Pith. sign in

REVIEW 4 major objections 7 minor 3 cited by

Pangu Ultra MoE: How to Train Your Big MoE on Ascend NPUs

T0 review · 4 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A 718-billion-parameter mixture-of-experts model can be trained on Ascend NPUs at 30.0% model FLOPs utilization, with benchmark scores comparable to DeepSeek R1.

desk verdict A credible industrial recipe for training a 718B MoE on Ascend NPUs, but the headline 30% MFU is unauditable without the FLOPs formula and the simulator re-validation on the final model is in-sample. read the letter →

arxiv 2505.04519 v1 pith:D5BVIOYI submitted 2025-05-07 cs.CL

Yehui Tang , Yichun Yin , Yaoyuan Wang , Hang Zhou , Yu Pan , Wei Guo , Ziyang Zhang , Miao Rang
show 66 more authors
Fangcheng Liu Naifu Zhang Binghan Li Yonghan Dong Xiaojun Meng Yasheng Wang Dong Li Yin Li Dandan Tu Can Chen Youliang Yan Fisher Yu Ruiming Tang Yunhe Wang Botian Huang Bo Wang Boxiao Liu Changzheng Zhang Da Kuang Fei Liu Gang Huang Jiansheng Wei Jiarui Qin Jie Ran Jinpeng Li Jun Zhao Liang Dai Lin Li Liqun Deng Peifeng Qin Pengyuan Zeng Qiang Gu Shaohua Tang Shengjun Cheng Tao Gao Tao Yu Tianshu Li Tianyu Bi Wei He Weikai Mao Wenyong Huang Wulong Liu Xiabing Li Xianzhi Yu Xueyu Wu Xu He Yangkai Du Yan Xu Ye Tian Yimeng Wu Yongbing Huang Yong Tian Yong Zhu Yue Li Yufei Wang Yuhang Gai Yujun Li Yu Luo Yunsheng Ni Yusen Sun Zelin Chen Zhe Liu Zhicheng Liu Zhipeng Tu Zilin Ding Zongyuan Zhan
This is my paper · ORCID
classification cs.CL
keywords MixtureofExpertsAscendNPUmodelFLOPsutilizationexpertparallelismloadbalancingsimulation-basedarchitecturesearchmemoryoptimizationlargelanguagetraining
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is a system-and-architecture recipe for training a near-trillion-parameter mixture-of-experts language model on Ascend NPUs rather than on GPU clusters. It claims that careful model-shape selection guided by a hardware-aware simulator, together with communication, memory, and load-balancing optimizations, lets a 718-billion-parameter MoE reach 30.0% model FLOPs utilization on 6K Ascend NPUs while producing benchmark scores comparable to DeepSeek R1. The central assertion is that the sparse-model efficiency promised by MoE can be materialized on a non-NVIDIA accelerator stack. If true, it means the practical ceiling on large-model training is not set by the choice of accelerator vendor, and that simulation can replace much of the expensive trial-and-error in choosing model configurations.

What carries the argument

The load-bearing machinery is a two-part design loop. First, a bottom-up simulator models operator throughput, communication, memory, and overlaps on Ascend 910B hardware, validated to 88.9% and 90.1% accuracy on two training runs, and is used to prune a search space of about 10,000 architectures down to the Pangu Ultra MoE configuration. Second, the system-side optimizations turn that configuration into a working 6K-NPU run: hierarchical EP all-to-all communication (a global inter-node AllGather followed by an intra-node All-to-All exchange) to cut inter-node traffic, an adaptive pipe overlap schedule that hides forward/backward communication behind computation, fine-grained recomputation plus tensor swapping to manage activation memory, fused operators, and a planner/executor pair that predicts expert load and physically moves experts across devices. These mechanisms jointly convert sparse-model dynamic routing from a throughput tax into a recovered efficiency gain.

What would settle it

Take the paper's own reported numbers—718B total parameters, roughly 39B activated per token, and 1.46M tokens per second—and recompute the FLOPs used per second, then divide by the number of Ascend NPUs times their peak floating-point throughput; if the resulting 30.0% MFU only holds under a lenient FLOPs formula, the comparison to prior MFU reports fails. Independently, rerunning the evaluation suite with a public harness against officially published DeepSeek R1 scores would settle whether the benchmark parity is real.

Watch

Extended reading notes

Core claim

The paper's central discovery is a repeatable training recipe for large sparse MoE models on Ascend NPUs. It combines simulation-based architecture search that screens roughly 10,000 candidate configurations by predicted throughput and selects 61 layers, hidden size 7680, 256 routed experts plus one shared expert, MLA attention, and a specific 5D parallelism plan (TP = 8, PP = 16, VPP = 2, EP = 4), with system optimizations including hierarchical EP all-to-all communication, adaptive pipe overlap, fine-grained recomputation with tensor swapping, and dynamic device-level expert placement. On this basis the authors report training Pangu Ultra MoE, 718B total parameters with 39B activated per token, at 30.0% MFU and 1.46M tokens per second on 6K Ascend NPUs, up from an 18.9% MFU baseline on 4K NPUs, with benchmark performance comparable to or better than DeepSeek R1 on most evaluated tasks, especially medical benchmarks. The paper also reports that dropless routing outperforms drop-and-pad increasingly as models grow, and that an EP-group auxiliary loss strikes the best performance/efficiency balance among sequence-, micro-batch-, EP-group-, and DP-group-level losses.

Load-bearing premise

The single load-bearing premise is that the efficiency metric behind the 30.0% figure (the share of peak computing power actually used) and the benchmark numbers produced by the authors' own testing are counted with the same conventions used for DeepSeek's published figures; if either convention differs, the headline comparison is not apples-to-apples.

Editorial extensions

If this is right

  • If the 30.0% MFU holds under the same accounting used for GPU reports, near-trillion-parameter MoE training is no longer tied to NVIDIA accelerators; other large vendors' stacks can host frontier-scale runs.
  • Simulation-based architecture search at 85-90% accuracy can substitute for much full-scale hyperparameter experimentation, so future model/hardware co-design can be done before hardware exists in quantity.
  • Dropless routing with device-level expert placement is preferable to token dropping at scale: with capacity factor 1.5, the 718B model dropped 8% of tokens versus 6% for the 20B model, meaning token dropping is increasingly harmful as models grow.
  • The EP-group auxiliary loss offers a middle ground between sequence-level and global-level balancing, yielding near-global-level loss with lower communication cost.
  • The reported system optimizations cumulatively raise MFU by 58.7% over baseline, implying similar headroom may exist on other under-optimized stacks for MoE training.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: because the paper never states the FLOPs formula behind MFU, a reader who re-derives efficiency from the stated 1.46M tokens per second, 718B parameters, and the Ascend peak compute could get a different number; a standardized formula would be needed for a fair cross-vendor comparison.
  • Editorial extension: the hierarchical all-to-all scheme and dynamic expert placement are developed for training, but the same mechanisms should carry over to MoE inference on the same hardware, where routing is still dynamic and memory is tighter.
  • Editorial extension: the simulator's 85-90% accuracy suggests the architecture-search workflow itself is portable; other accelerator vendors could run the same screening on their own compute-to-memory ratios, and the prediction that larger hidden sizes fit compute-heavy chips is directly testable.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 7 minor

Summary. This paper reports the design and training of Pangu Ultra MoE, a 718B-parameter Mixture-of-Experts language model with 61 layers, hidden size 7680, 256 routed experts, top-8 routing, a shared expert, MLA, and MTP, trained on 6K Ascend NPUs. The authors propose a simulation-based architecture search over approximately 10,000 configurations and select a model with 61 layers, hidden size 7680, 256 experts, using parallelism TP=8, PP=16, VPP=2, EP=4, MBS=2. On the systems side, they describe hierarchical EP All-to-All communication, adaptive pipe overlap, fine-grained recomputation, tensor swapping, and a planner/executor for dynamic expert placement. They report 30.0% MFU and 1.46M tokens/sec on 6K Ascend NPUs, a 58.7% relative MFU improvement over their baseline, and benchmark results across C-Eval, MMLU, AIME, GPQA, LiveCodeBench, and medical datasets that are said to be comparable to DeepSeek R1. They also include analyses of auxiliary-loss granularity, drop-and-pad versus dropless training, expert specialization, router scale, and expert co-activation.

Significance. If the reported efficiency and benchmark numbers hold, the paper is a significant engineering contribution: it provides one of the first public accounts of training a near-trillion-parameter MoE to near-frontier quality on a non-NVIDIA accelerator stack, and the system techniques (hierarchical EP all-to-all, adaptive pipe overlap, fine-grained recomputation, tensor swapping, dynamic placement) are described in enough detail to be useful to other systems groups. The authors are also transparent about simulation-accuracy limits and mark which benchmark scores were obtained by their own testing, which is more candid than many industrial reports. However, the paper does not ship code, checkpoints, training logs, or a full data recipe, and the central quantitative claims currently rest on undisclosed accounting choices: no MFU formula, no global batch size or sequence length, no evaluation protocol for asterisked benchmark scores, and an in-sample simulation validation. The methodology is sound in outline, but the headline numbers are not yet auditable.

major comments (4)
  1. [Section 1 / Section 5.1 / Table 5] The paper never states the FLOPs-per-token formula used to compute MFU. For the reported 718B MoE with 256 routed experts, top-8 activation, one shared expert, MLA, MTP, fine-grained recomputation, tensor swapping, and dynamic expert placement, the distinction between counting only activated routed-expert GEMMs and counting attention/shared-expert/MTP/recomputation FLOPs can shift the MFU by tens of percent. Section 5.1 does not report the global batch size, sequence length, total training tokens, or the peak FLOPs of the Ascend 910B used in the denominator, so the TPS value of 1.46M cannot be used to reconstruct or audit the 30.0% figure. In addition, Table 5's cumulative '58.7% relative increase' is presented as a baseline of 1, while the Section 1 baseline of 18.9% MFU is stated on 4K NPUs; because the cluster size and parallelism differ, the relative improvement is not a controlled comparison. The authors should disclose the exact MFU formula and all experimental inputs, or the central claim is not comparable to MFU numbers reported for GPU baselines.
  2. [Section 2.3 / Table 1] The simulation is validated on the final model it selected. The second row of Table 1 uses 'Pangu Ultra MoE Training' itself, giving 90.1% accuracy; this is an in-sample check and does not establish that the simulator correctly ranks the ~10,000 candidate configurations or the eight high-throughput models in Figure 4. The first row (4.2B, 128 NPUs) is out-of-sample but uses GQA, 4 experts, and a 6-layer model, so it does not exercise MLA, 256-expert routing, MTP, or the 6K-NPU communication patterns of the target configuration. The architecture-selection claim needs either a held-out validation on at least one additional 718B-class configuration or an explicit analysis of how simulation error varies across the candidate set, for example by comparing predicted versus measured performance of several selected candidates.
  3. [Section 5.2 / Table 6] The comparison to DeepSeek R1 is not fully auditable. Several DeepSeek numbers are asterisked as obtained by the authors' own testing, but the paper gives no evaluation protocol: prompt templates, decoding settings (temperature, top-p, sampling), answer extraction, number of runs or seeds, and contamination checks are all absent. Without these details and without run-to-run variability, differences such as AIME2024 81.3 versus 79.8 and LiveCodeBench 64.3 versus 65.9 cannot be distinguished from evaluation noise. The claim that Pangu Ultra MoE is comparable to DeepSeek R1 should be treated as provisional; the authors should provide a complete evaluation protocol or rely on official reported scores.
  4. [Section 4.4 / Table 5] The paper claims that dynamic expert placement gives a 'relative 10% improvement in MFU' and reduces device-level load imbalance by 80%-90%, but no ablation or measurement description is provided, and Table 5's cumulative MFU improvements do not list this item. If the Table 5 baseline is the same 18.9% MFU from Section 1, the arithmetic 18.9% x 1.587 = 30.0% already accounts for the full headline gain; adding a further 10% relative improvement would exceed the reported MFU unless it is already concealed inside one of Table 5's categories or the baselines differ. The authors should reconcile these numbers and provide an explicit ablation for the dynamic placement mechanism.
minor comments (7)
  1. [Page 1] The header 'Huawei Proprietary - Restricted Distribution' appears on a public arXiv submission; if the report is intended to be public, this marking should be removed or explained.
  2. [Section 2.4] There are typos: 'memomy access bandwidth' should be 'memory access bandwidth', and 'This prevents any bottleneck from happening' is ungrammatical and should be rephrased.
  3. [Section 4.1] The term 'Tiling = 4' in the CoC discussion is undefined; please explain what tiling denotes and how the 4x reduction in communication exposure is measured.
  4. [Figure 9 / Section 4.5] 'Umpermute' should be 'Unpermute' in the Figure 9 legend and in Section 4.5.
  5. [Section 5.2] The benchmark list in the opening paragraph omits C-SimpleQA, although it appears in Table 6; add it to the list.
  6. [Figure 12] The figure axis labels are rendered as '/uni...' escape sequences in the submitted text; the figure needs to be regenerated so expert IDs are legible.
  7. [References] References [58] and [59] duplicate the same Qwen2.5 technical report, and [51] and [52] duplicate the Megatron-LM citation; consolidate to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the post-hoc 718B simulator 're-validation' is in-sample but not load-bearing, and the undisclosed MFU formula is an auditability issue, not a circular derivation.

full rationale

The paper is an empirical systems report, not a derivation from first principles. The architecture-selection chain is: pilot MoE experiments on a 20B model choose expert granularity and shared-expert structure; a simulator built from operator-level hardware measurements and explicit hardware specifications searches roughly 10,000 configurations and selects Model 7; the chosen parallelism is then implemented and measured on 6K Ascend NPUs. The headline 30.0% MFU and 1.46M TPS are measured or engineering-derived quantities on the implemented system, not outputs produced by the simulator itself. The closest candidate for circularity is Section 2.3 and Table 1, where the simulator is 're-validated' on the 718B model after training: the paper explicitly says 'The end-to-end time in the second experiment with 718B experiments is acquired after actual Pangu Ultra MoE training on 6K NPUs, and we carry out simulations afterward to re-validate our simulation method.' This is a post-hoc consistency check, not an out-of-sample prediction, and the forward validation was done on the 4.2B pilot experiment before the architecture search. The 718B agreement therefore does not determine the architecture choice or the MFU claim by construction; it is in-sample evidence of simulator accuracy. The undisclosed FLOPs-per-token formula behind the 30.0% MFU and the asterisked self-run DeepSeek scores in Table 6 are auditability and evaluation-protocol concerns, not cases where a claimed result is equal to its inputs by definition. No load-bearing self-citation, imported uniqueness theorem, or renamed known result was found; the only same-author citation ([60], Pangu Ultra dense) supports an existing fused-operator implementation and is not load-bearing. The derivation chain is therefore self-contained: each quantitative claim is either measured, produced by a simulator with explicitly stated inputs, or validated against a separate small-scale experiment before use.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim relies on the Ascend NPU hardware model encoded in the simulator, the accuracy of that simulator (validated to 85-90% on two runs), and the assumption that benchmark comparisons and the undisclosed MFU flop-counting convention are standard and fair. The model hyperparameters and parallelism settings are selected by simulation and hand-tuning rather than derived from first principles.

free parameters (6)
  • Model architecture hyperparameters (61 layers, hidden 7680, 256 routed experts, expert intermediate 2048, top-8… = 61/7680/256/2048/8/1
    Selected by the simulator to maximize predicted training and inference throughput on Ascend NPUs; the pilot experiments (expert granularity, shared experts) only fix the MoE structure, not these final numbers.
  • Parallelism strategy (TP=8, PP=16, VPP=2, EP=4, MBS=2) = TP=8, PP=16, VPP=2, EP=4, MBS=2, Tiling=4
    Chosen by the same simulator as the best combination for the selected architecture; several other 5D parallelism combinations were evaluated but not reported in detail.
  • Auxiliary loss coefficient alpha = 1e-4 to 1e-2 tested; final value for Pangu Ultra MoE not stated
    The paper compares alpha=1e-2 and 1e-4 for the 20B baseline and says EP-group is a good compromise, but the final alpha used for the 718B model is not reported.
  • Expert capacity factor = 1.5
    Used in the drop-and-pad comparison experiments (Section 3.2); the capacity factor is a hand-set hyperparameter that directly affects the reported token drop rates.
  • Planner hyperparameters for dynamic expert placement = not specified
    Sliding window size, placement switch interval, and coefficient of variation threshold are not given, yet they determine the claimed 10% relative MFU improvement from load balancing.
  • Memory policy choices (MLA KV-only vs QKV recompute, permute recompute, prob swap) = KV-only recompute, permute recompute, activation recompute, prob swapping
    The combination is stated in Section 4.3 but the memory budgets and thresholds that trigger each policy are not reported, making the policy a hand-tuned configuration.
assumptions (4)
  • domain assumption The simulator's operator-level model of Ascend 910B compute, memory, and network behavior is accurate enough for comparative architecture search.
    Invoked in Section 2.2 and validated to 88.9% and 90.1% on two runs (Table 1). The in-sample validation on the final model limits independent support.
  • domain assumption MFU is computed with a standard FLOPs accounting convention that is comparable to DeepSeek's reported MFU.
    MFU 30.0% is the headline efficiency claim (Section 1, Section 4.5), but the exact FLOPs formula, sequence length, and batch size used in the MFU calculation are not disclosed.
  • domain assumption The evaluation protocol, including author-run benchmark results (marked with asterisks), is consistent across all compared models.
    Table 6 mixes publicly reported scores with the authors' own testing; if prompting or sampling settings differ, the 'comparable to DeepSeek R1' claim could be an artifact.
  • domain assumption The internal training data quality controls and the 3:1 reasoning to non-reasoning sample ratio are sufficient for the model to reach the reported capability.
    Data composition is described only at a high level in Section 5.1, yet the model's benchmark performance depends entirely on it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Pangu Ultra MoE: How to Train Your Big MoE on Ascend NPUs." pith.science (2026). https://pith.science/paper/D5BVIOYI

@misc{pith2026250504519,
  author       = {Pith},
  title        = {Pith review of: Pangu Ultra MoE: How to Train Your Big MoE on Ascend NPUs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D5BVIOYI}},
  note         = {Machine review of arXiv:2505.04519}
}
read the original abstract

Sparse large language models (LLMs) with Mixture of Experts (MoE) and close to a trillion parameters are dominating the realm of most capable language models. However, the massive model scale poses significant challenges for the underlying software and hardware systems. In this paper, we aim to uncover a recipe to harness such scale on Ascend NPUs. The key goals are better usage of the computing resources under the dynamic sparse model structures and materializing the expected performance gain on the actual hardware. To select model configurations suitable for Ascend NPUs without repeatedly running the expensive experiments, we leverage simulation to compare the trade-off of various model hyperparameters. This study led to Pangu Ultra MoE, a sparse LLM with 718 billion parameters, and we conducted experiments on the model to verify the simulation results. On the system side, we dig into Expert Parallelism to optimize the communication between NPU devices to reduce the synchronization overhead. We also optimize the memory efficiency within the devices to further reduce the parameter and activation management overhead. In the end, we achieve an MFU of 30.0% when training Pangu Ultra MoE, with performance comparable to that of DeepSeek R1, on 6K Ascend NPUs, and demonstrate that the Ascend system is capable of harnessing all the training stages of the state-of-the-art language models. Extensive experiments indicate that our recipe can lead to efficient training of large-scale sparse language models with MoE. We also study the behaviors of such models for future reference.

Figures

Figures reproduced from arXiv: 2505.04519 by the authors.

Figure 1
Figure 1. Experts number ablation. Since increasing [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. Simulation workflow for LLM performance and optimal search. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Simulated model normalized throughput of training and inference. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figures from the paper (10 more)
Figure 5
Figure 5. Figure 5: Comparison of the load balancing auxiliary loss and cross-entropy loss during training across [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Training loss curves for the 20B MoE models trained with drop-and-pad and dropless methods. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Analysis of token drop rates. During drop-and-pad training with the expert capacity factor set to [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Virtual pipeline stage partitioning matrix (2-row VPP Stages [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: MoE operator streams. Subfigure (a) shows the forward pass. Subfigure (b) shows the backward [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Comparison between before host_bound optimization and after. Compared with permute1, GMM [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]
Figure 11
Figure 11. Figure 11: Overview of dynamic device-level load balancing mechanism. The planner generates and periodi [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Expert specialization pattern of Pangu Ultra MoE. Each subgraph represents the token distribution [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: The outputs of routed and shared experts. We display the outputs after the 30th layer for better [PITH_FULL_IMAGE:figures/full_fig_p019_13.png]
Figure 14
Figure 14. Figure 14: Co-activation among experts in nine layers on a random 0.5% of the C4 validation data. We display [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MoX: Efficient MoE Routing on Direct-Connect Topologies

    cs.NI 2026-07 conditional novelty 6.0 of 10

    Static, demand-oblivious routing with token-aware multicast trees and precomputed per-link weights brings MoE traffic on direct-connect fabrics close to ideal switch performance.

  2. DTop-p MoE: Sparsity-Controlled Dynamic Top-p MoE for Foundation Model Pre-training

    cs.AI 2025-12 conditional novelty 6.0 of 10

    A PI-controller-tuned, layerwise-normalized Top-p router trains sparse MoE models that beat Top-k at matched average activated-expert count.

  3. Serving Large Language Models on Huawei CloudMatrix384

    cs.DC 2025-06 conditional novelty 6.0 of 10

    CloudMatrix384 with CloudMatrix-Infer reports prefill 6,688 tokens/s per NPU and decode 1,943 tokens/s per NPU on DeepSeek-R1, exceeding published NVIDIA-based efficiencies.

Reference graph

Works this paper leans on

92 extracted references · 41 canonical work pages · cited by 3 Pith papers

  1. [1]

    https://gitee.com/ascend/MindSpeed-LLM/blob/ master/docs/features/communication-over-computation.md

    CoC (Communication Over Computation). https://gitee.com/ascend/MindSpeed-LLM/blob/ master/docs/features/communication-over-computation.md

  2. [2]

    https://github.com/NVIDIA/Megatron-LM

    Megatron-LM. https://github.com/NVIDIA/Megatron-LM

  3. [3]

    https://gitee.com/ascend/MindSpeed

    MindSpeed. https://gitee.com/ascend/MindSpeed

  4. [4]

    The llama 4 herd: The beginning of a new era of natively multimodal ai innovation, 2025

    Meta AI. The llama 4 herd: The beginning of a new era of natively multimodal ai innovation, 2025

    2025

  5. [5]

    Gqa: Training generalized multi-query transformer models from multi-head checkpoints, 2023

    Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints, 2023

    2023

  6. [6]

    Program synthesis with large language models

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, et al. Program synthesis with large language models. arXiv preprint arXiv:2108.07732, 2021

    arXiv 2021

  7. [7]

    Training deep nets with sublinear memory cost, 2016

    Tianqi Chen, Bing Xu, Chiyuan Zhang, and Carlos Guestrin. Training deep nets with sublinear memory cost, 2016. arXiv preprint arXiv:1604.06174, 2016

    arXiv 2016

  8. [8]

    Prediction is all moe needs: Expert load distribution goes from fluctuating to stabilizing.arXiv preprint arXiv:2404.16914, 2024

    Peizhuang Cong, Aomufei Yuan, Shimao Chen, Yuxuan Tian, Bowen Ye, and Tong Yang. Prediction is all moe needs: Expert load distribution goes from fluctuating to stabilizing.arXiv preprint arXiv:2404.16914, 2024

    arXiv 2024

Show all 92 references
  1. [9]

    Damai Dai, Chengqi Deng, Chenggang Zhao, R. X. Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Y . Wu, Zhenda Xie, Y . K. Li, Panpan Huang, Fuli Luo, Chong Ruan, Zhifang Sui, and Wenfeng Liang. Deepseekmoe: Towards ultimate expert specialization in mixture-of-...

    2024 arXiv

  2. [10]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025

    DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, 2025

    2025

  3. [11]

    Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity

    William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1–39, 2022

    2022

  4. [12]

    Megablocks: Efficient sparse training with mixture-of-experts

    Trevor Gale, Deepak Narayanan, Cliff Young, and Matei Zaharia. Megablocks: Efficient sparse training with mixture-of-experts. Proceedings of Machine Learning and Systems, 5:288–304, 2023

    2023

  5. [13]

    Measuring massive multitask language understanding

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020

    2009 arXiv

  6. [14]

    Swapadvisor: Pushing deep learning beyond the gpu memory limit via smart swapping

    Chien-Chin Huang, Gu Jin, and Jinyang Li. Swapadvisor: Pushing deep learning beyond the gpu memory limit via smart swapping. In Proceedings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Operating Systems, pages 1341–1355, 2020. 20

    2020

  7. [15]

    Harder tasks need more experts: Dynamic routing in moe models

    Quzhe Huang, Zhenwei An, Nan Zhuang, Mingxu Tao, Chen Zhang, Yang Jin, Kun Xu, Liwei Chen, Songfang Huang, and Yansong Feng. Harder tasks need more experts: Dynamic routing in moe models. arXiv preprint arXiv:2403.07652, 2024

    2024

  8. [16]

    Le, Yonghui Wu, and Zhifeng Chen

    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. Gpipe: Efficient training of giant neural networks using pipeline parallelism, 2019

    2019

  9. [17]

    C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models

    Yuzhen Huang, Yuzhuo Bai, Zhihao Zhu, Junlei Zhang, Jinghan Zhang, Tangjun Su, Junteng Liu, Chuancheng Lv, Yikai Zhang, Yao Fu, et al. C-eval: A multi-level multi-discipline chinese evaluation suite for foundation models. Advances in Neural Information Processing Systems, 36:6...

    2023

  10. [18]

    Calculon: a methodology and tool for high-level co-design of systems and large language models

    Mikhail Isaev, Nic Mcdonald, Larry Dennison, and Richard Vuduc. Calculon: a methodology and tool for high-level co-design of systems and large language models. In Proceedings of the International Conference for High Performance Computing, Networking, Storage and Analysis, SC ’...

    2023

  11. [19]

    Deepspeed ulysses: System optimizations for enabling training of extreme long sequence transformer models, 2023

    Sam Ade Jacobs, Masahiro Tanaka, Chengming Zhang, Minjia Zhang, Shuaiwen Leon Song, Samyam Rajbhandari, and Yuxiong He. Deepspeed ulysses: System optimizations for enabling training of extreme long sequence transformer models, 2023

    2023

  12. [20]

    Livecodebench: Holistic and contamination free evaluation of large language models for code

    Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Armando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. arXiv preprint arXiv:2403.07974, 2024

    2024 arXiv

  13. [21]

    Mixtral of experts

    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, et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024

    2024 arXiv

  14. [22]

    What disease does this patient have? a large-scale open domain question answering dataset from medical exams

    Di Jin, Eileen Pan, Nassim Oufattole, Wei-Hung Weng, Hanyi Fang, and Peter Szolovits. What disease does this patient have? a large-scale open domain question answering dataset from medical exams. Applied Sciences, 11(14):6421, 2021

    2021

  15. [23]

    Scaling laws for neural language models

    Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models. arXiv preprint arXiv:2001.08361, 2020

    2001 arXiv

  16. [24]

    Reducing activation recomputation in large transformer models, 2022

    Vijay Korthikanti, Jared Casper, Sangkug Lym, Lawrence McAfee, Michael Andersch, Mohammad Shoeybi, and Bryan Catanzaro. Reducing activation recomputation in large transformer models, 2022

    2022

  17. [25]

    Le, Haruki Imai, Yasushi Negishi, and Kiyokuni Kawachiya

    Tung D. Le, Haruki Imai, Yasushi Negishi, and Kiyokuni Kawachiya. Tflms: Large model support in tensorflow by graph rewriting, 2019

    2019

  18. [26]

    Gshard: Scaling giant models with conditional computation and automatic sharding, 2020

    Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding, 2020

    2020

  19. [27]

    The depth-to-width interplay in self-attention

    Yoav Levine, Noam Wies, Or Sharir, Hofit Bata, and Amnon Shashua. The depth-to-width interplay in self-attention. arXiv preprint arXiv:2006.12467, 2020

    2006 arXiv

  20. [28]

    Minimax-01: Scaling foundation models with lightning attention

    Aonian Li, Bangwei Gong, Bo Yang, Boji Shan, Chang Liu, Cheng Zhu, Chunhao Zhang, Congchao Guo, Da Chen, Dong Li, et al. Minimax-01: Scaling foundation models with lightning attention. arXiv preprint arXiv:2501.08313, 2025

    2025 arXiv

  21. [29]

    Locmoe: A low-overhead moe for large language model training

    Jing Li, Zhijie Sun, Xuan He, Li Zeng, Yi Lin, Entong Li, Binfan Zheng, Rongqian Zhao, and Xin Chen. Locmoe: A low-overhead moe for large language model training. arXiv preprint arXiv:2401.13920, 2024

    2024 arXiv

  22. [30]

    Davinci: A scalable architecture for neural network computing

    Heng Liao, Jiajin Tu, Jing Xia, and Xiping Zhou. Davinci: A scalable architecture for neural network computing. In 2019 IEEE Hot Chips 31 Symposium (HCS), pages 1–44, 2019

    2019

  23. [31]

    Let’s verify step by step

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. In The Twelfth International Conference on Learning Representations, 2023

    2023

  24. [32]

    Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model

    Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai, Daya Guo, et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model. arXiv preprint arXiv:2405.04434, 2024. 21

    2024 arXiv

  25. [33]

    Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    2024 arXiv

  26. [34]

    Grin: Gradient-informed moe

    Liyuan Liu, Young Jin Kim, Shuohang Wang, Chen Liang, Yelong Shen, Hao Cheng, Xiaodong Liu, Masahiro Tanaka, Xiaoxia Wu, Wenxiang Hu, et al. Grin: Gradient-informed moe. arXiv preprint arXiv:2409.12136, 2024

    2024 arXiv

  27. [35]

    Codeforces

    MAA. Codeforces. American Invitational Mathematics Examination - AIME 2024, 2024. https:// maa.org/math-competitions/american-invitational-mathematics-examination-aime

    2024

  28. [36]

    Codeforces

    MAA. Codeforces. American Invitational Mathematics Examination - AIME 2025, 2025. https:// maa.org/math-competitions/american-invitational-mathematics-examination-aime

    2025

  29. [37]

    Training deeper models by gpu memory optimization on tensorflow

    Chen Meng, Minmin Sun, Jun Yang, Minghui Qiu, and Yang Gu. Training deeper models by gpu memory optimization on tensorflow. In Proc. of ML Systems Workshop in NIPS, volume 7, page 26, 2017

    2017

  30. [38]

    Mgpu-tsm: A multi-gpu system with truly shared memory

    Saiful A Mojumder, Yifan Sun, Leila Delshadtehrani, Yenai Ma, Trinayan Baruah, José L Abellán, John Kim, David Kaeli, and Ajay Joshi. Mgpu-tsm: A multi-gpu system with truly shared memory. arXiv preprint arXiv:2008.02300, 2020

    2008 arXiv

  31. [39]

    Olmoe: Open mixture-of-experts language models

    Niklas Muennighoff, Luca Soldaini, Dirk Groeneveld, Kyle Lo, Jacob Morrison, Sewon Min, Weijia Shi, Pete Walsh, Oyvind Tafjord, Nathan Lambert, et al. Olmoe: Open mixture-of-experts language models. arXiv preprint arXiv:2409.02060, 2024

    2024 arXiv

  32. [40]

    Efficient large-scale language model training on gpu clusters using megatron-lm

    Deepak Narayanan, Mohammad Shoeybi, Jared Casper, Patrick LeGresley, Mostofa Patwary, Vijay Korthikanti, Dmitri Vainbrand, Prethvi Kashinkunti, Julie Bernauer, Bryan Catanzaro, et al. Efficient large-scale language model training on gpu clusters using megatron-lm. In Proceedin...

    2021

  33. [41]

    Efficient large-scale language model training on gpu clusters using megatron-lm, 2021

    Deepak Narayanan, Mohammad Shoeybi, Jared Casper, Patrick LeGresley, Mostofa Patwary, Vijay Anand Korthikanti, Dmitri Vainbrand, Prethvi Kashinkunti, Julie Bernauer, Bryan Catanzaro, Amar Phanishayee, and Matei Zaharia. Efficient large-scale language model training on gpu clus...

    2021

  34. [42]

    Flexmoe: Scaling large-scale sparse pre-trained model training via dynamic device placement

    Xiaonan Nie, Xupeng Miao, Zilong Wang, Zichao Yang, Jilong Xue, Lingxiao Ma, Gang Cao, and Bin Cui. Flexmoe: Scaling large-scale sparse pre-trained model training via dynamic device placement. Proceedings of the ACM on Management of Data, 1(1):1–19, 2023

    2023

  35. [43]

    Medmcqa: A large-scale multi- subject multi-choice dataset for medical domain question answering

    Ankit Pal, Logesh Kumar Umapathi, and Malaikannan Sankarasubbu. Medmcqa: A large-scale multi- subject multi-choice dataset for medical domain question answering. In Conference on health, inference, and learning, pages 248–260. PMLR, 2022

    2022

  36. [44]

    Capuchin: Tensor-based gpu memory management for deep learning

    Xuan Peng, Xuanhua Shi, Hulin Dai, Hai Jin, Weiliang Ma, Qian Xiong, Fan Yang, and Xuehai Qian. Capuchin: Tensor-based gpu memory management for deep learning. In Proceedings of the Twenty-Fifth International Conference on Architectural Support for Programming Languages and Op...

    2020

  37. [45]

    Demons in the detail: On implementing load balancing loss for training specialized mixture-of-expert models

    Zihan Qiu, Zeyu Huang, Bo Zheng, Kaiyue Wen, Zekun Wang, Rui Men, Ivan Titov, Dayiheng Liu, Jingren Zhou, and Junyang Lin. Demons in the detail: On implementing load balancing loss for training specialized mixture-of-expert models. arXiv preprint arXiv:2501.11873, 2025

    2025 arXiv

  38. [46]

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. arXiv e-prints, 2019

    2019

  39. [47]

    Zero: Memory optimizations toward training trillion parameter models, 2020

    Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models, 2020

    2020

  40. [48]

    Gpqa: A graduate-level google-proof q&a benchmark

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R Bowman. Gpqa: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling, 2024

    2024

  41. [49]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. arXiv preprint arXiv:1701.06538, 2017

    2017 arXiv

  42. [50]

    Megatron-lm: Training multi-billion parameter language models using model parallelism

    Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. Megatron-lm: Training multi-billion parameter language models using model parallelism. arXiv preprint arXiv:1909.08053, 2019. 22

    1909 arXiv

  43. [52]

    Megatron-lm: Training multi-billion parameter language models using model parallelism, 2020

    Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. Megatron-lm: Training multi-billion parameter language models using model parallelism, 2020

    2020

  44. [53]

    Hunyuan-large: An open-source moe model with 52 billion activated parameters by tencent

    Xingwu Sun, Yanfeng Chen, Yiqing Huang, Ruobing Xie, Jiaqi Zhu, Kai Zhang, Shuaipeng Li, Zhen Yang, Jonny Han, Xiaobo Shu, et al. Hunyuan-large: An open-source moe model with 52 billion activated parameters by tencent. arXiv preprint arXiv:2411.02265, 2024

    2024 arXiv

  45. [54]

    Qwen1.5-moe: Matching 7b model performance with 1/3 activated parameters", February 2024

    Qwen Team. Qwen1.5-moe: Matching 7b model performance with 1/3 activated parameters", February 2024

    2024

  46. [55]

    Mmlu-pro: A more robust and challenging multi-task language understanding benchmark

    Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. In The Thirty-eight Conference on Neural Information Pr...

    2024

  47. [56]

    Skywork-moe: A deep dive into training techniques for mixture-of- experts language models

    Tianwen Wei, Bo Zhu, Liang Zhao, Cheng Cheng, Biye Li, Weiwei Lü, Peng Cheng, Jianhao Zhang, Xiaoyu Zhang, Liang Zeng, et al. Skywork-moe: A deep dive into training techniques for mixture-of- experts language models. arXiv preprint arXiv:2406.06563, 2024

    2024 arXiv

  48. [57]

    Clue: A chinese language understanding evaluation benchmark

    Liang Xu, Hai Hu, Xuanwei Zhang, Lu Li, Chenjie Cao, Yudong Li, Yechen Xu, Kai Sun, Dian Yu, Cong Yu, et al. Clue: A chinese language understanding evaluation benchmark. arXiv preprint arXiv:2004.05986, 2020

    2004 arXiv

  49. [59]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024

    2024 arXiv

  50. [60]

    Pangu ultra: Pushing the limits of dense large language models on ascend npus

    Yichun Yin, Wenyong Huang, Kaikai Song, Yehui Tang, Xueyu Wu, Wei Guo, Peng Guo, Yaoyuan Wang, Xiaojun Meng, Yasheng Wang, et al. Pangu ultra: Pushing the limits of dense large language models on ascend npus. arXiv preprint arXiv:2504.07866, 2025

    2025 arXiv

  51. [61]

    SmartMoE: Efficiently training sparsely-activated models through combining offline and online oarallelization

    Mingshu Zhai, Jiaao He, Zixuan Ma, Zan Zong, Runqing Zhang, and Jidong Zhai. SmartMoE: Efficiently training sparsely-activated models through combining offline and online oarallelization. In 2023 USENIX Annual Technical Conference (USENIX ATC 23), pages 961–975, Boston, MA, Ju...

    2023

  52. [62]

    {SmartMoE}: Efficiently training{Sparsely-Activated} models through combining offline and online parallelization

    Mingshu Zhai, Jiaao He, Zixuan Ma, Zan Zong, Runqing Zhang, and Jidong Zhai. {SmartMoE}: Efficiently training{Sparsely-Activated} models through combining offline and online parallelization. In 2023 USENIX Annual Technical Conference (USENIX ATC 23), pages 961–975, 2023

    2023

  53. [63]

    客户信用风险以外的风险

    Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. Instruction-following evaluation for large language models. arXiv preprint arXiv:2311.07911, 2023. 23 A Contributions and Acknowledgments Core Contributors Yehui Tang, Yi...

    2023 arXiv

  54. [64]

    **明确岗位需求** - 针对人大学生的专业优势（如法学、经济学、公共管理、新闻传播等）设计岗位，明确实习内容、技能 要求、工作时间和薪资待遇。-突出企业品牌和实习生的成长机会（如培训体系、转正通道）。

  55. [65]

    **联系校方资源** -提前与人 大就业指导中心或相关院系对接，确认校招流程、场地 安排和时间节点。-通过校内公众号、 院系公告栏、学生社群等渠道推广招聘信息。

  56. [66]

    企业代言人”参与宣讲。 **二、招聘流程设计** 1. ** 宣讲会或双选会** - 组织线 下/线上宣讲会，介绍企业文化、业务方向、实习生培养计划，并预留 互动问答时间。-可设置“快闪面试

    ** 准备宣传材料** -制作简洁清晰的招聘海报、宣传单页、公司介绍PPT等，突出企业核心竞争力和实习生福利。- 若有校 友资源，可邀请人大毕业的员工作为“企业代言人”参与宣讲。 **二、招聘流程设计** 1. ** 宣讲会或双选会** - 组织线 下/线上宣讲会，介绍企业文化、业务方向、实习生培养计划，并预留 互动问答时间。-可设置“快闪面试”环节，现场接收简历并初步筛选。

  57. [67]

    **简历筛选与面试** - 人大学生普遍综合素质较高，需注重简历中的学术背景、实践经历和技能匹配度。- 面试时侧重考察 逻辑思维、学习能力、抗压能力，避免过于强调经验（实习生更看重潜力）。

  58. [68]

    ** 灵活考核方式** -可设置案例分析、小组讨论等环节，观察候选人的团队协作和问题解决能力。- 对技术类岗位可增加笔 试或实操测试。 **三、吸引人大学子的关键点**

  59. [69]

    **职业发展价值** -强调实习与 未来职业发展的关联性，例如参与核心项目、接触行业资源、获得导师指导等。-提供清晰 的晋升路径（如转正机会或推荐信）。

  60. [70]

    **匹配学生特点** - 人大学生普遍关注社会价值，可突出企业社会责任、行业影响力或岗位的公共 政策相关性。- 灵活的工 作时间（如兼顾学业）、交通/餐补等细节需明确说明。

  61. [71]

    ** 薪资与福利竞争力** -提供符合市场水平的实习薪资（北京地区需考虑生活成本），或通过非现金福利（如免费培训、活 动参与机会）吸引学生。 **四、注意事项**

  62. [72]

    **遵守法律法规** -签订正规实习协议，明确工作内容、薪资、保密条款等，避免劳动 纠纷 。- 工作时间需符合《劳动 法》 规定，保护实习生权益。

  63. [73]

    **提升候选人体验** -及时反馈面试结果，即使拒绝也保持礼貌，维护企业形象。- 入职前做好岗前培训，帮助实习生快速 融入团队。

  64. [74]

    法律实务研讨会”，吸引法学实 习生，现场提供模拟法庭案例分析，增强互动。- **企 业B**：联合人大就业中心推出“实习+导师计划

    ** 长期关系维护** - 建立实习生人才库，对未录用的优秀学生保持联系，未来有 岗位时优先推荐。-通过校友网络或企业活 动持续扩大在校园的影响力。 **五、案例参考** - **企业A**：在人大法学院举办“法律实务研讨会”，吸引法学实 习生，现场提供模拟法庭案例分析，增强互动。- **企 业B**：联合人大就业中心推出“实习+导师计划”，由高管一对一指导，提升学生参与意愿。 通过以上策略，既能高效匹配优质实习生，也能在人大校园中树立良好的雇主品牌，为未来校招积累资源。 Pangu Ultra MoE : 在中国人民大学（人大）招聘实习生时，为确...

  65. [75]

    ** 学科侧重**：人大以人文社科见长，优势学科包括法学、经济学、管理学、新闻传播、社会学、公共管理等，理工科相对 较少。确保岗位与这些领域匹配，或明确交叉学科需求（如数据科学在经济/金融中的应用）。

  66. [76]

    ** 学生特质**：人大生源综合素质高，注重理论与实践结合，多数学生对政策研究、智库咨询、金融、法律等领域兴趣浓 厚。 **二、岗位设计与宣传策略**

  67. [77]

    打杂”岗位，突出专业性和成长性（如参与政策分析、课题研究、行业调研等）。- **命名**：避免笼统的“实习生

    ** 岗位吸引力**：- ** 内容**：避免“打杂”岗位，突出专业性和成长性（如参与政策分析、课题研究、行业调研等）。- **命名**：避免笼统的“实习生”，可细化为“金融分析实习生（宏观经济方向）”“ 公共 政策研究助理”等。- ** 转正机 会**：明确是否有留用可能，人大高年级学生（如研二、大三）更关注转正潜力。

  68. [78]

    ** 宣传渠道**：- ** 官方合作**：通过人大就业指导中心官网（http://career.ruc.edu.cn ）发布信息，申请线下宣讲会或双选 会。 - **校内渗透**：联系相关学院（如经济学院、法学院、商学院）的就业辅导员，利用院系微信群、公众号精准推送。- **学生社群**：与人大BBS（天地人大）、学生社团（如职业发展协会）合作，或通过校友内推。 — **三、招聘流程优化**

  69. [79]

    ** 时间安排**：- **避开考试周**：学期末 （6月、12月）学生专注考试，建议在学期中段（4-5月、9-10月）启动招聘。 - **反馈效率**：人大生通常投递多家企业，需缩短流程周期（如1周内反馈初筛结果）。

  70. [80]

    **考核方式**：- **笔试**：若需考察专业能力，可结合案例分析（如经济类岗位提供政策分析题）。- **面试**：建议安 排业务部门负责人参与，突出专业深度，避免泛泛而谈。 **四、待遇与法律合规**

  71. [81]

    ** 实习待遇**：- **薪资**：人大位于北京，建议日薪不低于150元（科研助理类岗位可适当调整）。 - **福利**：提供餐 补、交通补贴或远程办公灵活性，优秀者可承诺推荐信或实习奖励。

  72. [82]

    **协议规范**：-签订《实习协议》，明确工作内容、时长、保密条款及意外保险。-避免要求“强制坐班”，尊重学生课 表时间（如每周3天到岗）。 **五、品牌形象与长期合作**

  73. [83]

    ** 宣讲会设计**：-邀请人大校友分享职业经历，增强学生认同感。-提供行业前沿洞察（如政策解读、金融趋势），吸引 学术型学生。

  74. [84]

    ** 校企合作**：- 与学院共建实习基地 ，赞助学术竞赛或课题研究。- 定期开放企业参访，邀请学生参与短期项目（如周末 工作坊）。 **六、后续跟进** - **录用沟通**：明确入职培训计划，发送欢迎邮件并附上团队介绍。 - **反馈收集**：实习结束后向学生和导师征求意见，优化未来招聘策略。 **总结**：人大招聘需注重岗位与学科的匹配度、流程高效性及学生职业发展诉求。通过精准宣传、专业考核和长期合作，可 建立优质人才输送渠道。建议提前2-3个月联系校方，预留充足时间协调资源。 29 Table 13: This is an AIGC cas...

  75. [85]

    ** 应用场景不同**：- **DFT** ：在设计阶段嵌入测试结构（如扫描链、边界扫描JTAG），主要用于芯片制造完成后的**生 产测试**（检测制造缺陷）。 - **SSI** ：在芯片部署到系统后，通过系统级接口（如高速总线、专用调试接口）进行**运行时的功能验证、性能监控和故障 诊断**，属于**运维阶段的测试**。

  76. [86]

    ** 实现方式差异**： - **DFT** ：依赖芯片内部预置的测试逻辑（如扫描触发器、BIST引擎），需要占用芯片面积和功耗。 - **SSI** ：利用系统已有的通信接口（如PCIe、以太网或专用调试接口），通过协议解析和命令交互 实现测试，通常无需额外 硬件开销。

  77. [87]

    ** 测试目标**： - **DFT** ：主要检测制造相关的物理缺陷（如开路、短路、延迟故障）。 - **SSI** ：针对系统级问题，如信号完整性（SI）、电源噪声、软错误（如单粒子翻转）以及软件与硬件的交互故障。

  78. [88]

    ** 灵活性与动态性**： - **DFT** ：测试模式固定，需通过ATE（自动测试设备）执行，测试向量在制造阶段固化。 - **SSI** ：支持动态配置测试策略，可在系统运行过程中实时调整测试参数，甚至结合软件驱动实现自适应测试。 典型应用场景： - **复杂SoC的运维监测**：通过SSI实时监控芯片温度、电压、关键信号路径的完整性。 - **车载/航天电子**：在恶劣环境中检测瞬时故障并触发自修复机制。 - **数据中心芯片**：周期性扫描系统状态，预防因老化或软错误导致的宕机。 总结：SSI是DFT的补充而非替代，两者结合可覆盖芯片全生...

  79. [89]

    ** 核心思想** SSI是在系统级（如芯片集成到终端设备后）通过软件驱动或固件实现的测试模式，利用系统自身资源（如处 理器、内存、接口等）执行自测试功能，通常用于运行时故障检测或周期性健康监测。

  80. [90]

    ** 特点** - **动态测试**：在系统运行过程中实时或按需触发测试，支持在线故障诊断（如老化监测、功能异常检测）。- **软件驱动**：依赖软件算法或固件指令生成测试用例，无需额外硬件电路（或仅需少量硬件配合）。- **系统级 覆盖**：可 测试芯片与其他系统组 件（如传感器、电源管理模块）的交互 问题，而不仅是芯片内部结构。- **灵活性高**：测试逻辑可通 过软件更新，适配不同应用场景。

  81. [91]

    ** 典型应用** - 汽车电子中的功能安全（ISO 26262 ）要求系统实时自检。- 数据中心芯片的运行时健康监测（如温度、电压 异常检测）。- 物联网设备的低功耗周期性自测试。 **DFT （Design for Test ）**

  82. [92]

    ** 核心思想** DFT是在芯片设计 阶段插入硬件结构（如扫描链、BIST电路）， 以提升制造缺陷的检测效率，主要用于芯片 量产后的**静态测试**（如ATE测试）。

  83. [93]

    ** 特点** - ** 硬件依赖**：需额外设计测试电路（如扫描触发器、MBIST控制器），占用芯片面积和功耗。- **制造阶段测 试**：主要用于芯片出厂前的缺陷筛查（如stuck-at 故障、延迟故障）。- **结构化测试**：针对预定义的故障模型生成测试向 量，覆盖晶体管级或门级缺陷。- **高覆盖率**：通过自动化工具（如ATPG）生成测试模式，确保缺陷检测率。

  84. [94]

    ** 典型应用** - 芯片量产时的自动化测试（ATE）。- 扫描链测试（Scan Test）、内存BIST（Built-In Self-Test ）。- 边界扫描 （Boundary Scan ）用于PCB互联测试。 **SSI 与DFT的关键区别** | **维度** | **SSI** | **DFT** | | **测试阶段** | 系统运行时（动态测试）| 芯片制造后（静态测试）| | **实现方式** | 软件/固件驱动，利用系统资源| 硬件结构设计（扫描链、BIST等）| | **测试目标** |功能异常、环境干扰、老化问题|制造缺陷...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.