Pith. sign in

REVIEW 4 major objections 4 minor

Two-dimensional Sparse Parallelism for Large Scale Deep Learning Recommendation Model Training

T0 review · 4 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A two-dimensional layout for embedding tables lets recommendation training scale to 4,000 GPUs.

desk verdict Plausible and potentially important industrial systems result, but the abstract alone cannot support the scaling claim; the optimizer confound is real and should be checked in review. read the letter →

arxiv 2508.03854 v1 pith:EGYX4RY7 submitted 2025-08-05 cs.DC cs.LG

classification cs.DCcs.LG
keywords two-dimensionalsparseparallelismdeeplearningrecommendationmodelsembeddingtablesmodeldataall-to-allcommunicationAdaGradlarge-scaleGPUtraining
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

This paper claims that the bottleneck in training large recommendation models is how embedding tables are spread across GPUs, and that a two-dimensional arrangement fixes it. Rather than sharding every embedding table across all GPUs, the method shards within groups and replicates across groups, so communication stays inside smaller groups and peak memory drops. The paper pairs this layout with a modified AdaGrad optimizer to keep model quality unchanged. If the claim holds, training a trillion-parameter recommendation model can run on thousands of GPUs with near-linear speedup instead of choking on communication and memory imbalance.

What carries the argument

The central object is the two-dimensional parallelism layout of the embedding tables. It replaces the one-dimensional shard-everything-everywhere scheme with a grid: shard within a group, replicate across groups, so each GPU's all-to-all exchange involves only its group. The momentum-scaled row-wise AdaGrad is the second mechanism; it adjusts per-row adaptive learning rates with momentum to keep convergence on replicated rows close to the fully sharded baseline. Together they carry the paper's scalability and quality claims.

What would settle it

Run the same 2D layout with the standard (non-momentum) AdaGrad and the 1D fully-sharded layout with momentum-scaled AdaGrad; if the speedup and memory reduction track the optimizer rather than the layout, the central scalability claim fails. A simpler check: profile per-component time and memory (all-to-all, embedding lookup, optimizer update) on the 2D layout; if the all-to-all cost does not drop relative to the 1D baseline, the mechanism is not doing the claimed work.

Watch

Extended reading notes

Core claim

The paper proposes two-dimensional sparse parallelism as a replacement for fully sharded embedding tables in distributed recommendation model training. In this scheme, the GPU cluster is partitioned into groups: embedding tables are sharded across GPUs within each group (model parallelism), while the same sharded tables are replicated across groups (data parallelism). This keeps the all-to-all lookup communication within each group and removes the need for a global all-to-all, lowering both communication volume and peak activation memory. To prevent the replication from degrading convergence, the paper introduces momentum-scaled row-wise AdaGrad, which scales per-row adaptive learning rates by a momentum term. Across experiments, the method reaches near-linear scaling on up to 4,000 GPUs while matching the model quality of baseline training.

Load-bearing premise

The central claim rests on the assumption that the two-dimensional layout itself, rather than the new optimizer or the particular network hardware, is what reduces communication and memory costs, and that this reduction does not simply move the bottleneck elsewhere, such as to embedding activation memory or stragglers.

Editorial extensions

If this is right

  • Training runs on industrial-scale recommendation models can reach near-linear speedup with GPU count instead of being limited by the all-to-all communication bottleneck.
  • The same quality of model is achievable with the 2D layout plus momentum-scaled AdaGrad, so practitioners do not have to sacrifice accuracy for speed.
  • Peak memory per GPU is reduced, which allows larger embedding tables or bigger batch sizes to fit on the same hardware.

Reading between the lines

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

  • By extension, the same two-dimensional trick may apply to other models with large sparse tables, such as graph neural networks or language models with massive vocabularies, wherever a global all-to-all is the bottleneck.
  • If momentum-scaled AdaGrad is what preserves convergence, the layout could also be paired with other optimizers, but the optimizer's contribution would need separate verification.
  • A direct comparison of per-GPU memory and communication breakdowns between 2D and 1D layouts, with the optimizer held fixed, would settle whether the 2D layout or the optimizer drives the gains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a two-dimensional sparse parallelism strategy for large-scale deep learning recommendation model training. Instead of fully sharding embedding tables across all GPUs, the method adds data parallelism on top of model parallelism, aiming to improve all-to-all communication efficiency and reduce peak memory consumption. The abstract also introduces a momentum-scaled row-wise AdaGrad optimizer to mitigate quality loss from the shift in training paradigm. The empirical claims are nearly linear training speed scaling up to 4K GPUs and model performance parity compared with existing approaches.

Significance. If the empirical claims are substantiated, the work would address a recognized bottleneck in industrial-scale recommendation model training: the high communication and memory cost of sparse embeddings. The two-dimensional layout and the accompanying optimizer variant are plausible practical contributions, and the scale of the reported experiments (up to 4K GPUs) is ambitious. However, the significance cannot be assessed from the abstract alone. The central claims of near-linear scaling and performance parity require detailed experimental evidence, including baseline comparisons, datasets, and the full system architecture. The paper's value is conditional on the missing evidence.

major comments (4)
  1. [Abstract, scaling claim] The abstract states 'achieves nearly linear training speed scaling up to 4K GPUs' without specifying whether the scaling is weak or strong, the hardware and network topology (e.g., NVLink, InfiniBand), the baseline distributed training configuration, or the number of runs and variance. As written, this claim cannot be reproduced or independently assessed.
  2. [Abstract, optimizer confound] The momentum-scaled row-wise AdaGrad algorithm is introduced in the same abstract as the layout change and is described as mitigating 'performance losses associated with the shift in training paradigms', yet no ablation is reported that separates the effect of the 2D layout from the effect of the new optimizer. Without such an ablation, the attribution of the scaling and parity results to the 2D layout is not established.
  3. [Abstract, memory claim] The claim that adding data parallelism on top of model parallelism 'reduces peak memory consumption' is non-obvious because data parallelism replicates embedding tables across the data-parallel dimension. A rigorous memory accounting that includes embedding tables, activations, gradients, and optimizer states is needed to support this claim, and the abstract provides no such breakdown.
  4. [Abstract, performance parity] The assertion of 'maintaining model performance parity' is stated without defining the evaluation metric (e.g., AUC, NDCG, loss), the baseline systems used for comparison, or the statistical significance of the parity claim. 'Parity' is a strong claim and requires a clear experimental protocol that is not visible in the abstract.
minor comments (4)
  1. [Abstract, language] The phrase 'fully parallelism strategies' should be 'fully parallel strategies', and 'post significant scalability challenges' should be 'pose significant scalability challenges'.
  2. [Abstract, definition of 2D parallelism] The abstract uses 'two-dimensional sparse parallelism' without precisely defining the two dimensions; a figure or formal algorithm description in the full text would help readers understand the proposed layout.
  3. [Abstract, state-of-the-art claim] The statement 'setting a new state-of-the-art benchmark for recommendation model training' is unsupported by any named comparison systems or public benchmarks in the abstract.
  4. [Abstract, related work] The abstract does not mention prior work on large-scale DLRM training, such as Facebook's DLRM, HugeCTR, or embedding table sharding techniques; the full paper should provide these references to position the contribution.

Circularity Check

0 steps flagged · score 0.0 of 10

Abstract-only review finds no circular derivation to exhibit.

full rationale

The reviewable material is limited to the abstract, which contains no equations, no fitted parameters, no derived predictions, and no citation chain. The central claim is an empirical systems result: the proposed two-dimensional sparse parallelism achieves nearly linear training speed scaling up to 4K GPUs while maintaining model performance parity. This is a measured outcome, not a quantity derived from its own inputs. The abstract mentions a co-designed optimizer, momentum-scaled row-wise AdaGrad, but does not define it in terms of the target result or present any numerical fit that would make the scaling claim true by construction. Concerns about whether the layout, the optimizer, or the hardware environment causes the observed scaling are attribution questions about an empirical benchmark, not circular reasoning. Under the requirement that circularity be exhibited by quoting a specific reduction, no such reduction can be identified from the available text. The correct finding is therefore no significant circularity.

Assumptions & free parameters 2 free parameters · 1 assumptions · 0 invented entities

Only the abstract was available. The ledger lists parameters and assumptions that are implicit in the abstract; full details would require the complete manuscript.

free parameters (2)
  • 2D parallel split ratio (data-parallel versus model-parallel dimension) = not reported in abstract
    The choice of how many GPUs form data-parallel replicas versus model-parallel shards is a design parameter that affects memory and communication; abstract does not specify it.
  • Optimizer hyperparameters for momentum-scaled row-wise AdaGrad = not reported in abstract
    Step size and momentum-related scaling are needed to reproduce claimed convergence parity; not visible in abstract.
assumptions (1)
  • domain assumption Embedding table memory and lookup communication are the dominant scalability bottlenecks in large-scale DLRM training.
    The method is motivated by these bottlenecks; if they are not dominant on the target cluster, the approach may not generalize.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Two-dimensional Sparse Parallelism for Large Scale Deep Learning Recommendation Model Training." pith.science (2026). https://pith.science/paper/EGYX4RY7

@misc{pith2026250803854,
  author       = {Pith},
  title        = {Pith review of: Two-dimensional Sparse Parallelism for Large Scale Deep Learning Recommendation Model Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EGYX4RY7}},
  note         = {Machine review of arXiv:2508.03854}
}
read the original abstract

The increasing complexity of deep learning recommendation models (DLRM) has led to a growing need for large-scale distributed systems that can efficiently train vast amounts of data. In DLRM, the sparse embedding table is a crucial component for managing sparse categorical features. Typically, these tables in industrial DLRMs contain trillions of parameters, necessitating model parallelism strategies to address memory constraints. However, as training systems expand with massive GPUs, the traditional fully parallelism strategies for embedding table post significant scalability challenges, including imbalance and straggler issues, intensive lookup communication, and heavy embedding activation memory. To overcome these limitations, we propose a novel two-dimensional sparse parallelism approach. Rather than fully sharding tables across all GPUs, our solution introduces data parallelism on top of model parallelism. This enables efficient all-to-all communication and reduces peak memory consumption. Additionally, we have developed the momentum-scaled row-wise AdaGrad algorithm to mitigate performance losses associated with the shift in training paradigms. Our extensive experiments demonstrate that the proposed approach significantly enhances training efficiency while maintaining model performance parity. It achieves nearly linear training speed scaling up to 4K GPUs, setting a new state-of-the-art benchmark for recommendation model training.

Discussion (0). Continue with ORCID to comment.

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.