Papers for
gpu system architects
Papers whose findings have a practical use for this group, as judged from the abstract. Open a paper to read what it means in practice.
Ising machine chip cuts energy for context selection in AI models
p-DCIM: A 28nm Probabilistic SRAM DCIM based Ising Machine for Near-Memory Context Selection in LLM-RAG Pipeline
Abstract: We present a fabricated SRAM DCIM Ising machine that performs set-level context selection for large language model (LLM) with retrieval-augmented generation (RAG). In RAG, the evidence-selection stage is increasingly delegated to repeated language-model reranking, which adds substantial energy to the inference pipeline. We reformulate evidence selection as a quadratic unconstrained binary optimization (QUBO) problem and solve it on a dual-mode digital compute-in-memory (DCIM) macro: the same multiply-and-accumulate datapath computes the embedding-similarity scores that populate the QUBO coefficients and evaluates the per-spin Hamiltonian differences that drive the anneal. The annealing temperature is set by a digital plane-count control along with the SRAM pseudo-read random bit flips, giving graded cooling entirely from on-chip digital configuration. The macro is taped-out in a 28 nm process [0.16 mm2, 0.41 mW] and is proposed to be integrated as part of the logic base-die in hybrid high-bandwidth memory (HBM) and high-bandwidth flash (HBF) stacks in a GPU package, where only the selected passage identifiers return to the host. Replacing the language-model selection pass with the near-memory Ising solver yields a macro normalized selection-energy estimate more than six orders of magnitude lower than the LLM-based baseline and reduces the evaluated total energy per query by 67%, while maintaining comparable token-level F1 on the evaluated examples.
Flash RNN-T cuts memory use and speeds speech transcription
Flash RNN-T: Memory-Efficient Joint Network and Transducer Loss Computation
Abstract: The recurrent neural network transducer (RNN-T) remains a practical choice for streaming speech recognition, but training requires large intermediate tensors. We present Flash RNN-T, a joint and loss implementation combining specialized kernels with optional tiled recomputation over the full alignment space and vocabulary. Tiled and untiled Flash form the entire measured throughput-memory frontier on medium and long workloads. On H100, untiled Flash achieves 24-39% higher throughput with 26-56% less memory than the fastest baseline on these workloads wherever it fits. Tiling reduces memory further through recomputation. On medium, tiled Flash uses ninefold less memory than that baseline at 16% lower throughput when both use one backward call per batch. On long, it fits where all tested alternatives with that backward schedule run out of memory with 80 GB. These comparisons use each implementation's lowest-memory configuration reaching at least 95% of its peak throughput.
Memory budget increases can unexpectedly break neural network training
When More Memory Fails: Hidden Feasibility Holes in Dynamic Tensor Rematerialization
Abstract: Dynamic Tensor Rematerialization (DTR) trains neural networks under a memory budget by evicting tensors and recomputing them on demand. A natural expectation is that a budget large enough to complete training stays sufficient when it is increased. Using DTR's reference simulator, simrd, to replay the public DTR training traces, we show that this expectation fails. We call the exceptions feasibility holes: memory budgets at which an eviction policy runs out of memory although both a smaller and a larger budget complete, a form of non-monotone feasibility. Under a written sweep protocol with dated amendments, a 0.01-step budget grid shows no such hole for DTR on any of six confirmatory traces, although unresolved timeouts prevent it from establishing monotonicity. Finer sampling finds holes on InceptionV4 (three regions) and U-Net (two regions), in addition to one on the ResNet-32 development trace. Every hole is reproduced by three repeats and by unmodified simrd. For one InceptionV4 pair (0.2343 versus 0.235 of peak memory) we trace the failure to individual allocations. The larger budget avoids one eviction at an operator where the smaller budget was 173,479 bytes short. The extra data it carries forward forces a later eviction of a tensor needed three operators afterwards, and rebuilding that tensor triggers a recomputation cascade that cannot fit. Keeping that one tensor resident under the same budget prevents the failure. The two simple score changes we tested do not remove holes across traces; both add failing budgets on some traces. All results are from simulation, with no allocator model; behaviour on GPU hardware is untested.
Offloading gpu communication tasks frees cores for faster ai training
HOCCL: Offloading Collective Communication from GPU Cores to Accelerate Distributed Training
Abstract: Large language model training involves massive computation on GPU streaming multiprocessors (SMs), the primary compute units of GPUs. Since SMs host specialized accelerators such as Tensor Cores, their efficient utilization is critical to training efficiency. Unfortunately, existing collective communication systems compete with computation for SMs, as they consume SMs for communication-related data movement and synchronization operations. We observe that communication can, in principle, be driven by DMA engines, thereby eliminating SM involvement in communication. Based on this insight, we propose HOCCL, a zero-SM collective communication framework consisting of three components: a stream manager, a point-to-point (P2P) executor, and a collective scheduler. The stream manager preserves operator-level temporal ordering with other GPU kernels. The P2P executor enables zero-SM point-to-point communication, while the collective scheduler orchestrates P2P transfers to maximize bandwidth. Experiments show that HOCCL preserves near-peak communication performance, achieving within 3% of the state of the art on average, while eliminating communication occupancy on nearly 10% of total GPU SMs. By freeing SM resources for computation, HOCCL improves end-to-end training throughput by up to 5%.
SketchSSM reduces state access to speed up transformer decoding
SketchSSM: Write to the Full State, Read from a Compact Sketch
Abstract: Hybrid-attention models replace most softmax attention layers with linear attention, reducing KV-cache growth and enabling larger decode batches where recurrent state access becomes a major bottleneck. ReplaySSM amortizes state updates by buffering keys and values, but each new query still requires a full-state read even though the state remains unchanged between state updates. We observe that low-rank state-weighted query approximation accurately preserves state-read outputs. Although future queries are unknown, the basis vectors used to approximate them can be fixed offline. Based on this observation, we introduce SketchSSM, which preserves full-state updates while approximating reads. At each state update, SketchSSM reads the full state once to precompute outputs for these basis vectors, storing them in a compact sketch. Each subsequent decode step combines the sketch vectors with query-dependent coefficients to reconstruct the output without a full-state read. Across four Mamba-2-, GDN-, and KDA-based models, SketchSSM reduces state-access traffic by approximately 10x while largely preserving average accuracy across four decode benchmarks and recall on four RULER retrieval tasks. On one NVIDIA B300, linear-attention kernel speedups over the standard vLLM baseline reach 7.78x, 5.22x, and 5.20x for Mamba-2, GDN, and KDA, respectively, with up to 2.64x higher decode throughput on Nemotron 3 Super.
Hybrid memory setup boosts large language model session handling
Hot-Cold Tiering of HBM and High Bandwidth Flash for Agentic LLM Serving
Abstract: Large language model (LLM) serving is increasingly agentic, with multi-turn sessions that idle between actions yet must retain their full context. Limited GPU memory capacity forces inactive KV states to be evicted, so resuming a session incurs either costly recomputation or slow interconnect transfers. To address this, high bandwidth flash (HBF)-an on-package 3D-NAND memory offering orders-of-magnitude greater capacity than high bandwidth memory (HBM) at comparable read bandwidth-has emerged as a strong candidate. However, its high read energy and limited write endurance make it impractical to serve all KV traffic. Fortunately, our analysis shows that agentic KV states exhibit distinct access patterns: a small hot set is read for every decoding step, while a large cold pool is read only when a paused session resumes. Exploiting this, we place the hot set in HBM and the cold pool in HBF, forming a hot-cold KV hierarchy within the GPU memory tier. On agentic workloads with Qwen3-Coder-30B-A3B, our design delivers 14 ms time-between-tokens (TBT) and adds only $\approx$0.1 ms of resume latency on top of prefill, while hosting $24\times$ more concurrent sessions per GPU. By confining HBM to the hot set, our design also cuts read power by 7.6 kW per 8-GPU node relative to serving all KV from flash-establishing HBF as a cold-tier complement to HBM rather than its replacement.
Die scaling causes GPU resource asymmetry that impacts performance
Dissecting How Die Scaling Breaks GPU Fine-grained Scheduling
Abstract: Modern GPUs are no longer physically symmetric. Die scaling leads to both manufacturing-driven floorsweeping and cache and memory partitioning. The former creates chip-specific compute topologies, while the latter causes non-uniform memory access. These asymmetries are substantial. Topology-oblivious compute unit allocation can lead to up to 1.33x performance variation, while remote accesses increase HBM latency by up to 67% and nearly double L2 latency. However, these asymmetries are hidden behind the GPU's logical resource abstractions and can vary across chips. We develop lightweight characterization methods to uncover per-chip compute topology and memory affinity. We then use the discovered information to make existing fine-grained scheduling asymmetry-aware, considering not only how many resources are allocated but also which physical resources are assigned. Across full-GPU kernel execution, intra-application multiplexing, and inter-application co-location, asymmetry-aware scheduling improves mainstream kernels by up to 1.22x, multiplexed LLM inference by up to 14.3%, and avoids up to 1.33x performance variation.
Transformer models shrink channels to save computing and energy
Dense Structural Compression of Transformers via Gauge-Correct Channel Removal
Abstract: Inference energy per token drives the cost and carbon footprint of deployed transformers. It is dominated by dense matrix products that incur fused multiply-accumulate (FMA) operations and memory traffic. To reduce these computations while retaining dense tensors for high GPU throughput, we develop a methodology from first principles to adapt structural complexity during training to maximize inference utility per unit compute. Channel penalties drive entire tensor slices to zero to enable physical removal while preserving density and the network function. The natural approach, penalizing the norm of operator components acting through each channel, is provably destabilized by gauge freedom. We resolve this pathology with GaugeLasso: additive symmetric group-lasso penalties that recover a monotone function of product-norms when the network converges to gauge balance. Our equilibrium analysis enables per-channel calibration to correctly suppress slices that under-perform in inference utility per unit compute. Under adaptive pressure, the network reorganizes into depth-dependent structural profiles that can be far smaller than the architecture required to learn the task. On polynomial long division over $\mathbb{F}_{31}$, compute compresses from 148 to 255 times with perfect accuracy. On character-level language modeling, compressed models outperform the hand-designed baseline at equal FMA. On masked autoencoding, a compression trial exposes which axes were over-provisioned and which saturated, guiding a better second design. Compaction also accelerates training monotonically as the model progresses. Post-hoc pruning with the same utility ranking cannot reach these structures, showing that sustained pressure is central to discovery of efficient models. Retraining a discovered architecture recovers baseline quality on our statistical tasks, but fails on our exact algorithmic task.