pith. sign in

arxiv: 2605.11333 · v3 · pith:3VSRKNJ6new · submitted 2026-05-11 · 💻 cs.DC · cs.LG· cs.PF

MLCommons Chakra: Advancing Performance Benchmarking and Co-design using Standardized Execution Traces

Srinivas Sridharan , Theodor-Adrian Badea , Andy Balogh , Bradford M. Beckmann , Brian Coutinho , Louis Feng , Sheng Fu , Sanshan Gao
show 21 more authors
Mehryar Garakani Taekyung Heo David Kanter Josh Ladd Ziwei Li Winston Liu Changhai Man Dan Mihailescu Spandan More Joongun Park Ashwin Ramachandran Vinay Ramakrishnaiah Saeed Rashidi Vijay Janapa Reddi Puneet Sharma Phio Tian William Won Hanjiang Wu Huan Xu Jinsun Yoo Tushar Krishna
This is my paper

Pith reviewed 2026-05-20 22:13 UTC · model grok-4.3

classification 💻 cs.DC cs.LGcs.PF
keywords execution tracesdistributed MLperformance benchmarkinghardware-software co-designAI systemsworkload modelingsimulators
0
0 comments X

The pith

Chakra defines an open graph-based execution trace format to standardize observation and co-design of distributed AI/ML workloads.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Chakra as an ecosystem built around Chakra execution traces, which are graph representations that encode compute, memory, communication operations plus their data and control dependencies, timing, and resource constraints. This format aims to let researchers and engineers collect traces from real production clusters, then replay or simulate them consistently across different tools without relying on proprietary data. The goal is faster iteration on workload optimization and hardware-software co-design by making traces portable and interoperable. Industry adoption through MLCommons is presented as evidence that the approach can scale beyond individual labs.

Core claim

The central contribution is the Chakra execution trace, an open and interoperable graph-based representation of distributed AI/ML workloads that captures key operations such as compute, memory, and communication together with data and control dependencies, timing, and resource constraints, accompanied by tools for trace collection, analysis, generation, and use by simulators, emulators, and replay systems.

What carries the argument

Chakra execution trace (ET), a graph-based representation that encodes operations, dependencies, timing, and constraints to enable analysis and replay across independent tools.

If this is right

  • Traces collected on existing AI clusters can be reused to evaluate new hardware designs without re-running the full workload.
  • Different research groups can compare optimization results using identical ETs instead of custom trace formats.
  • Co-design loops can alternate between trace generation from software changes and simulation on proposed hardware.
  • Industry-wide benchmarking becomes more reproducible once tools adopt the common ET format.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If ETs become widely adopted, the cost of developing new performance tools could drop because each tool would only need to read one format rather than many proprietary ones.
  • The same graph structure might later support automated search for better workload mappings or communication schedules.
  • Extending the format to include power or thermal constraints could link performance traces directly to energy-efficiency studies.

Load-bearing premise

A single graph-based execution trace format can sufficiently capture the essential operations, dependencies, timing, and resource constraints of production-scale distributed ML workloads so that many different simulators and emulators can use the same traces effectively.

What would settle it

Run the same production-scale workload trace through several independent simulators and check whether they produce performance predictions that match each other and real hardware measurements within a small error margin.

Figures

Figures reproduced from arXiv: 2605.11333 by Andy Balogh, Ashwin Ramachandran, Bradford M. Beckmann, Brian Coutinho, Changhai Man, Dan Mihailescu, David Kanter, Hanjiang Wu, Huan Xu, Jinsun Yoo, Joongun Park, Josh Ladd, Louis Feng, Mehryar Garakani, Phio Tian, Puneet Sharma, Saeed Rashidi, Sanshan Gao, Sheng Fu, Spandan More, Srinivas Sridharan, Taekyung Heo, Theodor-Adrian Badea, Tushar Krishna, Vijay Janapa Reddi, Vinay Ramakrishnaiah, William Won, Winston Liu, Ziwei Li.

Figure 1
Figure 1. Figure 1: AI system SW-HW co-design flow [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: AI system SW-HW co-design flow. to reproduce behaviors in different environments. Simula￾tors and emulators are a dime-a-dozen across NPU compute and networking vendors, of varying degrees of fidelity. Each of these have their own custom formats for describing work￾loads and the AI platform architecture. This fragmentation creates barriers to platform-agnostic analysis and co-design, and limits the opportu… view at source ↗
Figure 2
Figure 2. Figure 2: Chakra Infrastructure Overview. to describe distributed AI workload performance behavior over an AI platform. Analogous to instruction and mem￾ory traces (Ranganathan & Victor), ETs record operator dimensions for compute and communication and their de￾pendencies while avoiding disclosure of model or dataset details. Software organizations can share ETs of internal workloads with hardware vendors, who can i… view at source ↗
Figure 3
Figure 3. Figure 3 [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Chakra ET visualization example. broad tasks: analysis, replay and simulation/emulation that are required at different times within the development cycle of AI platforms. The open schema enables interoperability across different stages and diverse open/proprietary tools. 4.1 Trace Analysis Chakra offers a range of open-source tools to help users vi￾sualize, analyze, and consume execution traces. We describ… view at source ↗
Figure 6
Figure 6. Figure 6: Normalized execution time breakdown across workloads for traces collected on the system mentioned in Sec. 5. For each workload, we show measured performance from Kineto (left) and the performance via trace reconstruction through Chakra (right). AllToAll AllGather ReduceScatter AllReduce Collective Communication Type 0.0 0.2 0.4 0.6 0.8 1.0 Total Duration (µs) 1e7 4.1× slower 4.4× slower 1.5× slower 9.7× sl… view at source ↗
Figure 7
Figure 7. Figure 7: Total collective communication runtime comparison at 400 Gb/s and 100 Gb/s InfiniBand. Measured on training Mixtral￾8×22B with 32 GPUs (four HGX-8×H200 nodes, TP/SP=4, EP=8) and the global batch size of 32. open-source tools like Genie (Yoo et al., 2026b) as well as commercial system emulators like Keysight AI Data Center Builder (Keysight Technologies, 2025), which now support the Chakra format for worklo… view at source ↗
Figure 8
Figure 8. Figure 8: GPU memory utilization for different LLM models dur￾ing one training step. Traces are aligned relative to the start of each epoch. Each model and its corresponding parallelization match the first entry (row) in [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Compute characteristics of the Mixtral-8x22-Chakra trace. (a) Most compute kernels complete within 2–102 µs. (b) The majority of nodes have 10–500 parent data dependencies. 5.2 Trace Replay Case Studies Replaying Chakra ETs on real systems allows reproduc￾ing the exact workload behavior either fully (replay both compute and comms operations) or partial replay (replay selective operations). The latter enabl… view at source ↗
Figure 10
Figure 10. Figure 10: Bus bandwidth per iteration when (a) All-Reduce (b) All-to-All (c) mixing All-to-All and All-Reduce in one time span. AllReduce1 AllReduce10 AllReduce2 AllReduce3 AllReduce4 AllReduce5 AllReduce6 AllReduce7 AllReduce8 AllReduce9 AllToAll1 AllToAll10 AllToAll2 AllToAll3 AllToAll4 AllToAll5 AllToAll6 AllToAll7 AllToAll8 AllToAll9 Percentile 100 80 60 40 20 0 Completion Time (ms) 5 10 15 20 25 30 35 40 45 50… view at source ↗
Figure 11
Figure 11. Figure 11: Mixing collectives results of CDF. Result. The experiment revealed a significant performance anomaly when interleaving All-Reduce and All-to-All col￾lectives. While both collectives show stable performance in isolation ( [PITH_FULL_IMAGE:figures/full_fig_p010_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Communication time for different network topology and bandwidth with Mixtral 8x7B target. connected. Additionally, we test bandwidths ranging from 75 GB/s to 900 GB/s. The Mixtral 8×7B model serves as the workload for this evaluation. Results [PITH_FULL_IMAGE:figures/full_fig_p011_12.png] view at source ↗
Figure 5
Figure 5. Figure 5 [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 14
Figure 14. Figure 14: Distribution of token routing among two expert parallel rank for each model layer. The input has six tokens and the model used is Mixtral 8x7B with 32 layers. 0 4 8 12 16 20 24 28 Model Layer ID 110 120 130 140 150 160 170 180 190 KV Transfer Duration (µs) Send (prefill) Recv (decode) [PITH_FULL_IMAGE:figures/full_fig_p012_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Runtime breakdown of the KV cache transfer for in￾ferencing Llama3-8B between one prefill and decode GPU. The captured trace denotes the per-layer (32 layers for Llama3-8B) send and receive latency between two GPUs. 5.5.3 KV-Cache Transfer In inference, when disaggregating prefill and decode stages on different GPUs (Patel et al., 2024; Zhong et al., 2024; Bambhaniya et al., 2026), it introduces unique po… view at source ↗
read the original abstract

The fast pace of artificial intelligence~(AI) innovation demands an agile methodology for observation, reproduction and optimization of distributed machine learning~(ML) workload behavior in production AI systems and enables efficient software-hardware~(SW-HW) co-design for future systems. We present Chakra, an open and portable ecosystem for performance benchmarking and co-design. The core component of Chakra is an open and interoperable graph-based representation of distributed AI/ML workloads, called Chakra execution trace~(ET). These ETs represent key operations, such as compute, memory, and communication, data and control dependencies, timing, and resource constraints. Additionally, Chakra includes a complementary set of tools and capabilities to enable the collection, analysis, generation, and adoption of Chakra ETs by a broad range of simulators, emulators, and replay tools. We present analysis of Chakra ETs collected on production AI clusters and demonstrate value via real-world case studies. Chakra has been adopted by MLCommons and has active contributions and engagement across the industry, including but not limited to NVIDIA, AMD, Meta, Keysight, HPE, and Scala, to name a few.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces Chakra, an open and portable ecosystem for performance benchmarking and co-design of distributed AI/ML workloads. Its core is the Chakra execution trace (ET), a graph-based representation capturing compute, memory, and communication operations along with data/control dependencies, timing, and resource constraints. Complementary tools support collection, analysis, generation, and adoption of ETs by simulators, emulators, and replay tools. The work includes analysis of production AI cluster traces and real-world case studies, and notes adoption by MLCommons with contributions from NVIDIA, AMD, Meta, Keysight, HPE, and Scala.

Significance. A standardized, interoperable ET format with supporting tools could meaningfully advance reproducible benchmarking and SW-HW co-design for large-scale AI systems by enabling portable workload reproduction across independent tools. The open-source nature, MLCommons adoption, and broad industry engagement are concrete strengths that increase the likelihood of impact if the representation proves sufficiently accurate. However, the current lack of quantitative reproduction fidelity metrics and validation against non-deterministic production effects limits the assessed significance.

major comments (2)
  1. [Abstract] Abstract: the central claim that ETs enable effective analysis, reproduction, and co-design across simulators rests on the representation of timing and resource constraints, yet the text provides no explicit mechanisms (e.g., stochastic timing distributions or feedback loops) for non-deterministic effects such as network jitter or collective algorithm selection. This is load-bearing because a purely static graph risks under-specifying contention at scale, directly affecting the asserted interoperability value.
  2. [Analysis and case studies] The analysis of production traces and real-world case studies section: no quantitative results, error bounds, or fidelity comparisons (e.g., simulated vs. measured runtime or bandwidth utilization) are reported. Without such data the utility claim for co-design remains plausible but unverified, weakening the soundness assessment.
minor comments (1)
  1. The ET schema and encoding details would benefit from an explicit table or figure early in the manuscript to aid readers in understanding the graph structure and attribute fields.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed review. The comments on non-deterministic effects and quantitative fidelity metrics identify important opportunities to strengthen the presentation of Chakra's capabilities. We have revised the manuscript accordingly and address each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that ETs enable effective analysis, reproduction, and co-design across simulators rests on the representation of timing and resource constraints, yet the text provides no explicit mechanisms (e.g., stochastic timing distributions or feedback loops) for non-deterministic effects such as network jitter or collective algorithm selection. This is load-bearing because a purely static graph risks under-specifying contention at scale, directly affecting the asserted interoperability value.

    Authors: The Chakra ET is intentionally a faithful recording of an observed execution, so the included timing values and dependencies already embed the non-deterministic effects (including network jitter) that occurred during trace collection. The static graph structure is deliberate: it guarantees portability and interoperability across independent simulators, emulators, and replay tools. Different simulators are free to overlay stochastic timing distributions, feedback loops, or collective-selection models on top of the provided timings and constraints. We have revised the abstract and added a short clarifying paragraph in the ET specification section to make this extensibility explicit while preserving the core static representation. revision: partial

  2. Referee: [Analysis and case studies] The analysis of production traces and real-world case studies section: no quantitative results, error bounds, or fidelity comparisons (e.g., simulated vs. measured runtime or bandwidth utilization) are reported. Without such data the utility claim for co-design remains plausible but unverified, weakening the soundness assessment.

    Authors: We agree that quantitative fidelity metrics would strengthen the soundness assessment. The original manuscript prioritized qualitative analysis and evidence of industry adoption. In the revised version we have added a dedicated subsection that reports simulated-versus-measured runtime differences, bandwidth utilization comparisons, and associated error bounds for representative workloads drawn from the production traces. These results were obtained by replaying Chakra ETs through available simulators and contrasting the outputs with direct cluster measurements. revision: yes

Circularity Check

0 steps flagged

No circularity: proposal of new ET format with no derivation chain

full rationale

The paper presents Chakra as a new open graph-based execution trace representation and supporting ecosystem for ML workload benchmarking and co-design. No equations, fitted parameters, predictions, or first-principles derivations are described that could reduce to self-defined inputs or self-citation chains. The central claims rest on the definition and adoption of the ET format itself, collection from production clusters, and case studies, all of which are independent of any prior author equations or fitted quantities. This is a standardization and tooling contribution rather than a closed mathematical derivation, so no circular steps exist.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

This is a systems and benchmarking paper whose contribution is the definition of a new trace format and tool ecosystem rather than mathematical derivations or fitted models. No free parameters appear in the abstract. The main invented element is the Chakra execution trace representation itself.

invented entities (1)
  • Chakra execution trace (ET) no independent evidence
    purpose: Graph-based representation capturing compute, memory, communication operations, dependencies, timing, and resource constraints for distributed ML workloads
    New standardized format introduced to enable interoperability across tools and hardware; no independent falsifiable evidence provided in the abstract.

pith-pipeline@v0.9.0 · 5858 in / 1357 out tokens · 40398 ms · 2026-05-20T22:13:59.141512+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

101 extracted references · 101 canonical work pages · 8 internal anchors

  1. [1]

    vLLM Github , howpublished =

    work page

  2. [2]

    Ultra Ethernet Consortium , howpublished =

    work page

  3. [3]

    2026 , eprint=

    Flint: Compiler Enabled Cluster-Free Design Space Exploration for Distributed ML , author=. 2026 , eprint=

    work page 2026

  4. [5]

    2024 , howpublished =

    Man, C , title =. 2024 , howpublished =

    work page 2024

  5. [6]

    2025 , eprint=

    Lumos: Efficient Performance Modeling and Estimation for Large-scale LLM Training , author=. 2025 , eprint=

    work page 2025

  6. [7]

    Jia, Zhihao and Zaharia, Matei and Aiken, Alex , booktitle=

    work page

  7. [8]

    2020 , organization=

    Rajbhandari, Samyam and Rasley, Jeff and Ruwase, Olatunji and He, Yuxiong , booktitle=. 2020 , organization=

    work page 2020

  8. [9]

    Ranganathan, Parthasarathy and Lee Victor , url =

    work page

  9. [10]

    Manpreet Singh Minhas , url =

    work page

  10. [11]

    2023 , eprint=

    Chakra: Advancing Performance Benchmarking and Co-design using Standardized Execution Traces , author=. 2023 , eprint=

    work page 2023

  11. [12]

    2020 , organization=

    Rashidi, Saeed and Sridharan, Srinivas and Srinivasan, Sudarshan and Krishna, Tushar , booktitle=. 2020 , organization=

    work page 2020

  12. [13]

    IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS) , year=

    William Won and Taekyung Heo and Saeed Rashidi and Srinivas Sridharan and Sudarshan Srinivasan and Tushar Krishna , title=. IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS) , year=

    work page

  13. [14]

    2011 , publisher=

    Rodrigues, Arun F and Hemmert, K Scott and Barrett, Brian W and Kersey, Chad and Oldfield, Ron and Weston, Marlo and Risen, Rolf and Cook, Jeanine and Rosenfeld, Paul and Cooper-Balis, Elliot and others , journal=. 2011 , publisher=

    work page 2011

  14. [15]

    2019 , organization=

    Wang, Fei and Chen, Guoyang and Zhang, Weifeng and Rompf, Tiark , booktitle=. 2019 , organization=

    work page 2019

  15. [16]

    Proceedings of the 22nd USENIX Symposium on Networked Systems Design and Implementation , articleno =

    Wang, Xizheng and Li, Qingxu and Xu, Yichi and Lu, Gang and Li, Dan and Chen, Li and Zhou, Heyang and Zheng, Linkang and Zhang, Sen and Zhu, Yikai and Liu, Yang and Zhang, Pengcheng and Qian, Kun and He, Kunling and Gao, Jiaqi and Zhai, Ennan and Cai, Dennis and Fu, Binzhang , title =. Proceedings of the 22nd USENIX Symposium on Networked Systems Design a...

    work page 2025

  16. [17]

    Santhanam, Keshav and Krishna, Siddharth and Tomioka, Ryota and Harris, Tim and Zaharia, Matei , journal=

    work page

  17. [18]

    Schaarschmidt, Michael and Grewe, Dominik and Vytiniotis, Dimitrios and Paszke, Adam and Schmid, Georg Stefan and Norman, Tamara and Molloy, James and Godwin, Jonathan and Rink, Norman Alexander and Nair, Vinod and others , journal=

    work page

  18. [19]

    and Lee, Jaewon and Lundell, John and Kim, Changkyu and Kejariwal, Arun and Owens, John D

    Lin, Zhongyi and Feng, Louis and Ardestani, Ehsan K. and Lee, Jaewon and Lundell, John and Kim, Changkyu and Kejariwal, Arun and Owens, John D. , booktitle=. Building a Performance Model for Deep Learning Recommendation Model Training on GPUs , year=

    work page

  19. [20]

    DreamShard: Generalizable Embedding Table Placement for Recommender Systems , url =

    Zha, Daochen and Feng, Louis and Tan, Qiaoyu and Liu, Zirui and Lai, Kwei-Herng and Bhushanam, Bhargav and Tian, Yuandong and Kejariwal, Arun and Hu, Xia , booktitle =. DreamShard: Generalizable Embedding Table Placement for Recommender Systems , url =

    work page

  20. [21]

    Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    Autoshard: Automated embedding table sharding for recommender systems , author=. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    work page

  21. [23]

    Mingyu Liang and Wenyin Fu and Louis Feng and Zhongyi Lin and Pavani Panakanti and Shengbao Zheng and Srinivas Sridharan and Christina Delimitrou , year=

    work page

  22. [24]

    FacebookResearch , title =

    work page

  23. [25]

    Facebook Research , title =

    work page

  24. [26]

    Zenodo , year=

    TensorFlow , author=. Zenodo , year=

    work page

  25. [27]

    2017 , howpublished =

    ONNX , title =. 2017 , howpublished =

    work page 2017

  26. [28]

    Advances in neural information processing systems , volume=

    Pytorch: An imperative style, high-performance deep learning library , author=. Advances in neural information processing systems , volume=

    work page

  27. [29]

    Proceedings of the 53rd IEEE/ACM International Symposium on Computer Architecture (ISCA '26) , year =

    Scalable Synthesis of LLM Benchmarks through Symbolic Tensor Graphs , author =. Proceedings of the 53rd IEEE/ACM International Symposium on Computer Architecture (ISCA '26) , year =

    work page

  28. [30]

    2024 , eprint=

    Splitwise: Efficient generative LLM inference using phase splitting , author=. 2024 , eprint=

    work page 2024

  29. [32]

    2026 , month = dec, howpublished =

    vLLM , title =. 2026 , month = dec, howpublished =

    work page 2026

  30. [33]

    Keysight AI Data Center Builder , year =

    work page

  31. [34]

    2015 , eprint=

    Deep Residual Learning for Image Recognition , author=. 2015 , eprint=

    work page 2015

  32. [35]

    2019 , eprint=

    Deep Learning Recommendation Model for Personalization and Recommendation Systems , author=. 2019 , eprint=

    work page 2019

  33. [36]

    2024 , eprint=

    DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models , author=. 2024 , eprint=

    work page 2024

  34. [37]

    2024 , eprint=

    The Llama 3 Herd of Models , author=. 2024 , eprint=

    work page 2024

  35. [38]

    2024 , eprint=

    Mixtral of Experts , author=. 2024 , eprint=

    work page 2024

  36. [39]

    2020 , eprint=

    Language Models are Few-Shot Learners , author=. 2020 , eprint=

    work page 2020

  37. [40]

    2024 , eprint=

    DistServe: Disaggregating Prefill and Decoding for Goodput-optimized Large Language Model Serving , author=. 2024 , eprint=

    work page 2024

  38. [41]

    Chakra Execution Traces: Benchmarking Network Performance Optimization , year =

    work page

  39. [42]

    2023 , month = jul, howpublished =

    Chakra: Advancing Benchmarking and Co-design for Future AI Systems , author =. 2023 , month = jul, howpublished =

    work page 2023

  40. [43]

    2026 , eprint=

    Understanding and Optimizing Multi-Stage AI Inference Pipelines , author=. 2026 , eprint=

    work page 2026

  41. [44]

    The 13th International Conference on Learning Representations (ICLR) , year=

    LayerDAG: A Layerwise Autoregressive Diffusion Model of Directed Acyclic Graphs for System , author=. The 13th International Conference on Learning Representations (ICLR) , year=

    work page

  42. [45]

    2025 , urldate =

    Kineto: A CPU. 2025 , urldate =

    work page 2025

  43. [46]

    2025 , publisher =

    Bortok, Alex , title =. 2025 , publisher =

    work page 2025

  44. [47]

    2025 , publisher =

    Bergeron, Matt and Kumar, Ashutosh , title =. 2025 , publisher =

    work page 2025

  45. [48]

    2025 , publisher =

    Wareing, Richard and Graf, Tyler , title =. 2025 , publisher =

    work page 2025

  46. [49]

    ASTRA-sim: Scalable System-Level Simulation Framework for Large-Scale Machine Learning Systems , author =

    work page

  47. [50]

    2025 , month = feb, url =

    deepseek-ai , title =. 2025 , month = feb, url =

    work page 2025

  48. [52]

    2024 , pages=

    Won, William and Rashidi, Saeed and Srinivasan, Sudarshan and Krishna, Tushar , booktitle=. 2024 , pages=

    work page 2024

  49. [53]

    TACOS: Topology-Aware Collective Algorithm Synthesizer for Distributed Machine Learning , year=

    Won, William and Elavazhagan, Midhilesh and Srinivasan, Sudarshan and Gupta, Swati and Krishna, Tushar , booktitle=. TACOS: Topology-Aware Collective Algorithm Synthesizer for Distributed Machine Learning , year=

    work page

  50. [54]

    2021 ACM/IEEE 48th Annual International Symposium on Computer Architecture (ISCA) , pages=

    Enabling compute-communication overlap in distributed deep learning training platforms , author=. 2021 ACM/IEEE 48th Annual International Symposium on Computer Architecture (ISCA) , pages=. 2021 , organization=

    work page 2021

  51. [55]

    Proceedings of the 52nd Annual International Symposium on Computer Architecture , pages=

    FRED: A Wafer-scale Fabric for 3D Parallel DNN Training , author=. Proceedings of the 52nd Annual International Symposium on Computer Architecture , pages=

    work page

  52. [56]

    The Scala Compute Platform , author =

    work page

  53. [57]

    Georgia Tech AI Makerspace , author =

    work page

  54. [58]

    Astra-sim: Scalable system-level simulation framework for large-scale machine learning systems

    ASTRA-sim . Astra-sim: Scalable system-level simulation framework for large-scale machine learning systems. https://astra-sim.github.io/, 2025. Accessed: 2025-10-30

    work page 2025

  55. [59]

    MIST: A Co-Design Framework for Heterogeneous, Multi-Stage LLM Inference

    Bambhaniya, A. R., Wu, H., Subramanian, S., Srinivasan, S., Kundu, S., Yazdanbakhsh, A., Elavazhagan, M., Kumar, M., and Krishna, T. Understanding and optimizing multi-stage ai inference pipelines, 2026. URL https://arxiv.org/abs/2504.09775

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  56. [60]

    and Kumar, A

    Bergeron, M. and Kumar, A. Accelerating AI Hardware NPI - Clusterless Validation of GPUs and Networking . https://youtu.be/-PRs1eVF3nY?si=sYl3P0tSsmEAJOL2, 2025. OCP Global Summit

    work page 2025

  57. [61]

    Methodology and Observation of Congestion Control Impact on MoE Training Job Completion Time

    Bortok, A. Methodology and Observation of Congestion Control Impact on MoE Training Job Completion Time . https://youtu.be/nLSDrgvu-qw?si=EJnOJ__zB35delA1, 2025. OCP Global Summit

    work page 2025

  58. [62]

    Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, ...

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  59. [63]

    DeepSeekMoE: Towards Ultimate Expert Specialization in Mixture-of-Experts Language Models

    Dai, D., Deng, C., Zhao, C., Xu, R. X., Gao, H., Chen, D., Li, J., Zeng, W., Yu, X., Wu, Y., Xie, Z., Li, Y. K., Huang, P., Luo, F., Ruan, C., Sui, Z., and Liang, W. Deepseekmoe: Towards ultimate expert specialization in mixture-of-experts language models, 2024. URL https://arxiv.org/abs/2401.06066

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  60. [64]

    Holistic trace analysis

    FacebookResearch. Holistic trace analysis. https://github.com/facebookresearch/HolisticTraceAnalysis. Accessed: 2025-09-28

    work page 2025

  61. [65]

    [PyTorch] Integrate Execution Graph Observer into PyTorch Profiler

    Feng, L. [PyTorch] Integrate Execution Graph Observer into PyTorch Profiler . URL https://github.com/pytorch/pytorch/pull/75358

    work page

  62. [66]

    Georgia tech ai makerspace

    Georgia Institute of Technology . Georgia tech ai makerspace. https://coe.gatech.edu/academics/ai-for-engineering/ai-makerspace, 2026. Accessed: 2026-04-06

    work page 2026

  63. [67]

    Characterizing the efficiency of distributed training: A power, performance, and thermal perspective

    Go, S., Park, J., More, S., Wu, H., Wang, I., Jezghani, A., Krishna, T., and Mahajan, D. Characterizing the efficiency of distributed training: A power, performance, and thermal perspective. In Proceedings of the 58th IEEE/ACM International Symposium on Microarchitecture, MICRO '25, pp.\ 626–642, New York, NY, USA, 2025. Association for Computing Machiner...

    work page doi:10.1145/3725843.3756111 2025

  64. [68]

    Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., Yang, A., Fan, A., Goyal, A., Hartshorn, A., Yang, A., Mitra, A., Sravankumar, A., Korenev, A., Hinsvark, A., Rao, A., Zhang, A., Rodriguez, A., Gregerson, A., Spataru, A., Roziere, B., Biron, B., Tang, B., Chern, B., Caucheteux...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  65. [69]

    Deep Residual Learning for Image Recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition, 2015. URL https://arxiv.org/abs/1512.03385

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  66. [70]

    Mixtral of Experts

    Jiang, A. Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D. S., de las Casas, D., Hanna, E. B., Bressand, F., Lengyel, G., Bour, G., Lample, G., Lavaud, L. R., Saulnier, L., Lachaux, M.-A., Stock, P., Subramanian, S., Yang, S., Antoniak, S., Scao, T. L., Gervet, T., Lavril, T., Wang, T., Lacroix, T., and Sayed, W. E. Mixtral...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  67. [71]

    Infrastructure Graph

    Keysight. Infrastructure Graph . URL https://infragraph.dev/

    work page

  68. [72]

    Keysight ai data center builder

    Keysight Technologies . Keysight ai data center builder. https://github.com/Keysight/kai-dc-builder/releases/download/v1.0.2/Keysight.AI.Data.Center.Builder.Solution.Brief.pdf, 2025. Solution brief. Accessed: 2026-03-31

    work page 2025

  69. [73]

    Layerdag: A layerwise autoregressive diffusion model of directed acyclic graphs for system

    Li, M., Shitole, V., Chien, E., Man, C., Wang, Z., Zhang, Y., Krishna, T., Li, P., et al. Layerdag: A layerwise autoregressive diffusion model of directed acyclic graphs for system. In The 13th International Conference on Learning Representations (ICLR), 2024

    work page 2024

  70. [74]

    Mystique: Enabling Accurate and Scalable Generation of Production AI Benchmarks

    Liang, M., Fu, W., Feng, L., Lin, Z., Panakanti, P., Zheng, S., Sridharan, S., and Delimitrou, C. Mystique: Enabling Accurate and Scalable Generation of Production AI Benchmarks . In Proceedings of the 50th Annual International Symposium on Computer Architecture (ISCA), 2023

    work page 2023

  71. [75]

    Lumos: Efficient performance modeling and estimation for large-scale llm training, 2025

    Liang, M., Kassa, H. T., Fu, W., Coutinho, B., Feng, L., and Delimitrou, C. Lumos: Efficient performance modeling and estimation for large-scale llm training, 2025. URL https://arxiv.org/abs/2504.09307

    work page arXiv 2025

  72. [76]

    symbolic\_tensor\_graph: A symbolic tensor graph generator for astra-sim

    Man, C. symbolic\_tensor\_graph: A symbolic tensor graph generator for astra-sim. https://github.com/astra-sim/symbolic_tensor_graph, 2024. Accessed: 2025-10-29

    work page 2024

  73. [77]

    Scalable synthesis of llm benchmarks through symbolic tensor graphs

    Man, C., Park, J., Wu, H., Xu, H., Sridharan, S., and Krishna, T. Scalable synthesis of llm benchmarks through symbolic tensor graphs. In Proceedings of the 53rd IEEE/ACM International Symposium on Computer Architecture (ISCA '26), 2026

    work page 2026

  74. [78]

    Chakra execution traces: Benchmarking network performance optimization

    Meta . Chakra execution traces: Benchmarking network performance optimization. https://engineering.fb.com/2023/09/07/networking-traffic/chakra-execution-traces-benchmarking-network-performance-optimization/, September 2023. Accessed: 2026-03-31

    work page 2023

  75. [79]

    Minhas, M. S. Computational Graphs in PyTorch and TensorFlow . URL https://towardsdatascience.com/computational-graphs-in-pytorch-and-tensorflow-c25cc40bdcd1. Accessed: 2025-10-01

    work page 2025

  76. [80]

    Chakra: Advancing benchmarking and co-design for future ai systems

    MLCommons . Chakra: Advancing benchmarking and co-design for future ai systems. https://mlcommons.org/2023/07/chakra-advancing-benchmarking-and-co-design-for-future-ai-systems/, July 2023. Accessed: 2026-03-31

    work page 2023

  77. [81]

    MLPerf Storage Benchmark

    MLPerf. MLPerf Storage Benchmark . URL https://mlcommons.org/benchmarks/storage/

    work page

  78. [82]

    Deep Learning Recommendation Model for Personalization and Recommendation Systems

    Naumov, M., Mudigere, D., Shi, H.-J. M., Huang, J., Sundaraman, N., Park, J., Wang, X., Gupta, U., Wu, C.-J., Azzolini, A. G., Dzhulgakov, D., Mallevich, A., Cherniavskii, I., Lu, Y., Krishnamoorthi, R., Yu, A., Kondratenko, V., Pereira, S., Chen, X., Chen, W., Rao, V., Jia, B., Xiong, L., and Smelyanskiy, M. Deep learning recommendation model for persona...

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  79. [83]

    NVIDIA NeMo

    NeMo . NVIDIA NeMo . https://www.nvidia.com/en-us/ai-data-science/products/nemo/. Accessed: 2025-10-01

    work page 2025

  80. [84]

    Splitwise: Efficient generative llm inference using phase splitting, 2024

    Patel, P., Choukse, E., Zhang, C., Shah, A., Íñigo Goiri, Maleki, S., and Bianchini, R. Splitwise: Efficient generative llm inference using phase splitting, 2024. URL https://arxiv.org/abs/2311.18677

    work page arXiv 2024

Showing first 80 references.