Pith.

Distributed, Parallel, and Cluster Computing · arXiv:2611.17206 v1 · submitted 12 Nov 2026 · 127 references

PACER: GPU Excursion Management for ML Workloads

Rutwik Jain, Nathaniel Morris, Yiwei Jiang, Shidhartha Das, Gary Kao, Indrani Paul, Shivaram Venkataraman, Srilatha Manne

show 1 more author

Matthew D. Sinclair

The paper says

“Variability is an inherent characteristic of modern GPU systems, arising from a combination of process variation, dynamic environmental conditions, and power management decisions. Prior work has characterized inter-GPU variability and found consistent performance differences across GPUs.”

more of the abstract

However, we observe that modern workloads, particularly large language models (LLMs), exhibit significant intra-GPU variability. Their short-running kernels interact poorly with the coarse granularity of dynamic power management mechanisms, leading to unpredictable performance. A key source of this intra-GPU transient variability is power excursions, also referred to as electrical design power peaks (EDPp), where power temporarily exceeds the GPU's rated power limits. While power management mechanisms are designed to ensure reliable operation despite the presence of EDPp sustained power peaks cause throttling that detrimentally impacts overall application performance. To the best of our knowledge, our work is the first to characterize the magnitude and frequency of EDPp, showing that frequent, significant power overshoots are especially common in ML workloads, which combine high power draw, short kernels, and rapidly changing compute demands. To address this, we design and leverage hardware and firmware mechanisms to provide consistent power management behavior that reduces intra-GPU variability due to EDPp and improves overall performance. We evaluate our techniques across a range of popular workloads. Our approach improves performance by 3-12% over state-of-the-art GPU power management and comes within 3% of an oracular solution.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 13 Nov 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. Solid characterization of GPU power excursions with a practical frequency-capping solution; the 3-12% improvement is real but the offline-to-online transfer is the weakest link. This line is wrong

4 major objections · 7 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 127 references say.

The paper says, at §4.2.1, Eq. (1); §6.2.2, Table 4; §6.2.2, Figure 14

The review did not quote this passage.

Pith says: Objection 1 of 4 · major

The central transfer claim—that offline p90 frequency caps remain near-optimal online—is not directly validated. The paper reports aggregate TTFT improvements (Figure 14) but does not present online per-kernel effective-frequency distributions, online EDP_p counts under PACER, or a sensitivity sweep around the chosen caps.

more of objection 1

Without such evidence, the reader cannot assess whether the caps are actually optimal or merely non-harmful. The authors acknowledge that effective frequency is context-dependent (§3.3) and that diminishing returns occur for long inputs (§6.2.2, sentence b319), but these acknowledgments do not substitute for a direct validation of the transfer. I recommend adding either (a) a comparison of online per-kernel p90 frequencies against the offline-derived caps, or (b) a cap-sensitivity study showing that small deviations from the chosen cap do not yield better TTFT.

The paper says, at §6.2.2, Figure 14; §5.2

The review did not quote this passage.

Pith says: Objection 2 of 4 · major

The online evaluation is limited to a single request trace (LongBench, 500 requests) and a narrow QPS range (0.1–0.3). The authors justify the QPS bound by saturation analysis (§5.2), but the trace size is small and the workload diversity is limited. The paper would benefit from evaluation on additional traces (e.g., ShareGPT, Azure LLM traces) and at higher loads to demonstrate robustness. The current results may not generalize to production traffic patterns with different input-length distributions or burstiness.

The paper says, at §4.2.1, Eq. (1); §6.2.1, Figure 10

The review did not quote this passage.

Pith says: Objection 3 of 4 · major

The oracle comparison (Figure 10) is performed only for single-GPU offline inference, not for the multi-GPU online serving scenario where the headline 12% improvement is reported. The claim 'within 3% of an oracular solution' is therefore established only offline. The online oracle would need to account for multi-GPU synchronization, RCCL overhead, and dynamic request mixing. The authors should either provide an online oracle estimate or qualify the claim to reflect the offline-only validation.

The paper says, at §4.2.1, Eq. (1); §6.2.2, Table 4

The review did not quote this passage.

Pith says: Objection 4 of 4 · major

The selection of top-K kernels (K such that they constitute >50% of runtime) is heuristic and may miss kernels that are infrequent but critical for EDP_p. The paper does not report the sensitivity of the predicted cap to the choice of K. A robustness analysis (e.g., varying K from 3 to all kernels) would strengthen the methodology.

7 minor objections

The paper says, at §1, Figure 1b

Pith says: minor objection 1 of 7

The caption of Figure 1b refers to 'CU Frequency (1ms)' and 'CU Frequency (50ms)', but the legend in the figure uses 'CU Frequency (a)' and 'CU Frequency (b)'. Please align the legend with the caption.

The paper says, at §3.1, Figure 3

Pith says: minor objection 2 of 7

The CDF in Figure 3 includes 'LLaMA2 fine-tuning' in the text but the legend lists only inference workloads. Clarify whether fine-tuning was included and, if so, report its configuration.

show 5 more

The paper says, at §4.2.1, Eq. (2)

Pith says: minor objection 3 of 7

The EWMA smoothing factor α is set to 0.3 for all experiments. The paper should discuss the sensitivity of results to this choice or justify why 0.3 is appropriate.

The paper says, at §5.1, Table 3

Pith says: minor objection 4 of 7

The table lists 'TTFT SLO' for online models but does not define how T_mean is computed. Specify whether it is the mean TTFT over all requests in the trace or over a subset.

The paper says, at §6.2.2, Figure 14

Pith says: minor objection 5 of 7

The y-axis label 'TTFT Improv (%)' is ambiguous. Clarify that it is percentage improvement over baseline (higher is better).

The paper says, at §7

Pith says: minor objection 6 of 7

The related work section mentions Kakolyris et al. [45] but does not discuss how PACER's approach differs in terms of objective (performance vs. energy) and phase (prefill vs. decode). A brief comparison would help position the contribution.

The paper says, at §2, Table 1a

Pith says: minor objection 7 of 7

The OCP specification table lists '1×TDP Indefinitely SW'. It would be helpful to note that this is a steady-state limit, not a peak.

What would settle it

Rerun the online deployment (LLaMA3.1-405B, TP=8, 0.2 QPS, same fixed-seed LongBench stream) with Stage 1's tracing enabled during the uncapped baseline, and recompute Eq. 1 per bucket from the in-situ per-kernel effective-frequency distributions, comparing the results to Table 4. Then A/B serve with the offline caps versus the in-situ recomputed caps. If any recomputed cap differs from Table 4 by more than ~150 MHz, if serving with the in-situ caps improves p50 TTFT by more than 3 percentage points over the Table 4 caps, or if PACER's online runs still show >1.3×TDP events at rates comparable to the baseline, the offline-profiling transfer fails and the 3–12% claim is not robust to the online context change.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 13 Nov 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic partly agrees 3 min 39 s · 33,987 tokens read, 24,981 written
  3. Editor wrote the verdict above 10 s · 35,000 tokens read, 661 written
  4. Referee major revision: 4 major objections, 7 minor 49 s · 34,427 tokens read, 1,459 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

This paper is worth reading if you care about GPU power management for LLM serving, but the headline 3–12% improvement is narrower than it sounds. The main contribution is a systematic characterization of EDP_p (power excursions above TDP) across ML, HPC, and graph workloads on AMD MI300X, plus a two-stage firmware solution (PACER) that caps GPU frequency based on offline profiling of per-kernel p90 effective frequencies.

What is actually new: the characterization is the first I have seen that quantifies how often and how high these excursions go for modern LLM inference, and the observation that low-magnitude EDP_p (up to ~1.3×TDP) is beneficial for computational sprinting while higher excursions trigger throttling that hurts performance. The PACER algorithm itself is a straightforward application of frequency capping, but the offline profiling to pick per-bucket caps based on kernel composition is a practical engineering contribution. The hardware DPE for sub-microsecond power estimation is also a real piece of work, though it is evaluated only indirectly.

What the paper does well: the experimental methodology is solid—three trials, comparisons against oracle and realistic baselines, explicit discussion of the 1 ms power measurement limitation, and a clear breakdown of why MoE models (DeepSeek) benefit less. The authors are honest about the offline profiling cost being a one-time hyperscaler expense. The kernel composition analysis in Figure 13 and the bell-curve behavior of TTFT improvement with input length are genuinely informative.

Soft spots: the stress-test concern about the transfer from offline p90 frequencies to online deployment is real and not fully addressed. The paper validates caps only through aggregate TTFT; it does not report online per-kernel frequency distributions or EDP_p counts under the caps. The "within 3% of an oracle" claim is established offline (Figure 10), not in the online setting where the 12% headline comes from. The online evaluation is limited to three models, one QPS bound (0.3), and no code release. These are acknowledged limitations, but they mean the central transfer claim is load-bearing and not stress-tested. The DPE estimation error (up to 12% NRMSE) is noted but its impact on the clamping threshold is not explored.

Who this is for: systems researchers working on GPU power management, especially at hyperscalers. The characterization alone is a useful reference. The PACER algorithm is practical but not theoretically deep.

Recommendation: this deserves peer review. The characterization is reproducible in principle (though no artifacts), the methodology is sound, and the limitations are stated. A good referee would ask for online per-kernel validation and a sensitivity analysis on the profiling assumptions, but the paper is not fatally flawed.

The strongest objection

PACER's central claim is a transfer claim: caps chosen by offline profiling (Section 4.2.1, Eq. 1) must remain near-optimal when deployed online. The weakest link is that the per-kernel p90 effective frequencies are measured in the uncapped baseline's specific DPM and power state, while Section 3.3 asserts that "the same GEMM computations may experience different power fluctuations and EDP_p, depending on what kernel precedes it, and how it interacts with the DPM" — effective frequency is context-dependent. Online deployment changes exactly those conditions: TP=8 introduces RCCL kernels into the profiled composition (Figure 13), Poisson load sustains back-to-back prefill near TDP, the EWMA of Eq. 2 mixes buckets, and chunked prefill inserts unprofiled boundaries. The paper validates the caps only through aggregate TTFT (Figure 14); it reports no online per-kernel frequency distributions, no online EDP_p counts under the caps, and no cap-sensitivity sweep. Two observed gaps make the concern concrete: online p50 gains (9–12% for LLaMA3.1-405B) run above the dominant bucket's offline gain (7.9% for 8k–16k tokens in Table 4) at every load, and the "within 3% of an oracle" bound is established only offline (Figure 10), not where the 12% headline figure arises. The acknowledged mitigations — b354 (one-time profiling cost), the uniformly nonnegative per-bucket gains in Table 4, and the flat kernel-duration curves in Figure 12 — bound the cost of modest mis-calibration, but they do not establish that the p90 statistic itself survives the context change.

Good-faith read: the end-to-end measurements are clean — a fixed-seed request stream identical for baseline and PACER, on production MI300X hardware, with major limitations stated explicitly (b98/b99 for power granularity, b319 for diminishing returns, b354 for profiling cost). The offline results (Figure 10) support the 3–4% improvement range and the near-oracle gap, and Table 4's per-bucket gains are uniformly nonnegative. The concern raised here is therefore not about whether the measured improvements occurred, but about whether the headline range transfers: the mechanism-level link between Stage 1's offline p90 caps and the online conditions is supported only by aggregate outcomes. The paper's own Section 3.3 makes effective frequency context-dependent, so that link is structurally unsecured, and the gap between online p50 gains and offline per-bucket gains suggests load-dependent amplification that is neither decomposed nor characterized. A stronger objection is not warranted because nothing contradicts the measured end-to-end A/B results, and the flatness of Figure 12 plus the nonnegative Table 4 cells indicate the heuristic is not knife-edged. The reader's weakest assumption (offline profiling representativeness) is in the same territory; this version differs by pinning the failure mode to the Section 3.3 context-dependence and by adding the online-versus-offline gain discrepancy and the missing online EDP_p evidence as concrete symptoms. The recommendation stays UNCHANGED: the reader's CONDITIONAL verdict already conditions on broader evaluation, and this analysis reinforces condition (1) rather than changing the verdict. The absence of a released artifact additionally limits independent verification, but that is a reproducibility limitation, not a soundness threat.

The referee's report

If the results hold, this work provides a practical, deployable solution to a real problem in GPU-accelerated ML serving: power excursions causing performance variability and throttling. The characterization of EDP_p across workloads is valuable. The firmware-only approach (PACER) requires no hardware changes and is demonstrated on production hardware, making it potentially impactful for hyperscalers. The paper is well-structured and the claims are supported by experiments on a relevant platform.

Pith's reading of the paper (a machine's summary, not the authors')

Modern GPUs can briefly exceed their rated power limit, a phenomenon called electrical design power peaks (EDP_p). These peaks occur when the GPU's power management system cannot react quickly enough to rapid changes in workload, especially in large language models (LLMs) that alternate between compute-heavy and memory-heavy kernels every few hundred microseconds. The paper first measures EDP_p across many workloads and finds that LLMs suffer from frequent, high-magnitude peaks that cause performance-degrading throttling. The authors acknowledge that their 1 ms power measurement granularity dampens events shorter than 1 ms, so very short, very high peaks are not fully captured. The authors then develop two solutions. First, a hardware digital power estimator (DPE) that monitors power at microsecond scales and can clamp peaks to a configurable threshold. Second, a firmware solution called PACER that sets an upper limit on GPU clock frequency based on offline profiling of each model's kernel behavior. The frequency cap is chosen per input length bucket to allow compute-intensive kernels to run fast while preventing the extreme peaks that trigger throttling. Evaluated on AMD Instinct MI300X GPUs, PACER improves time-to-first-token by 3-12% over default power management and comes within 3% of an oracle that knows the optimal frequency for each kernel. The benefits are largest for dense LLMs like LLaMA-3.1-405B, while mixture-of-experts models see smaller gains due to their more fractured kernel composition.

Core claim. PACER, a firmware frequency-capping solution built on offline kernel profiling, improves performance by 3-12% over state-of-the-art, production-grade GPU power management and comes within 3% of an oracular solution on AMD Instinct MI300X GPUs (Abstract, Section 4, Section 8).

The premise it rests on. The offline profiling step (Stage 1, Section 4.2.1) assumes that the kernel composition and per-kernel p90 effective frequencies measured during baseline runs are representative of online serving conditions with varying batch sizes, request rates, and input distributions. The authors acknowledge this as a one-time cost for hyperscalers (Section 7, sentence b354), but they do not evaluate how distribution shift affects the optimality of the precomputed frequency caps.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Variability is an inherent characteristic of modern GPU systems, arising from a combination of process variation, dynamic environmental conditions, and power management decisions. Prior work has characterized inter-GPU variability and found consistent performance differences across GPUs. However, we observe that modern workloads, particularly large language models (LLMs), exhibit significant intra-GPU variability. Their short-running kernels interact poorly with the coarse granularity of dynamic power management mechanisms, leading to unpredictable performance. A key source of this intra-GPU transient variability is power excursions, also referred to as electrical design power peaks (EDPp), where power temporarily exceeds the GPU's rated power limits. While power management mechanisms are designed to ensure reliable operation despite the presence of EDPp sustained power peaks cause throttling that detrimentally impacts overall application performance. To the best of our knowledge, our work is the first to characterize the magnitude and frequency of EDPp, showing that frequent, significant power overshoots are especially common in ML workloads, which combine high power draw, short kernels, and rapidly changing compute demands. To address this, we design and leverage hardware and firmware mechanisms to provide consistent power management behavior that reduces intra-GPU variability due to EDPp and improves overall performance. We evaluate our techniques across a range of popular workloads. Our approach improves performance by 3-12% over state-of-the-art GPU power management and comes within 3% of an oracular solution.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-14

The authors' main claim, as the classifier picked it out 0.91

“Our approach improves performance by 3-12% over state-of-the-art GPU power management and comes within 3% of an oracular solution.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.96 · states limitation 0.16 · human participants 0.02 · controlled comparison 0.86 · randomized assignment 0.01 · states quantitative result 0.98 · compares with prior results 0.96 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Variability is an inherent characteristic of modern GPU systems, arising from a combination of process variation, dynamic environmental conditions, and power management decisions. background 1.00 0.09 0.02 asserted 1.00
Prior work has characterized inter-GPU variability and found consistent performance differences across GPUs. background 1.00 0.03 0.02 asserted 1.00
However, we observe that modern workloads, particularly large language models (LLMs), exhibit significant intra-GPU variability. result 0.67 0.92 0.08 asserted 1.00
Their short-running kernels interact poorly with the coarse granularity of dynamic power management mechanisms, leading to unpredictable performance. background 0.46 0.53 0.04 asserted 1.00
A key source of this intra-GPU transient variability is power excursions, also referred to as electrical design power peaks (EDP$_{p}$), where power temporarily exceeds the GPU's rated power limits. background 0.86 0.47 0.06 asserted 0.98
While power management mechanisms are designed to ensure reliable operation despite the presence of EDP$_{p}$ sustained power peaks cause throttling that detrimentally impacts overall application performance. background 0.97 0.12 0.06 asserted 0.99
To the best of our knowledge, our work is the first to characterize the magnitude and frequency of EDP$_{p}$, showing that frequent, significant power overshoots are especially common in ML workloads, which combine high power draw, short kernels, and rapidly changing compute demands. contribution 0.70 0.96 0.16 asserted 0.99
To address this, we design and leverage hardware and firmware mechanisms to provide consistent power management behavior that reduces intra-GPU variability due to EDP$_{p}$ and improves overall performance. contribution 0.51 0.95 0.26 asserted 1.00
We evaluate our techniques across a range of popular workloads. method 0.85 0.90 0.02 asserted 0.81
Our approach improves performance by 3-12% over state-of-the-art GPU power management and comes within 3% of an oracular solution. result 1.00 0.96 0.66 asserted 1.00

What kind of paper the classifier reports this is: method or tool 0.98 · evidence from observation or experiment 1.00

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-14: 1 request, 7,607 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (14)
Figure 1 of the paper
Figure 1. (a) Thermal design power (TDP) of successive generations of AMD Instinct and NVIDIA datacenter
Figure 2 of the paper
Figure 2. Time slice showing kernels from two transformer layers of the prefill phase for LLaMA3-8B inference
Figure 3 of the paper
Figure 3. Cumulative Distribution (CDF) of EDP𝑝 above TDP for HPC, ML, and graph analytics applications on a single AMD Instinct MI300X GPU. until either maximum board power or maximum frequency is reached. This results in higher frequencies when running medium- or light-density compute operations and lower frequencies when running compute intensive applications with high switching capacitance and high power. Similar to many CPU power management algorithms, GPU PMFW uses past behavior as an indicator of future requirements. If application behavior does not change or changes slowly over many…
Figure 4 of the paper
Figure 4. (a) EDP𝑝 distribution and (b) cumulative distribution of kernel durations for prefill and decode phases of LLaMA3-8B inference with varying batch sizes (bsz) on an AMD Instinct MI300X GPU. any excursions for them. Therefore,
Figure 5 of the paper
Figure 5. Intra-kernel variation in normalized C𝐴𝐶 and arithmetic intensity in operations/s for GEMMs within LLaMA3-8B inference. While shorter kernels exacerbate EDP𝑝 due to fine-grained fluctuations in activity, making kernels longer through techniques like kernel fusion [4, 25, 56, 70, 102] or megaKernels [91, 92] does not resolve the problem. Longer kernels also incur intra-kernel power variability. For example, the computation in fused kernels (e.g., a compute-intensive GEMM and a memory￾intensive elementwise operation) differs significantly. Moreover, even within the GEMM phase, compute…
Figure 6 of the paper
Figure 6. (a) Timeseries showing absolute prediction error (left Y) and normalized measured and predicted C
Figure 7 of the paper
Figure 7. PACER’s two stages. Offline, a baseline inference run is traced per batched-token bucket to recover
Figure 8 of the paper
Figure 8. (a) Cumulative distribution of EDP𝑝 when boost management clamps power to different TDP levels; (b) Reduction in %EDP𝑝 at different clamping levels; (c) Timeseries showing varying power values for LLaMA3-8B inference with and without clamping; and (d) Performance impact relative to the unclamped baseline case (%). All results are for the prefill portion of LLaMA3-8B inference with 4096 input tokens and a batch size of 1. first show results for LLaMA3-8B inference (4096 input tokens, 128 output tokens, and batch size 1) where the DPE estimation error is about 6%. Then we show results…
Figure 9 of the paper
Figure 9. %EDP𝑝 and performance impact at different clamp levels for (a) varying batch size and input lengths for LLaMA3-8B inference and (b) different models with bsz 1 and 4096 input tokens. Mixtral-8x7B showing more high magnitude EDP𝑝 events than the LLaMA models. Mixtral-8x7B in particular suffers severe performance degradation when the threshold is limited to 1.0×TDP, reinforcing EDP𝑝 ’s sprinting benefits. Regardless, a clamping threshold of 1.2-1.3×TDP is sufficient to eliminate high-magnitude EDP𝑝 for all these models, while providing near-identical performance to the baseline. Thus,…
Figure 10 of the paper
Figure 10. %TTFT improvement over uncapped (2100 MHz
Figure 11 of the paper
Figure 11. Reduction in high-magnitude EDP𝑝 (normalized to 2100MHz run) with E2E Capping for a subset of configurations from
Figure 12 of the paper
Figure 12. Kernel duration scaling with frequency caps for the top 3 kernels (by kernel duration), normalized
Figure 13 of the paper
Figure 13. Prefill kernel composition by input-length bucket, normalized to the total prefill kernel time.
Figure 14 of the paper
Figure 14. Online serving results. Top row: reduction in TTFT with PACER over baseline at each percentile

Figures are the authors', reproduced from the paper.

References (127)
  1. [1] AMD. 2023. AMD CDNA™3 Architecture. (2023). https://www.amd.com/content/dam/amd/en/documents/instinct- tech-docs/white-papers/amd-cdna-3-white-paper.pdf on Pith
  2. [2] AMD. 2025. AMD CDNA™4 Architecture. (2025). https://www.amd.com/content/dam/amd/en/documents/instinct- tech-docs/white-papers/amd-cdna-4-architecture-whitepaper.pdf on Pith
  3. [3] AMD. 2024. ROCm SMI LIB Documentation, Release 7.2.0. Version 7.2.0. (2024). https://rocm.docs.amd.com/en/docs- 6.2.0/about/release-notes.html on Pith
  4. [4] Jeremy Appleyard, Tomás Kociský, and Phil Blunsom. 2016. Optimizing Performance of Recurrent Neural Networks on GPUs.CoRR, abs/1604.01946. http://arxiv.org/abs/1604.01946 on PithPith's reviewarXiv
  5. [5] Meena Arunachalam, Miro Hodak, Wei-Ting Liao, Karan Verma, Ean Garvey, Kumar Deepak, Giuseppe Franco, and Eliot Li. 2025. Reproducing the AMD Instinct™GPUs MLPerf Inference v5.0 Submission. (2025). https://rocm.blogs. amd.com/artificial-intelligence/reproducing-amd-mlperf-infe... on Pith
  6. [6] Nick Baker, Vishesh Vikas, Andrew Lemmon, Chase Fortin, Justin Conzola, Arthur Boutry, and John Austin on Pith
  7. [7] Claude Bernard, Michael C. Ogilvie, Thomas A. Degrand, Carleton E. Detar, Steven A. Gottlieb, A. Krasnitz, R.L. Sugar, and D. Toussaint. 1991. Studying Quarks and Gluons On MIMD Parallel Computers.Int. J. High Perform. Comput. Appl., 5, 4, (Dec. 1991), 61–70. doi:10.1177/10943... on Pithdoi:10.1177/109434209100500406
  8. [8] Srikant Bharadwaj, Shomit Das, Kaushik Mazumdar, Bradford M. Beckmann, and Stephen Kosonocky. 2023. Predict; Don’t React for Enabling Efficient Fine-Grain DVFS in GPUs. InProceedings of the 28th ACM International Conference on Architectural Support for Programming Languages an... on Pithdoi:10.1145/3623278.3624756
  9. [9] Paolo Boldi, Marco Rosa, Massimo Santini, and Sebastiano Vigna. 2011. Layered label propagation: a multiresolution coordinate-free ordering for compressing social networks. InProceedings of the 20th international conference on World Wide Web. Sadagopan Srinivasan, Krithi Ramam... on Pith
  10. [10] Paolo Boldi and Sebastiano Vigna. 2004. The WebGraph framework I: Compression techniques. InProc. of the Thirteenth International World Wide Web Conference (WWW 2004). ACM Press, Manhattan, USA, 595–601 on Pith
  11. [11] Cadence System Analysis. 2021. Power Electronics Reliability and Temperature Dependency. https://resources. system-analysis.cadence.com/blog/msa2021-power-electronics-reliability-and-temperature-dependency. (2021) on Pith
  12. [12] Mithun Chandrasekhar. 2019. AMD Radeon Community Update: More Control Over GPU Power and Performance, Enhanced Thermal Monitoring, Maximized Performance. https://community.amd.com/t5/gaming/amd-radeon- community-update-more-control-over-gpu-power-and/ba-p/418629. (2019) on Pith
  13. [13] Tzu-Tao Chang, Benjamin Yuanyang Hong, Kiet Pham, and Shivaram Venkataraman. 2027. TeDiServe: High SLO Attainment Serving for Diffusion Language Models. In22nd European Conference on Computer Systems(EuroSys). doi:10.1145/3842654.3848543 on Pithdoi:10.1145/3842654.3848543
  14. [14] Dimitrios Chasapis, Marc Casas, Miquel Moretó, Martin Schulz, Eduard Ayguadé, Jesus Labarta, and Mateo Valero on Pith
  15. [15] Dimitrios Chasapis, Miquel Moretó, Martin Schulz, Barry Rountree, Mateo Valero, and Marc Casas. 2019. Power Efficient Job Scheduling by Predicting the Impact of Processor Manufacturing Variability. InProceedings of the ACM International Conference on Supercomputing(ICS ’19), 2... on Pithdoi:10.1145/3330345.3330372
  16. [16] Esha Choukse, Brijesh Warrier, Scot Heath, Luz Belmont, April Zhao, Hassan Ali Khan, Brian Harry, Matthew Kappel, Russell J. Hewett, Kushal Datta, Yu Pei, Caroline Lichtenberger, John Siegler, David Lukofsky, Zaid Kahn, Gurpreet Sahota, Andy Sullivan, Charles Frederick, Hien T... on PithPith's reviewarXiv
  17. [17] J. Coplin and M. Burtscher. 2016. Energy, Power, and Performance Characterization of GPGPU Benchmark Programs. InIEEE International Parallel and Distributed Processing Symposium Workshops(IPDPSW). (May 2016), 1190–1199. doi:10.1109/IPDPSW.2016.164 on Pithdoi:10.1109/ipdpsw.2016.164
  18. [18] Xingqi Cui, Chieh-Jan Mike Liang, Jiarong Xing, and Haoran Qiu. 2025. From Models to Operators: Rethinking Autoscaling Granularity for Large Generative Models. (2025). https://arxiv.org/abs/2511.02248 arXiv: 2511.02248 [cs.DC] on PitharXiv
  19. [19] Abhishek Das, Serkan Ozdemir, Gokhan Memik, Joseph Zambreno, and Alok Choudhary. 2007. Mitigating the Effects of Process Variations: Architectural Approaches for Improving Batch Performance. InWorkshop on Architectural Support for Gigascale Integration(ASGI) on Pith
  20. [20] Daniele De Sensi, Tiziano De Matteis, Konstantin Taranov, Salvatore Di Girolamo, Tobias Rahn, and Torsten Hoefler on Pith
show 80 more
  1. [21] Abhimanyu Dubey et al. 2024. The Llama 3 Herd of Models.arXiv preprint arXiv:2407.21783. https://arxiv.org/abs/ 2407.21783 arXiv: 2407.21783[cs.AI] on PithPith's reviewarXiv
  2. [22] Bishwajit Dutta, Vignesh Adhinarayanan, and Wu-chun Feng. 2018. GPU Power Prediction via Ensemble Machine Learning for DVFS Space Exploration. InProceedings of the 15th ACM International Conference on Computing Frontiers(CF). Association for Computing Machinery, Ischia, Italy,... on Pith
  3. [23] Kaijie Fan, Biagio Cosenza, and Ben Juurlink. 2019. Predictable GPUs Frequency Scaling for Energy and Performance. InProceedings of the 48th International Conference on Parallel Processing(ICPP 2019) Article 52. Association for Computing Machinery, Kyoto, Japan, 10 pages.isbn:... on Pithdoi:10.1145/3337821.3337833
  4. [24] Songchun Fan, Seyed Majid Zahedi, and Benjamin C. Lee. 2016. The Computational Sprinting Game. InProceedings of the Twenty-First International Conference on Architectural Support for Programming Languages and Operating Systems(ASPLOS ’16). Association for Computing Machinery, ... on Pithdoi:10.1145/2872362.2872383
  5. [25] Jiří Filipovič, Matúš Madzin, Jan Fousek, and Ludundefinedk Matyska. 2015. Optimizing CUDA Code by Kernel Fusion: Application on BLAS.The Journal of Supercomputing, 71, 10, (Oct. 2015), 3934–3957. doi:10.1007/s11227-015- 1483-z on Pithdoi:10.1007/s11227-015-
  6. [26] R. Ge, R. Vogt, J. Majumder, A. Alam, M. Burtscher, and Z. Zong. 2013. Effects of Dynamic Voltage and Frequency Scaling on a K20 GPU. In42nd International Conference on Parallel Processing(ICPP), 826–833. doi:10.1109/ICPP. 2013.98 on Pithdoi:10.1109/icpp
  7. [27] Aaron Grenat, Sanjay Pant, Ravinder Rachala, and Samuel Naffziger. 2014. 5.6 adaptive clocking system for improved power efficiency in a 28nm x86-64 microprocessor. In2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC), 106–107. doi:10.11... on Pithdoi:10.1109/isscc.2014.6757358
  8. [28] João Guerreiro, Aleksandar Ilic, Nuno Roma, and Pedro Tomás. 2019. DVFS-aware application classification to improve GPGPUs energy efficiency.Parallel Computing, 83, 93–117. doi:https://doi.org/10.1016/j.parco.2018.02.001 on Pithdoi:10.1016/j.parco.2018.02.001
  9. [29] Chao Guo and Zhonghai Lu. 2023. Damage-based Lifetime Modeling for Power Electronic Devices. InPHM Society Asia-Pacific Conferencenumber 1. Vol. 4 on Pith
  10. [30] Hilbert Hagedoorn. 2025. Rubin Ultra GPU to Use Advanced Microchannel Cooling for its 2300W TDP. https: //www.guru3d.com/story/rubin-ultra-gpu-to-use-advanced-microchannel-cooling-for-its-2300w-tdp/. (Oct. 2025) on Pith
  11. [31] Gabriel Hautreux and Etienne Malaboeuf. 2023. Reducing HPC energy footprint for large scale GPU accelerated workloads. InProceedings of the SC ’23 Workshops of the International Conference on High Performance Computing, Network, Storage, and Analysis(SC-W ’23). Association for... on Pithdoi:10.1145/3624062.3624268
  12. [32] Felix Hoffmann, Stefan Schmitt, and Nando Kaminski. 2023. Lifetime Modeling of SiC MOSFET Power Modules During Power Cycling Tests at Low Temperature Swings. In35th International Symposium on Power Semiconductor Devices and ICs(ISPSD), 294–297. doi:10.1109/ISPSD57135.2023.10147533 on Pithdoi:10.1109/ispsd57135.2023.10147533
  13. [33] Ke Hong, Xiuhong Li, Lufang Chen, Qiuli Mao, Guohao Dai, Xuefei Ning, Shengen Yan, Yun Liang, and Yu Wang on Pith
  14. [34] R. D. Hornung, Jeff A. Keasler, and M. B. Gokhale. 2011. Hydrodynamics Challenge Problem, Lawrence Livermore National Laboratory. Tech. rep. LLNL-TR-490254. Lawrence Livermore National Laboratory, Livermore, CA, 1–28. PACER: GPU Excursion Management for ML Workloads 23 on Pith
  15. [35] Farzad Hosseinabadi, Sajib Chakraborty, Sachin Kumar Bhoi, Guenter Prochart, Dino Hrvanovic, and Omar Hegazy on Pith
  16. [36] Yanping Huang, Youlong Cheng, Ankur Bapna, Orhan Firat, Mia Xu Chen, Dehao Chen, HyoukJoong Lee, Jiquan Ngiam, Quoc V. Le, Yonghui Wu, and Zhifeng Chen. 2019. GPipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism. InProceedings of the 33rd International... on Pith
  17. [37] Ranggi Hwang, Jianyu Wei, Shijie Cao, Changho Hwang, Xiaohu Tang, Ting Cao, and Mao Yang. 2025. Pre-Gated MoE: An Algorithm-System Co-Design for Fast and Scalable Mixture-of-Expert Inference. InProceedings of the 51st Annual International Symposium on Computer Architecture(ISC... on Pithdoi:10.1109/isca59077.2024.00078
  18. [38] Yuichi Inadomi, Tapasya Patki, Koji Inoue, Mutsumi Aoyagi, Barry Rountree, Martin Schulz, David Lowenthal, Yasutaka Wada, Keiichiro Fukazawa, Masatsugu Ueda, Masaaki Kondo, and Ikuo Miyoshi. 2015. Analyzing and Mitigating the Impact of Manufacturing Variability in Power-Constr... on Pithdoi:10.1145/2807591.2807638
  19. [39] Intel Corporation. 2003. Voltage regulator module (vrm) and enterprise voltage regulator-down (evrd) 10.0 design guidelines. https://www.intel.com/content/dam/www/public/us/en/documents/design-guides/voltage-regulator- module-enterprise-voltage-regulator-down-10-0-guidelines.p... on Pith
  20. [40] A comprehensive overview of reliability assessment strategies and testing of power electronics converters. IEEE Open Journal of Power Electronics, 5, 473–512. doi:10.1109/OJPEL.2024.3379294 on Pithdoi:10.1109/ojpel.2024.3379294
  21. [41] Rutwik Jain, Brandon Tran, Keting Chen, Matthew D. Sinclair, and Shivaram Venkataraman. 2024. PAL: A Variability- Aware Policy for Scheduling ML Workloads in GPU Clusters. InProceedings of the International Conference for High Performance Computing, Networking, Storage, and An... on Pith
  22. [42] N. E. Jerger, A. Kannan, Z. Li, and G. H. Loh. 2014. NoC Architectures for Silicon Interposer Systems: Why Pay for more Wires when you Can Get them (from your interposer) for Free? In47th Annual IEEE/ACM International Symposium on Microarchitecture(MICRO). (Dec. 2014), 458–470... on Pithdoi:10.1109/micro.2014.61
  23. [43] Albert Q. Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, Gianna Lengyel, Guillaume Bour, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne... on PithPith's reviewarXiv
  24. [44] Zhixing Jiang, Justin Garrigus, Allison Seigler, Ethan Syed, Yanlun Huang, Mehdi Sadi, Tawfik Rahal-Arabi, and Lizy Kurian John. 2026. Exploration of LLM Workload Reliability based on di/dt effects and Voltage Droops. InIEEE International Symposium on High-Performance Comp Arc... on Pith
  25. [45] Rutwik Jain, Yiwei Jiang, Matthew D. Sinclair, and Shivaram Venkataraman. 2026. Minos: Systematically Classifying Performance and Power Characteristics of GPU Workloads on HPC Clusters. (2026). https://arxiv.org/abs/2604.03591 arXiv: 2604.03591[cs.DC] on PithPith's reviewarXiv
  26. [46] Aditya K. Kamath, Ramya Prabhu, Jayashree Mohan, Simon Peter, Ramachandran Ramjee, and Ashish Panwar on Pith
  27. [47] Ajaykumar Kannan, Natalie Enright Jerger, and Gabriel H Loh. 2015. Enabling Interposer-based Disintegration of Multi-core Processors. In48th Annual IEEE/ACM International Symposium on Microarchitecture(MICRO). IEEE, 546–558 on Pith
  28. [48] Ian Karlin. 2012. LULESH Programming Model and Performance Ports Overview. Tech. rep. LLNL-TR-608824. Lawrence Livermore National Laboratory, Livermore, CA, (Dec. 2012), 1–17 on Pith
  29. [49] Ian Karlin, Abhinav Bhatele, Jeff Keasler, Bradford L. Chamberlain, Jonathan Cohen, Zachary DeVito, Riyaz Haque, Dan Laney, Edward Luke, Felix Wang, David Richards, Martin Schulz, and Charles Still. 2013. Exploring traditional and emerging parallel programming models using a p... on Pith
  30. [50] Andreas Kosmas Kakolyris, Dimosthenis Masouros, Sotirios Xydis, and Dimitrios Soudris. 2024. SLO-Aware GPU DVFS for Energy-Efficient LLM Inference Serving.IEEE Computer Architecture Letters, 23, 2, 150–153. doi:10.1109/ LCA.2024.3406038 on Pith
  31. [51] Donggyu Kim, Jerry Zhao, Jonathan Bachrach, and Krste Asanović. 2019. Simmani: runtime power modeling for arbitrary rtl with automatic signal selection. InProceedings of the 52nd Annual IEEE/ACM International Symposium 24 Jain et al. on Microarchitecture(MICRO-52). Association... on Pithdoi:10.1145/3352460.3358322
  32. [52] POD-Attention: Unlocking Full Prefill-Decode Overlap for Faster LLM Inference. InProceedings of the 30th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2 (ASPLOS). Association for Computing Machinery, Rotterdam, Ne... on Pith
  33. [53] Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. InProceedings of the ACM SIGOPS 29th Symposium on Operating Sy... on Pith
  34. [54] Lawrence Livermore National Labs. 2020. CORAL-2 Benchmarks. https://asc.llnl.gov/coral-2-benchmarks. (2020) on Pith
  35. [55] Jingwen Leng, Alper Buyuktosunoglu, Ramon Bertran, Pradip Bose, Yazhou Zu, and Vijay Janapa Reddi. 2021. Predictive Guardbanding: Program-Driven Timing Margin Reduction for GPUs.IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 40, 1, 171–184. doi:... on Pithdoi:10.1109/tcad.2020.2992684
  36. [56] Ian Karlin, Jeff Keasler, and Rob Neely. 2013. LULESH 2.0 Updates and Changes. Tech. rep. LLNL-TR-641973. Lawrence Livermore National Laboratory, Livermore, CA, (Aug. 2013), 1–9 on Pith
  37. [57] Zhuohan Li, Lianmin Zheng, Yinmin Zhong, Vincent Liu, Ying Sheng, Xin Jin, Yanping Huang, Zhifeng Chen, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. 2023. AlpaServe: Statistical Multiplexing with Model Parallelism for Deep Learning Serving. In17th USENIX Symposium on Operati... on Pith
  38. [58] Ivana F Kovacevic-Badstuebner, Johann W Kolar, Uwe Schilling, H Chung, H Wang, F Blaabjerg, and M Pecht. 2015. Modelling for the Lifetime Prediction of Power Semiconductor Modules.Reliability of Power Electronic Converter Systems, 103–140 on Pith
  39. [59] Gabriel H. Loh, Natalie Enright Jerger, Ajaykumar Kannan, and Yasuko Eckert. 2015. Interconnect-Memory Chal- lenges for Multi-chip, Silicon Interposer Systems. InProceedings of the 2015 International Symposium on Memory Systems(MEMSYS ’15). ACM, Washington DC, DC, USA, 3–10.is... on Pithdoi:10.1145/2818950.2818951
  40. [60] Lingxiao Ma, Zhiqiang Xie, Zhi Yang, Jilong Xue, Youshan Miao, Wei Cui, Wenxiang Hu, Fan Yang, Lintao Zhang, and Lidong Zhou. 2020. Rammer: Enabling Holistic Deep Learning Compiler Optimizations with rTasks. In14th USENIX Symposium on Operating Systems Design and Implementatio... on Pith
  41. [61] Alexandre Marie, Bernardo Cougo, Loic Renaudie, Trésor Kaounodji Koladoum, and Jean-Pierre Fradin. 2024. Precise 3D Modelling of SiC Die Temperature Oscillation for Lifetime Prediction of Power Modules used in DC/AC Power Converters. In30th International Workshop on Thermal In... on Pithdoi:10.1109/therminic62015.2024.10732697
  42. [62] Ao Li, Bojian Zheng, Gennady Pekhimenko, and Fan Long. 2020. Automatic Horizontal Fusion for GPU Kernels. (2020). arXiv: 2007.01277[cs.DC] on PithPith's reviewarXiv
  43. [63] Xinxin Mei, Ling Sing Yung, Kaiyong Zhao, and Xiaowen Chu. 2013. A Measurement Study of GPU DVFS on Energy Conservation. InProceedings of the Workshop on Power-A ware Computing and Systems(HotPower) Article on Pith
  44. [64] Qunyou Liu, Darong Huang, Marina Zapater, and David Atienza. 2025. GreenLLM: SLO-Aware Dynamic Frequency Scaling for Energy-Efficient LLM Serving.arXiv preprint arXiv:2508.16449. https://arxiv.org/abs/2508.16449 arXiv: 2508.16449[cs.PF] on PithPith's reviewarXiv
  45. [65] Nathaniel Morris, Siva Meenakshi Renganathan, Christopher Stewart, Robert Birke, and Lydia Chen. 2016. Sprint Ability: How Well Does Your Software Exploit Bursts in Processing Capacity? InIEEE International Conference on Autonomic Computing(ICAC), 173–178. doi:10.1109/ICAC.2016.61 on Pithdoi:10.1109/icac.2016.61
  46. [66] Nathaniel Morris, Christopher Stewart, Robert Birke, Lydia Chen, and Jaimie Kelley. 2017. Early Work on Modeling Computational Sprinting. InProceedings of the 2017 Symposium on Cloud Computing(SoCC ’17). Association for Computing Machinery, Santa Clara, California, 661.isbn: 9... on Pithdoi:10.1145/3127479.3132691
  47. [67] Rajib Nath and Dean Tullsen. 2015. The CRISP Performance Model for Dynamic Voltage and Frequency Scaling in a GPGPU. InProceedings of the 48th International Symposium on Microarchitecture(MICRO). Association for Computing Machinery, Waikiki, Hawaii, 281–293.isbn: 9781450340342... on Pithdoi:10.1145/2830772.2830826
  48. [68] Peter Mattson, Christine Cheng, Cody Coleman, Greg Diamos, Paulius Micikevicius, David A. Patterson, Hanlin Tang, Gu-Yeon Wei, Peter Bailis, Victor Bittorf, David Brooks, Dehao Chen, Debojyoti Dutta, Udit Gupta, Kim M. Hazelwood, Andrew Hock, Xinyuan Huang, Bill Jia, Daniel Ka... on PithPith's reviewarXiv
  49. [69] NVIDIA. 2024. NVIDIA Blackwell Architecture Technical Brief. https://resources.nvidia.com/en-us-blackwell- architecture/blackwell-architecture-technical-brief. (2024) on Pith
  50. [70] Association for Computing Machinery, Farmington, Pennsylvania, 5 pages.isbn: 9781450324588. doi:10.1145/ 2525526.2525852 on Pith
  51. [71] Ashish Mishra and Nilay Khare. 2015. Analysis of DVFS Techniques for Improving the GPU Energy Efficiency. http://file.scirp.org/pdf/ojee_2015121415504865.pdf. (2015) on Pith
  52. [72] Hiroshi Onodera, Nobuyuki Shishido, Daisuke Asari, Hiroshi Isono, and Wataru Saito. 2024. Effect of Solder Junction Void Variation in Power Semiconductor Package on Power Cycle Lifetime.Microelectronics Reliability, 161, 115471. doi:https://doi.org/10.1016/j.microrel.2024.115471 on Pithdoi:10.1016/j.microrel.2024.115471
  53. [73] Muhammad Osama, Duane Merrill, Cris Cecka, Michael Garland, and John D. Owens. 2023. Stream-K: Work-centric Parallel Decomposition for Dense Matrix-Matrix Multiplication on the GPU. (2023). https://arxiv.org/abs/2301.03598 arXiv: 2301.03598[cs.DS] on PithPith's reviewarXiv
  54. [74] Sangyoung Park, Jaehyun Park, Donghwa Shin, Yanzhi Wang, Qing Xie, Massoud Pedram, and Naehyuck Chang on Pith
  55. [75] Cedric Nugteren, Gert-Jan van den Braak, and Henk Corporaal. 2014. Roofline-Aware DVFS for GPUs. InProceed- ings of International Workshop on Adaptive Self-Tuning Computing Systems(ADAPT). Association for Computing Machinery, Vienna, Austria, 8–10.isbn: 9781450325141. doi:10.1... on Pithdoi:10.1145/2553062.2553067
  56. [76] Pratyush Patel, Esha Choukse, Chaojie Zhang, Aashaka Shah, Íñigo Goiri, Saeed Maleki, and Ricardo Bianchini. 2025. Splitwise: Efficient Generative LLM Inference Using Phase Splitting. InProceedings of the 51st Annual International Symposium on Computer Architecture(ISCA). IEEE... on Pithdoi:10.1109/isca59077.2024.00019
  57. [77] NVIDIA. 2025. NVIDIA cuDNN: GPU Accelerated Deep Learning. https://developer.nvidia.com/cudnn. (2025) on Pith
  58. [78] OCP Open Accelerator Infrastructure (OAI) Workstreams. 2022. OCP Universal Baseboard (UBB) Design Specification v1.5. https://www.opencompute.org/documents/universal-baseboard-design-specification-v1p5-final-20220223- docx-pdf. (2022) on Pith
  59. [79] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. 2023. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. (2023). https: //arxiv.org/abs/2307.01952 arXiv: 2307.01952[cs.CV] on PithPith's reviewarXiv
  60. [80] Arun Raghavan, Yixin Luo, Anuj Chandawalla, Marios Papaefthymiou, Kevin P. Pipe, Thomas F. Wenisch, and Milo M. K. Martin. 2012. Computational Sprinting. InIEEE International Symposium on High-Performance Computer Architecture, 1–12. doi:10.1109/HPCA.2012.6169031 on Pithdoi:10.1109/hpca.2012.6169031
  61. [81] Tawfik Rahal-Arabi, Paul Van der Arend, Ashish Jain, Mehdi Saidi, Rashad Oreifej, Sriram Sundaram, Srilatha Manne, Indrani Paul, Rajit Seahra, Frank Helms, Esha Choukse, Nithish Mahalingam, Brijesh Warrier, and Ricardo Bianchini on Pith
  62. [82] Vijay Janapa Reddi, Christine Cheng, David Kanter, Peter Mattson, Guenther Schmuelling, Carole-Jean Wu, Brian Anderson, Maximilien Breughe, Mark Charlebois, William Chou, Ramesh Chukka, Cody Coleman, Sam Davis, Pan Deng, Greg Diamos, Jared Duke, Dave Fick, J. Scott Gardner, It... on Pith
  63. [83] Pratyush Patel, Esha Choukse, Chaojie Zhang, Inigo Goiri, Brijesh Warrier, Nithish Mahalingam, and Ricardo Bianchini. 2024. Characterizing Power Management Opportunities for LLMs in the Cloud. InASPLOS. (Apr. 2024). https://doi.org/10.1145/3620666.3651329 on Pithdoi:10.1145/3620666.3651329
  64. [84] Skyler J Saleh, Samuel Naffziger, Milind S Bhagavat, and Rahul Agarwal. 2020. GPU Chiplets Using High Bandwidth Crosslinks. US Patent App. 16/456,287. (Dec. 2020) on Pith
  65. [85] Pratyush Patel, Zibo Gong, Syeda Rizvi, Esha Choukse, Pulkit Misra, Thomas Anderson, and Akshitha Sriraman on Pith
  66. [86] Christian Schwabe, Niklas Seltner, and Thomas Basler. 2024. Reliability of SiC MOSFETs in the High Cycle Fatigue Regime under Fast Power Pulses.Solid State Phenomena, 361, 7–12 on Pith
  67. [87] Suchita Pati, Shaizeen Aga, Nuwan Jayasena, and Matthew Sinclair. 2025. GOLDYLOC: Global Optimizations & Lightweight Dynamic Logic for Concurrency.ACM Trans. Archit. Code Optim., 22, 2, Article 78, (July 2025), 28 pages. doi:10.1145/3730584 on Pithdoi:10.1145/3730584
  68. [88] Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. 2020. Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism. (2020). https://arxiv. org/abs/1909.08053 arXiv: 1909.08053[cs.CL] on PithPith's reviewarXiv
  69. [89] Prasoon Sinha, Akhil Guliani, Rutwik Jain, Brandon Tran, Matthew D Sinclair, and Shivaram Venkataraman. 2022. Not All GPUs Are Created Equal: Characterizing Variability in Large-Scale, Accelerator-Rich Systems. InProceedings of the International Conference on High Performance ... on Pith
  70. [90] Alan Smith, Gabriel H. Loh, Michael J. Schulte, Mike Ignatowski, Samuel Naffziger, Mike Mantor, Nathan Kalyana- sundharam, Vamsi Alla, Nicholas Malaya, Joseph L. Greathouse, Eric Chapman, and Raja Swaminathan. 2024. Realizing the AMD Exascale Heterogeneous Processor Vision : I... on Pithdoi:10.1109/isca59077.2024.00068
  71. [91] Optimizing GPU Data Center Power. InIEEE Asia Pacific Conference on Circuits and Systems(APCCAS), 358–362. doi:10.1109/APCCAS62602.2024.10808935 on Pithdoi:10.1109/apccas62602.2024.10808935
  72. [92] Benjamin Spector, Jordan Juravsky, Stuart Sul, Dylan Lim, Dan Fu, Simran Arora, and Chris Ré. 2025. Look Ma, No Bubbles! Designing a Low-Latency Megakernel for Llama-1B. https://hazyresearch.stanford.edu/blog/2025-05-27- no-bubbles. Accessed: 2025-11-16. (2025) on Pith
  73. [93] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2021. High-resolution image synthesis with latent diffusion models. (2021). eprint: 2112.10752 (cs.CV) on PithPith's reviewarXiv
  74. [94] Jovan Stojkovic, Chaojie Zhang, Inigo Goiri, Josep Torrellas, and Esha Choukse. 2025. DynamoLLM: Designing LLM Inference Clusters for Performance and Energy Efficiency. InIEEE International Symposium on High Performance Computer Architecture(HPCA), 1348–1362. doi:10.1109/HPCA6... on Pithdoi:10.1109/hpca61900.2025.00102
  75. [95] Vito Savino. 2025. Rethinking Power for AI Processors. InPower Systems Design. (Apr. 2025). https : / / www. powersystemsdesign.com/articles/rethinking-power-for-ai-processors/159/22717 on Pith
  76. [96] Bo Su, Junli Gu, Li Shen, Wei Huang, Joseph L. Greathouse, and Zhiying Wang. 2014. PPEP: Online Performance, Power, and Energy Prediction Framework and DVFS Space Exploration. In2014 47th Annual IEEE/ACM International Symposium on Microarchitecture, 445–457. doi:10.1109/MICRO.2014.17 on Pithdoi:10.1109/micro.2014.17
  77. [97] Thomas Scogland, Jonathan Azose, David Rohr, Suzanne Rivoire, Natalie Bates, and Daniel Hackenberg. 2015. Node Variability in Large-Scale Power Measurements: Perspectives from the Green500, Top500 and EEHPCWG. In 26 Jain et al. Proceedings of the International Conference for H... on Pith
  78. [98] Burak Topcu, Deniz Karabacak, and Isil Oz. 2025. Demystifying Power and Performance Variations in GPU Systems through Microarchitectural Analysis.Computer Science and Information Systems, 22, 2, 533–561. doi:https://doi.org/ 10.2298/CSIS240722021T on Pithdoi:10.2298/csis240722021t
  79. [99] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023. LLaMA: Open and Efficient Foundation La... on PithPith's reviewarXiv
  80. [100] Daniel Velicka, Ondrej Vysocky, and Lubomir Riha. 2025. Methodology for gpu frequency switching latency measurement. (2025). https://arxiv.org/abs/2502.20075 arXiv: 2502.20075[cs.DC] on PithPith's reviewarXiv

The first 100 of 127 references are listed here. All of them are in this paper's citation record.

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment