapter/high-performance-computing-
for-understanding- natural-language/273400
3.
Mustafa F, Gilbert A. Scalable Data Architectures for
Generative AI: A Comparison of AWS and Google
Cloud Solutions [Internet]. Unpublished; 2024 [cited
2025
Jun
28].
Available
from:
https://rgdoi.net/10.13140/RG.2.2.26378.07364
4.
Rafsanjani H, Marwazi A, Sitompul D. Thermal
Management and Power Optimization in Modern
CPU and GPU Architectures.
5.
Katal A, Dahiya S, Choudhury T. Energy efficiency in
cloud computing data center: a survey on hardware
technologies.
Cluster
Comput.
2022
Feb
1;25(1):675
–
705.
6.
CPUs Versus GPUs | SpringerLink [Internet]. [cited
2025
Jun
28].
Available
from:
https://link.springer.com/chapter/10.1007/978-981-
97-9251-1_9
7.
Thermal intelligence: exploring AI’s role in optimizing
thermal systems
–
a review | Interactions [Internet].
[cited
2025
Jun
28].
Available
from:
https://link.springer.com/article/10.1007/s10751-
024-02122-6
8.
A Survey of Cloud-Based GPU Threats and Their
Impact on AI, HPC, and Cloud Computing.
9.
A systematic review of scheduling approaches on
multi-tenancy cloud platforms - ScienceDirect
[Internet]. [cited 2025 Jun 28]. Available from:
https://www.sciencedirect.com/science/article/abs/
pii/S0950584920302214
10.
Ajayi R. Integrating IoT and cloud computing for
continuous process optimization in real time
systems. Int J Res Publ Rev. 2025 Jan;6(1):2540
–
58
The American Journal of Applied Sciences
126
https://www.theamericanjournals.com/index.php/tajas
11.
Real-Time Thermal Map Characterization and
Analysis for Commercial GPUs with AI Workloads |
IEEE Conference Publication | IEEE Xplore [Internet].
[cited
2025
Jun
28].
Available
from:
https://ieeexplore.ieee.org/abstract/document/110
14443
12.
An Overview of Thermal and Mechanical Design,
Control, and Testing of the World’s
Most Powerful
and Fastest Supercomputer | J. Electron. Packag. |
ASME Digital Collection [Internet]. [cited
2025
Jun
28].
Available
from:
https://asmedigitalcollection.asme.org/electronicpac
kaging/article- abstract/143/1/011005/1082291/An-
Overview-of-Thermal-and-Mechanical-Design
13.
Design, Operation and Maintenance of Direct and
Indirect Evaporative Cooling Systems in Data Center
Thermal Management - ProQuest [Internet]. [cited
2025
Jun
28].
Available
from:
https://www.proquest.com/openview/eed632faea3
362b05f921b8213a5b9ab/1?pq-
origsite=gscholar&cbl=18750&diss=y
14.
Harnessing Machine Learning in Dynamic Thermal
Management in Embedded CPU-GPU Platforms |
ACM Transactions on Design Automation of
Electronic Systems [Internet]. [cited 2025 Jun 28].
Available
from:
https://dl.acm.org/doi/full/10.1145/3708890
15.
Bagai R. Comparative Analysis of AWS Model
Deployment Services. IJCTT. 2024 May 30;72(5):102
–
10.
16.
Zhang J, Zhang W, Xu J. Bandwidth-efficient multi-
task AI inference with dynamic task importance for
the Internet of Things in edge computing. Computer
Networks. 2022 Oct 24;216:109262.
17.
Ansar W, Goswami S, Chakrabarti A. A Survey on
Transformers in NLP with Focus on Efficiency
[Internet]. arXiv; 2024 [cited 2025 Jun 28]. Available
from:
http://arxiv.org/abs/2406.16893
18.
A survey on data center cooling systems: Technology,
power consumption modeling and control strategy
optimization - ScienceDirect [Internet]. [cited 2025
Jun
28].
Available
from:
https://www.sciencedirect.com/science/article/abs/
pii/S1383762121001739
19.
Knebel FP. Designing and implementing digital twins
with cloud and edge computing: challenges and
opportunities. Projetando e implementando Gêmeos
Digitais com computação em nuvem e de borda:
desafios e oportunidades [Internet]. 2024 [cited 2025
Jun
28];
Available
from:
https://lume.ufrgs.br/handle/10183/276593
20.
Advances in Numerical Modeling for Heat Transfer
and
Thermal
Management:
A
Review
of
Computational Approaches and Environmental
Impacts [Internet]. [cited 2025 Jun 28]. Available
from:
https://www.mdpi.com/1996-
21.
SOUND AND NOISE: MEASUREMENT AND DESIGN
GUIDANCE
–
HANDBOOK OF HUMAN FACTORS AND
ERGONOMICS - Wiley Online Library [Internet]. [cited
2025
Jun28].Availablefrom:
https://onlinelibrary.wiley.com/doi/abs/10.1002/978
1119636113.ch18
22.
GPU Devices for Safety-Critical Systems: A Survey |
ACM Computing Surveys [Internet]. [cited 2025 Jun
28].
Available
from:
https://dl.acm.org/doi/abs/10.1145/3549526
23.
Chowdhury U, Rodriguez J, Tradat M, Soud Q,
Wallace S, O”Brien D, et al. Acoustics Analysis of Air
and Hybrid Cooled Data Center. In: 2024 23rd IEEE
Intersociety
Conference
on
Thermal
and
Thermomechanical
Phenomena
in
Electronic
Systems (ITherm) [Internet]. 2024 [cited 2025 Jun
28].
p.
1
–
11.
Available
from:
https://ieeexplore.ieee.org/abstract/document/1070
9368
24.
Distributed
Cloud
Computing
Infrastructure
Management [Internet]. [cited 2025 Jun 28].
Available
from:
https://www.scirp.org/journal/paperinformation?p
aperid=143462
25.
TAPAS: Thermal- and Power-Aware Scheduling for
LLM Inference in Cloud Platforms | Proceedings of
the 30th ACM International Conference on
Architectural Support for Programming Languages
and Operating Systems, Volume 2 [Internet]. [cited
2025
Jun
28].
Available
from:
https://dl.acm.org/doi/abs/10.1145/3676641.37160
25
26.
Enhancing Reservoir Modeling and Simulation
Through Artificial Intelligence and
Machine
Learning: A Smart Proxy Modeling Approach -
ProQuest [Internet]. [cited 2025 Jun
28].
Availablefrom:
https://www.proquest.com/openview/f3886ad9ad2f
556a032981db194fcfa4/1?pq-
The American Journal of Applied Sciences
127
https://www.theamericanjournals.com/index.php/tajas
origsite=gscholar&cbl=18750&diss=y
27.
Ganesh P, Chen Y, Lou X, Khan MA, Yang Y, Sajjad H,
et al. Compressing Large-Scale Transformer-Based
Models: A Case Study on BERT. Transactions of the
Association for Computational Linguistics. 2021 Sep
21;9:1061
–
80
28.
Synchronizing Object Detection: Applications,
Advancements and Existing Challenges | IEEE
Journals & Magazine | IEEE Xplore [Internet]. [cited
2025
Jun
28].
Available
from:
https://ieeexplore.ieee.org/abstract/document/1049
9817
29.
Aalborg University, Liu J. Automatic Analysis of
People in Thermal Imagery [Internet] [Ph.d]. Aalborg
University; 2022 [cited 2025 Jun 28]. Available from:
https://vbn.aau.dk/en/publications/automatic-
analysis-of-people-in-thermal-imagery
30.
AI-Enabling Workloads on Large-Scale GPU-
Accelerated
System:
Characterization,
Opportunities, and Implications | IEEE Conference
Publication | IEEE Xplore [Internet]. [cited 2025 Jun
28].
Available
from:
https://ieeexplore.ieee.org/abstract/document/977
3216/
31.
Tapis: An API Platform for Reproducible, Distributed
Computational Research | SpringerLink [Internet].
[cited
2025
Jun
28].
Available
from:
https://link.springer.com/chapter/10.1007/978-3-
030-73100-7_61
