{"id":"c990f21f-8e48-4875-a0a2-6d5332afbf46","arxiv_id":"2506.07069","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A 3D Gaussian Splatting accelerator uses axis-shared rasterization and a trained MLP to replace sorting, reporting large speedups over edge GPUs with about 1 dB PSNR loss.","lead":"This paper designs a specialized chip that renders 3D Gaussian Splatting scenes faster by sharing math across nearby pixels and replacing depth sorting with a small trained network. If the results hold, real-time augmented reality, robotics, and autonomous driving on low-power devices become more feasible, though the paper reports inconsistent speedup numbers that need scrutiny.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Depth-only MLP transmittance cannot in general reproduce exact sorted alpha blending; the 'negligible quality loss' claim rests on a per-scene co-training procedure and is not established for arbitrary checkpoints or occlusion-heavy scenes.","rationale":"The axis-shared factorization of the 2D Gaussian exponent is mathematically sound and the reconfigurable-PE story is plausible; if the quality assumption held, the speedup and energy numbers would be credible. The load-bearing uncertainty is exactly the neural transmittance: Eq. (4) is not equivalent to Eq. (3) for arbitrary scenes, and the paper's own training details in Sec. 4.2 show that the MLP cannot be trained from scratch and cannot be trained with frozen Gaussians. Thus the final quality is produced by a co-adapted scene representation, not by a drop-in sorting replacement. The evaluation also reports per-scene PSNR gaps up to 1.55 dB, so 'negligible' is not yet established. The proposed test quantifies the frozen-Gaussian failure; if it fails as the paper suggests, the quality-preservation claim must be scoped to their co-training pipeline. This does not invalidate the hardware contributions, but it leaves the central combined claim conditional, matching the reader's verdict.","tokens_in":21210,"tokens_out":10096,"duration_ms":108972,"concrete_test":"Use the released code and pretrained checkpoints. For each MipNeRF-360 scene, train the 10-parameter MLP with the pretrained Gaussians frozen, which is the configuration the paper's Sec. 4.2 says does not work, and measure held-out-view PSNR/SSIM/LPIPS against exact sorted alpha blending on the same Gaussians. If the average PSNR gap exceeds about 1 dB, or any occlusion-heavy scene (e.g., Bonsai or Room) exceeds about 3 dB, then the MLP is not a quality-preserving sorting replacement; the Table 1 results come from co-trained Gaussian modifications and the 23.4 to 27.8x speedup claim must be rescoped to the full co-training pipeline. If instead the frozen-Gaussian gap is within 1 dB across all scenes, the concern is refuted and the central claim stands as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (4) replaces exact front-to-back alpha blending, C = sum T_i alpha_i c_i with T_i = prod_{j<i}(1 - alpha_j), by a normalized weighted sum with weight F(d_i) alpha_i. Exact transmittance T_i depends on the accumulated opacity of all preceding Gaussians, not only on the depth d_i of the i-th Gaussian. A 10-parameter MLP with scalar depth input can at best learn a scene-specific average depth-decay curve; it cannot represent order-dependent interactions. The paper's own Sec. 4.2 admits this: co-training from scratch fails, and freezing the pretrained Gaussians while training only the MLP does not work. Therefore the quality numbers in Table 1 are achieved only after fine-tuning the Gaussians for 15,000 additional epochs with a reduced learning rate, starting from a sorting-based checkpoint. This means the comparison is not a drop-in replacement of sorting for a fixed scene model; it is a new co-trained model. The reported 0.95 dB average PSNR gap (26.50 vs 27.45) is already non-negligible, and per-scene gaps reach 1.43 to 1.55 dB on Room and Bonsai. Because the MLP is per-scene, the gap on unseen scenes or views with denser occlusion is not bounded by the current evaluation. The axis-shared rasterization and hardware ideas may be correct, but the central 'preserves rendering quality' claim for the sorting-free method depends on this unvalidated generalization.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an algorithm-hardware co-design for 3D Gaussian splatting (3DGS). It makes three main contributions: (i) axis-oriented rasterization, which factors the per-pixel Gaussian exponent computation into shared X-axis and Y-axis terms and combines them in each PE, reducing the per-PE arithmetic from 8 multipliers and 4 adders to about 2.25 multipliers and 2.19 adders when including the support PE lines; (ii) a neural sorting scheme that replaces explicit depth sorting with a two-layer, 10-parameter MLP mapping depth d_i to a decay factor F(d_i), used in a normalized weighted sum (Eq. 4); and (iii) a 16x16 reconfigurable PE array that supports both rasterization and MLP inference, together with a fine-grained interleaved pipeline and a pi-trajectory tile schedule. Experiments on six MipNeRF-360 scenes report PSNR 26.50 on average versus 27.45 for a sorted baseline, and hardware evaluations report 23.4-27.8x speedup and 28.8-51.4x energy savings over an edge GPU, plus 1.94-2.39x area-efficiency and 1.36-1.89x energy-efficiency gains over GSCore.","tokens_in":21578,"tokens_out":8143,"duration_ms":80667,"significance":"The axis-shared rasterization identity is a legitimate and clearly explained arithmetic saving: the per-PE MAC reduction is substantial, and the hardware design is presented in enough detail (RTL synthesis, cycle-accurate simulator cross-validated with RTL, CACTI and DRAMPower estimates) to be credible as an engineering contribution. The design-space exploration of the MLP and the ablation study isolating each optimization are also strengths. The paper is less convincing on the algorithmic side: the claim that neural sorting 'preserves rendering quality' is the load-bearing premise for the whole co-design, and that premise currently rests on a per-scene co-training procedure whose comparison baseline is not matched in training budget. If the quality claim can be re-benchmarked and appropriately scoped, the hardware contributions would merit publication.","major_comments":[{"comment":"The central algorithmic claim that Eq. (4) can replace exact front-to-back alpha blending with negligible quality loss is not backed by a correctness argument or a controlled experiment. Exact transmittance in Eq. (3) is T_i = prod_{j<i}(1-alpha_j), which depends on the accumulated opacity of every preceding Gaussian and on their order. A scalar function F(d_i) of depth alone cannot represent such order-dependent interactions in general. The paper itself reports in Sec. 4.2 that co-training the MLP and Gaussians from scratch fails and that freezing the Gaussians while training only the MLP also fails, which shows that the method is not a drop-in sorting replacement for a fixed scene model. The quality numbers in Table 1 are therefore the result of a co-trained model with an additional 15,000 iterations of fine-tuning, not a demonstration that sorting can be omitted without quality loss. Please report the frozen-Gaussian baseline quantitatively and compare against the sorted baseline at matched total training budget.","section":"Sec. 4.1-4.2, Eq. (4)"},{"comment":"The quality comparison is confounded by training protocol. The baseline is trained for 7,000 epochs, while the neural-sorting model starts from that checkpoint and is trained for a further 15,000 epochs with the Gaussian learning rate scaled by 0.01, clone and split disabled, and the best-PSNR checkpoint selected. This gives the proposed method a longer and differently regularized training schedule, so the reported average gap of 0.95 dB PSNR and the per-scene gaps of 1.43-1.55 dB on Room and Bonsai cannot be attributed to the sorting-free formulation alone. The comparison with the sort-free method [16] is also a reproduction under the authors' own framework; the reproduction protocol is not described in enough detail to assess fairness. Please provide per-scene PSNR curves over training time and the exact settings used for the [16] baseline.","section":"Sec. 6.1 and Table 1"},{"comment":"Generalization to unseen scenes is not evaluated. The MLP F(d_i) is trained per scene on the same reconstruction loss used to report quality, so the 'negligible quality loss' claim in the abstract and Sec. 6.5 is only demonstrated on the six training scenes. Scenes with more complex occlusion (Room, Bonsai) already show the largest gaps. To support the architecture-algorithm co-design claim, the authors should either evaluate cross-scene transfer (e.g., train the MLP on one scene and test on held-out scenes or held-out views with denser occlusion) or explicitly limit the claim to per-scene fine-tuned deployment.","section":"Sec. 6.1-6.2"},{"comment":"The headline speedup numbers mix algorithmic and hardware contributions in a way that could mislead. The sorting+rasterization speedup of 23.4-27.8x drops to 7.8-11.5x end-to-end once the edge GPU executes projection, and the comparison with GSCore is reported as area efficiency and energy efficiency rather than end-to-end latency. Please state clearly whether the GSCore comparison includes projection and memory system effects, and report end-to-end latency/energy for both designs under the same workload. Without this, the claimed advantage over prior accelerators is not directly comparable.","section":"Sec. 6.5, Fig. 17"}],"minor_comments":[{"comment":"The arXiv abstract and the full-text title differ ('Axis-Shared Rasterization and Order-independent Transmittance' vs. 'Neural Sorting and Axis-Oriented Rasterization'), and the abstract states a MAC reduction of up to 38% while Sec. 3 states 63%; please align these.","section":"Title/Abstract/Sec. 3"},{"comment":"Eq. (3) uses alpha_i inside the product for T_i; it should be alpha_j (or alpha_{j}) to be correct.","section":"Eq. (3)"},{"comment":"There are numerous typos and inconsistent terms, e.g., 'PNSR' for PSNR, 'redundency', 'minized', and alternating 'GScore'/'GSCore'; please proofread.","section":"Throughout"},{"comment":"'7000epochs' is presumably 'iterations' or 'steps' (the original 3DGS trains for 30k iterations); please clarify the training schedule and use the same unit throughout.","section":"Sec. 6.1"},{"comment":"In Fig. 13, the PSNR differences across activation functions are small (roughly 25.1-25.9 dB); please indicate whether these differences are consistent across scenes or only shown for Garden.","section":"Fig. 13"},{"comment":"The paper states the code is open source at a GitHub URL containing 'ISCA26'; if the manuscript is intended for journal review, the URL should not pre-commit to a specific conference.","section":"Abstract/Code availability"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the manuscript appears to be an extended version of a conference submission (the GitHub URL references ISCA26); if it is under concurrent review elsewhere, that is a policy issue. Also, the split between algorithm and hardware claims may make the paper borderline for a graphics journal: the hardware evaluation is solid, but the algorithmic contribution needs substantial strengthening before acceptance. The reproduction of [16] rather than comparison with original numbers is another fairness concern that the authors should address."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the rasterization factorization is real and the hardware ideas have merit, but the 'neural sorting' claim as stated—preserving quality with negligible loss—does not survive a close read. The paper is worth refereeing because the axis-shared part is sound and the honest reporting of the MLP's training fragility is informative.\n\nWhat's new: Axis-shared rasterization is a clean algebraic factorization of the Gaussian exponent into X/Y quadratic terms and a cross term, precomputing shared terms and broadcasting them. That's a correct common-subexpression elimination; the 63% MAC reduction is arithmetically consistent (8 MUL/4 ADD down to 2 MUL/2 ADD per PE with amortized line overhead). The reconfigurable PE array and interleaved pipeline are well-motivated, and Fig. 16's speedup breakdown is useful. The pi-trajectory tile schedule improving cache hit rate from 43% to 62% is a modest but real gain.\n\nWhere it gets soft: The neural sorting replacement rests on a per-scene, 10-parameter MLP mapping scalar depth to a decay factor. Exact front-to-back blending makes transmittance depend on accumulated opacities of all preceding Gaussians, not just the current depth. A per-scene fit can approximate average depth-decay behavior, but it has no mechanism for order-dependent occlusion. The paper concedes co-training from scratch fails and freezing Gaussians while training only the MLP fails; the working recipe is 15k extra epochs of fine-tuning from a sorted pretrained checkpoint with clone/split disabled and best-PSNR selection. That is a new co-trained model, not a drop-in replacement for sorting, and there's no evidence the MLP generalizes to unseen scenes or denser occlusions. The Table 1 gap of 0.95 dB average, up to 1.55 dB per scene, is not 'negligible' though it is modest. Also, the abstract's 23.4–27.8x speedup and 28.8–51.4x energy savings are the rasterization+sorting partial numbers; the end-to-end figures in the full text are 7.8–11.5x and 4.1–8.8x. That distinction needs to be in the abstract.\n\nBottom line: the paper is a serious hardware co-design with one correct core idea and one overreaching algorithmic claim. A good referee can separate those. I'd send it to review—the axis-shared rasterization and the honest reporting of the MLP's training fragility are worth a public record—but with the expectation that the quality claims and the per-scene nature of the MLP be reworked.","headline":"Rasterization factorization is a real contribution; the neural sorting is a per-scene fit, not a general replacement, and the abstract overstates partial speedups as end-to-end.","tokens_in":22102,"tokens_out":3802,"would_cite":true,"duration_ms":33750,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that 3D Gaussian splatting can run in real time on edge hardware by replacing explicit depth sorting with a tiny learned decay function and reusing shared axis terms in rasterization, achieving 23.4–27.8x speedup and…","keywords":["3D Gaussian Splatting","hardware accelerator","neural sorting","order-independent transparency","axis-shared rasterization","reconfigurable PE array","tile scheduling"],"falsifier":"Take a scene with many overlapping semi-transparent Gaussians at similar depths (foliage, glass layers, smoke), fine-tune the MLP with the paper's recipe, and render novel views; if the average PSNR gap to the full sorted baseline exceeds the roughly 1 dB reported here, or if artifacts such as see-through occluders appear, the order-independent decay assumption is falsified. A more direct test is to compare the learned $F(d_i)$ against the true transmittance $\\prod_{j<i}(1-\\alpha_j)$ for a fixed checkpoint: systematic mismatch in high-occlusion tiles would show the MLP is learning a scene-specific heuristic rather than a general transmittance model.","tokens_in":21033,"feed_emoji":"⚡","tokens_out":6528,"duration_ms":67413,"temperature":0.7,"pith_summary":"The paper tries to establish that 3D Gaussian splatting (3DGS) can be made practical on resource-constrained edge devices by removing two bottlenecks: redundant per-pixel arithmetic in rasterization and the expensive explicit depth sort that feeds alpha blending. It does this by precomputing and broadcasting terms shared along image X- and Y-axes, and by replacing the sort with a two-layer MLP that maps each Gaussian's depth to a decay factor used in an order-independent weighted sum. If correct, a 16x16 reconfigurable FP16 processing array can render MipNeRF-360 scenes at more than 50 FPS end-to-end using 1.64W, while staying within about 1 dB PSNR of the original sorting-based algorithm (26.50 vs 27.45 on average) and beating both edge GPUs and prior dedicated accelerators in speed and energy.","feed_headline":"Tiny neural net replaces depth sorting in Gaussian splatting","feed_subtitle":"Edge chip renders 3D scenes at 50+ FPS with up to 27x speedup, 51x energy savings.","key_machinery":"Two mechanisms carry the argument. Axis-shared rasterization: for a $16\\times16$ tile, dedicated X-PE and Y-PE lines compute the X-axis term, Y-axis term, squared terms, and cross-term coefficient once, then broadcast them to a $16\\times16$ PE array where each PE only combines the received values (2 multiply, 2 add, 1 exponential per pixel instead of 8 multiply, 4 add). Neural sorting: a two-layer MLP with Leaky ReLU hidden activation and exponential output takes depth $d_i$ and outputs the decay factor $F(d_i)$; its 6 MACs fit inside the same reconfigurable PE array, which switches between rasterization mode and sorting mode. Supporting hardware mechanisms include a fine-grained interleaved pipeline that overlaps the memory-bound MLP sorting of the next subtile with rasterization of the current one, and a generalized $\\pi$-trajectory tile schedule (a Hilbert-like ordering) that raises the Gaussian feature cache hit rate from 43% to 62%.","core_discovery":"The central claim is that sorting in 3DGS exists only to compute a per-Gaussian decay factor (the transmittance), and that this decay can be learned rather than explicitly ordered. A tiny MLP with only 10 parameters and 6 MACs maps Gaussian depth $d_i$ to a decay $F(d_i)$, replacing the front-to-back product $T_i = \\prod_{j<i}(1-\\alpha_j)$ with $F(d_i)$ inside the weighted sum $C = \\sum_i F(d_i)\\alpha_i c_i / \\sum_i F(d_i)\\alpha_i$. The paper reports that, when trained from a pretrained sorting-based checkpoint with clone/split disabled and with the MLP learning faster than the Gaussians, this order-independent blending produces rendering quality close to the baseline and better than the prior sort-free weighted-sum approach (average PSNR 26.50 vs 25.43). Alongside this, the paper claims that axis-oriented rasterization reduces the $\\alpha$-computation MAC count by up to 63% with no loss of parallelism.","pith_inferences":["If the central claim holds, the 'sorting is just a decay function' idea may transfer to other order-dependent compositing pipelines, such as particle systems or layered volume rendering, but the depth-to-decay MLP would need per-scene retraining because the paper only demonstrates per-scene fits.","The reported end-to-end speedup (7.8–11.5x) is much smaller than the 23.4–27.8x speedup for rasterization plus sorting, because Gaussian projection still runs on the edge GPU; a full-system benefit would require accelerating projection too, which the paper explicitly leaves to the GPU.","The requirement to start from a pretrained sorting-based checkpoint and to disable clone/split means the method is a fine-tuning recipe, not a from-scratch training pipeline; scenes trained from scratch with the MLP did not converge in the paper's experiments, so adoption in training-from-scratch workflows would need a different initialization strategy."],"forward_implications":["Rendering quality stays within roughly 1 dB PSNR of fully sorted 3DGS while removing the sorting unit from the hardware pipeline; rasterization alone runs at 16.9–20.4x the speed of the edge GPU, exceeding 150 FPS on the tested scenes.","A dedicated 16x16 FP16 array at 3.85 mm^2 and 1.64 W in 28nm can meet real-time (over 50 FPS end-to-end) rendering on MipNeRF-360 scenes, with 23.4–27.8x speedup and 28.8–51.4x energy savings over an edge GPU.","Replacing per-pixel parallel rasterization with axis-shared, broadcast terms cuts the alpha-computation cost from 8 multiplications and 4 additions to 2 and 2 per pixel, reducing total multiplier count by about 63% while keeping one Gaussian-per-cycle throughput.","The interleaved pipeline makes sorting latency effectively negligible: even a naive pipelined version is 102–525x faster than GPU sorting, and the interleaved version reaches roughly 2000x by hiding memory-bound MLP evaluation behind rasterization.","The generalized pi-trajectory tile schedule reduces off-chip energy by 2.56x versus a design without cache and by 1.23x versus the Z-trajectory, because it exploits horizontal, vertical, and hierarchical locality in tile traversal."],"supporting_citations":[{"why":"Defines the original 3DGS rendering pipeline, the alpha-blending equation (3), and the pretrained checkpoints and quality baseline that neural sorting must match.","marker":"[18]"},{"why":"Prior sort-free Gaussian splatting via weighted sum rendering; supplies the comparison baseline (average PSNR 25.43) and the F(d)-weighted sum form that the paper extends with a learned MLP.","marker":"[16]"},{"why":"Weighted blended order-independent transparency; provides the weighted-sum formulation and the motivation that fixed depth-decay functions require scene-specific tuning.","marker":"[24]"},{"why":"GSCore, the prior 3DGS accelerator; supplies the conventional PE-array rasterization and bitonic sorting pipeline that the paper analyzes as inefficient and compares against for area and energy efficiency.","marker":"[22]"},{"why":"MipNeRF-360 dataset; provides the real-world scenes (garden, bicycle, stump, bonsai, counter, kitchen, room) used in all algorithm and hardware evaluations.","marker":"[2]"},{"why":"Bitonic sorting networks; the sorting hardware whose $O(k\\log^2 k)$ area scaling motivates the paper's decision to eliminate explicit sorting.","marker":"[17]"},{"why":"gsplat library; provides the GPU-based 3DGS implementation used as the algorithm baseline for training, inference, and profiling.","marker":"[38]"}],"fun_headline_variants":["MLP learns transmittance, avoids sorting in 3DGS","Axis-shared rasterization cuts 3DGS MACs by 38%","Sort-free 3DGS using MLP-based transmittance","Tiny 10-param MLP replaces Gaussian sort","1.33-1.88x speedup with order-independent 3DGS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single per-scene 10-parameter function of depth alone, trained from a sorted-checkpoint initialization, can reproduce the visual effect of exact front-to-back alpha blending on arbitrary tiles and views; if it cannot on scenes with complex occlusion or view-dependent transparency, the 'negligible quality loss' claim fails even though the rasterization hardware would still work.","fun_headline_variants_meta":{"raw":{"variants":["MLP learns transmittance, avoids sorting in 3DGS","Axis-shared rasterization cuts 3DGS MACs by 38%","Sort-free 3DGS using MLP-based transmittance","Tiny 10-param MLP replaces Gaussian sort","1.33-1.88x speedup with order-independent 3DGS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000807,"raw_usage":{"total_tokens":3594,"prompt_tokens":1049,"completion_tokens":2545,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":665,"completion_tokens_details":{"reasoning_tokens":2448}},"tokens_in":665,"tokens_out":2545,"duration_ms":19397,"temperature":1.0,"reasoning_tokens":2448,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:42:44.847827+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a scene with many overlapping semi-transparent Gaussians at similar depths (foliage, glass layers, smoke), fine-tune the MLP with the paper's recipe, and render novel views; if the average PSNR gap to the full sorted baseline exceeds the roughly 1 dB reported here, or if artifacts such as see-through occluders appear, the order-independent decay assumption is falsified. A more direct test is to compare the learned $F(d_i)$ against the true transmittance $\\prod_{j<i}(1-\\alpha_j)$ for a fixed checkpoint: systematic mismatch in high-occlusion tiles would show the MLP is learning a scene-specific heuristic rather than a general transmittance model.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Weighted blended order-independent transparency; provides the weighted-sum formulation and the motivation that fixed depth-decay functions require scene-specific tuning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Bitonic sorting networks; the sorting hardware whose $O(k\\log^2 k)$ area scaling motivates the paper's decision to eliminate explicit sorting."}],"review_version":1}