Searcharxiv⌕ Search

arXiv subjects

Kevin Bhimani

Publications and source records attributed to Kevin Bhimani.

2 recordsLinked to original sources

Benchmarking LLM Inference on Desktop-Scale NVIDIA Grace Blackwell Clusters

Two NVIDIA DGX Spark GB10 systems connect directly over 200 GbE to serve larger language models and increase inference capacity, but the benefits depend on how they are deployed. We benchmark 17 open-weight checkpoints spanning dense and mixture-of-experts architectures, 8B to 235B parameters and four weight precisions on one and two RDMA-connected systems. Using vLLM, we compare tensor parallelism (TP=2), pipeline parallelism (PP=2) and a two-replica baseline at concurrency 1 to 64, with three independent launches per configuration. TP=2 delivers a median 1.59x peak-throughput gain over one Spark, versus 1.39x for PP=2 and has higher throughput in 93 of 96 matched comparisons with separated 95% Welch intervals. It reduces unloaded inter-token latency to 0.52x to 0.73x of one Spark, whereas PP=2 provides little reduction. TP=2 also achieves the highest goodput under a strict 100 ms inter-token latency objective wherever both strategies are supported and the objective is met. For models fitting on one Spark, TP=2 generally favors low-concurrency serving, while replication offers higher throughput at high concurrency. One Spark uses less GPU energy per output token at matched concurrency. Across three precision ladders, NVFP4 reduces unloaded latency more than adding a second Spark at BF16. NCCL microbenchmarks and comparisons with an earlier TCP campaign suggest that collective communication overhead contributes to the advantage of TP=2. These results provide deployment guidance for desktop-scale inference.

cs.DC↗

CycleGAN-Driven Transfer Learning for Electronics Response Emulation in High-Purity Germanium Detectors

High-Purity Germanium (HPGe) detectors are a key technology for rare-event searches such as neutrinoless double-beta decay (\ensuremath{0νββ}) and dark matter experiments. Pulse shapes from these detectors vary with interaction topology and thus encode information critical for event classification. Pulse shape simulations (PSS) are essential for modeling analysis cuts that distinguish signal events from backgrounds and for generating reliable simulations of energy spectra. Traditional PSS methods rely on a series of first-principles corrections to replicate the effect of readout electronics, requiring challenging fits over large parameter spaces and often failing to accurately model the data. We present a neural network architecture, the Cyclic Positional U-Net (https://github.com/aobol/CPU-Net), that performs translations of simulated pulses so that they closely resemble measured detector signals. Using a Cycle Generative Adversarial Network (CycleGAN) framework, this {Response Emulation Network} (REN) learns a data-driven mapping between simulated and measured pulses with high fidelity, without requiring a predetermined response model. We use data from a High-Purity Germanium (HPGe) detector with an inverted-coaxial point contact (ICPC) geometry to show that \texttt{CPU-Net} effectively captures and reproduces critical pulse shape features, allowing more realistic simulations without detector-specific tuning. \texttt{CPU-Net} achieves up to a factor-of-four improvement in distribution-level agreement for pulse shape parameter reconstruction, while preserving the topology-dependent information required for pulse-shape discrimination.

physics.ins-det↗