The DOI matches its Crossref record; 13 of the 13 references it cites were found in a registry or an index.
AI and Memory Wall
Open accessAbstract
The availability of unprecedented unsupervised training data, along with neural scaling laws, has resulted in an unprecedented surge in model size and compute requirements for serving/training LLMs. However, the main performance bottleneck is increasingly shifting to memory bandwidth. Over the past 20 years, peak server hardware FLOPS has been scaling at 3.0×/2yrs, outpacing the growth of DRAM and interconnect bandwidth, which have only scaled at 1.6 and 1.4 times every 2 years, respectively. This disparity has made memory, rather than compute, the primary bottleneck in AI applications, particularly in serving. Here, we analyze encoder and decoder Transformer models and show how memory bandwidth can become the dominant bottleneck for decoder models. We argue for a redesign in model architecture, training, and deployment strategies to overcome this memory limitation.
Overview
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“These observations are consistent with Gholami et al.'s (2024) findings on AI's restricted memory capacity and its tendency to produce narrative incoherence in longer texts.”
From How Creative Is AI Writing? Generative and Collaborative AI in Japanese Fiction (2026) · cites this paper as Gholami et al.'s (2024)
Experimental — accuracy not yet measuredThe check wasn’t confident either way, so it gives no verdict — never disputed.
In the citing paper
“the relative weakness of zero-shot generation is consistent with research on AI's limitations in sustaining narrative coherence (Gholami et al., 2024)”
From How Creative Is AI Writing? Generative and Collaborative AI in Japanese Fiction (2026) · cites this paper as Gholami et al., 2024
Experimental — accuracy not yet measuredThe check wasn’t confident either way, so it gives no verdict — never disputed.
In the citing paper
“Gholami et al. (2024) identify restricted memory capacity as one of the primary challenges of AI-assisted fiction writing: as a story progresses, the model tends to lose track of earlier plot points and character details, which result in narrative inconsistencies.”
From How Creative Is AI Writing? Generative and Collaborative AI in Japanese Fiction (2026) · cites this paper as Gholami et al. (2024)
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173 · cited 3×
“We examine two different variations of the Trans-former architecture: the encoder architecture (e.g., BERT [7]), which concurrently processes all tokens, and the decoder architecture (e.g., GPT [3, 33]), which runs auto-regressively to process and generate one token at each iteration.”
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325 · cited 3×
“We examine two different variations of the Trans-former architecture: the encoder architecture (e.g., BERT [7]), which concurrently processes all tokens, and the decoder architecture (e.g., GPT [3, 33]), which runs auto-regressively to process and generate one token at each iteration.”
Language Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372 · cited 4×
“Deploying recent SOTA models [4, 5, 16, 37] such as GPT-3 [3] or large recommendation systems [29] is quite challenging, as they require distributed-memory deployment for inference.”
InfluentialWilliams, S., Waterman, A., & Patterson, D. A. (2009) · Communications of the ACM · doi:10.1145/1498765.1498785 · Cited by 2,798 · cited 1×
“This is due to the higher memory operations and lower arithmetic intensity of matrix-vector operations inherent in the auto-regressive inference of GPT (see Figure 3c).”
Wulf, W. A., & McKee, S. A. (1995) · ACM SIGARCH Computer Architecture News · doi:10.1145/216585.216588 · Cited by 1,625 · cited 1×
“They predicted that the diverging speed of improvement of how fast computations can be performed versus how fast data can be fetched is going to create a “memory wall” issue [25, 42].”
A Survey of Quantization Methods for Efficient Neural Network Inference
Gholami, A., Kim, S., Dong, Z., Yao, Z., Mahoney, M. W., & Keutzer, K. (2022) · Low-Power Computer Vision · doi:10.1201/9781003162810-13 · Cited by 1,219
Self-citationTraining Compute-Optimal Large Language Models
Hoffmann, J., Borgeaud, S., Mensch, A., Buchatskaya, E., Cai, T., Rutherford, E., Casas, D. D. L., Hendricks, L. A., Welbl, J., Clark, A., et al. (2022) · Advances in Neural Information Processing Systems 35 · doi:10.52202/068431-2176 · Cited by 932 · cited 1×
“In this work, we re-examined this trend by studying more recent data, with a particular focus on hardware used to train AI models, as well as the characteristics of the computations used for training/serving them.”
Method(2020) · doi:10.1109/sc41405.2020.00024 · Cited by 928 · cited 1×
“A promising line of work for that is the Zero framework from Microsoft, which showed how one can train 8 × bigger models with the same memory capacity by removing/sharding redundant optimization state variables [2, 34].”
Hoefler, T., Alistarh, D., Ben-Nun, T., Dryden, N., & Peste, A. (2021) · ETH Zurich · doi:10.3929/ethz-b-000519202 · Cited by 423
ADAHESSIAN: An Adaptive Second Order Optimizer for Machine Learning
Yao, Z., Gholami, A., Shen, S., Mustafa, M., Keutzer, K., & Mahoney, M. W. (2021) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v35i12.17275 · Cited by 257
Self-citation
Cited by
Sorted by citations
Efficient Memory Management for Large Language Model Serving with PagedAttention
Kwon, W., Li, Z., Zhuang, S., Sheng, Y., Zheng, L., Yu, C. H., Gonzalez, J. E., Zhang, H., & Stoica, I. (2023) · arXiv · doi:10.48550/arxiv.2309.06180 · Cited by 1,802
Efficient Memory Management for Large Language Model Serving with PagedAttention
Kwon, W., Li, Z., Zhuang, S., Sheng, Y., Zheng, L., Yu, C. H., Gonzalez, J., Zhang, H., & Stoica, I. (2023) · Proceedings of the 29th Symposium on Operating Systems Principles · doi:10.1145/3600006.3613165 · Cited by 1,817
(2021) · doi:10.1145/3458817.3476205 · Cited by 257
Towards sustainable AI: a comprehensive framework for Green AI
Tabbakh, A., Amin, A., Islam, M., Mahmud, G., Chowdhury, I. K., & Mukta, S. H. (2024) · Discover Sustainability · doi:10.1007/s43621-024-00641-4 · Cited by 168
ANT: Exploiting Adaptive Numerical Data Type for Low-bit Deep Neural Network Quantization
Guo, C., Zhang, C., Leng, J., Liu, Z., Yang, F., Liu, Y., Guo, M., & Zhu, Y. (2022) · IEEE/ACM International Symposium on Microarchitecture (MICRO) · doi:10.1109/micro56248.2022.00095 · Cited by 122
“1 × every two years) [29].”
Metadata
- Authors
- Gholami, A., Yao, Z., Kim, S., Hooper, C., Mahoney, M. W., & Keutzer, K.
- Affiliations
- International Computer Science Institute (US)
- Snowflake Inc. (United States) (US)
- University of California, Berkeley (US)
- Venue
- IEEE Micro
- Published
- 2024-05
- Volume
- 44
- Issue
- 3
- Pages
- 33-39
- ISSN
- 0272-1732 (ISSN-L)
- 1937-4143
- OpenAlex
- W4393145114
- Cited by
- 369
- References
- 13
- Citations by year
- 2026: 143 · 2025: 166 · 2024: 47
- FWCI
- 110.55
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- Locations
- IEEE Micro (publishedVersion)
- arXiv (Cornell University) (submittedVersion, cc-by)
- KAIST Institutional Repository (KAIST) (submittedVersion)
- arXiv (Cornell University) (cc-by)
- Funders
- National Science Foundation
- U.S. Department of Energy
- Intel Corporation
- Samsung Advanced Institute of Technology
- Korea Foundation for Advanced Studies
- Samsung
- Office of Naval Research
- Topics
- Physical Sciences › Computer Science › Hardware and Architecture › Parallel Computing and Optimization Techniques
- Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Keywords
- memory wall, memory bandwidth, large language models, transformer models, DRAM scaling, interconnect bandwidth, model serving
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
