Tootfinder

A Full-Stack Characterization of High-Bandwidth Flash for KV-Centric LLM Serving
Zhuoran Li, Zhuohang Bian, Xin Huang, Yibo Zhao, Guangyu Sun, Youwei Zhuo
https://arxiv.org/abs/2608.11668 https://arxiv.org/pdf/2608.11668 https://arxiv.org/html/2608.11668
arXiv:2608.11668v1 Announce Type: new
Abstract: High-Bandwidth Flash (HBF) stacks NAND behind a wide, package-local interface, giving flash-scale capacity with far better read latency and bandwidth than an SSD. This makes it tempting to keep an SSD-style Mooncake KV-offloading stack and swap only the backing tier for HBF. We test that substitution with an extended TokenSim, four complete two-hour Qwen-Bailian production traces, five dense and mixture-of-experts models, and H100/B200 profiles. Serving gets worse, not better, and a cost-benefit model explains why. A faster far tier helps only when read I/O is the serving bottleneck, reads outweigh writes, and delivered bandwidth is sustainable. All three must hold together, and transient KV fails every one. The package trade that buys flash costs GPU near-tier capacity and bandwidth, so average end-to-end latency rises 2--5.5x and maximum SLO goodput falls 1.1--2.7x across H100 and B200. Serving is almost insensitive to HBF's own read/write latency, and base-die near-memory compute does not raise the flash tier's share of the critical path. The two-tier hierarchy keeps reuse in the near tier and hands HBF a write-heavy stream, so writes outnumber reads on every trace. A 3D-ICE model shows that stream drives the stack to its thermal limit well below peak bandwidth, and a TLC tier wears out sooner than a capacity-matched SSD pool. The faster device yields a slower system because the package gives up more than the medium returns. HBF is not the problem; using it as a faster SSD for transient KV is. It belongs in serving as a selective, reuse-aware, write-budgeted, and thermally coordinated tier, not as a drop-in SSD replacement.
toXiv_bot_toot

Sensitivity of silicon-to-water dose conversion and Bragg-peak metrics to stopping-power datasets in proton dosimetry
F. Matias, J. M. B. Shorto, P. de Vera, R. Garcia-Molina, I. Abril, T. F. Silva, J. Pereira, H. Yoriyaz
https://arxiv.org/abs/2607.16441 https://arxiv.org/pdf/2607.16441 https://arxiv.org/html/2607.16441
arXiv:2607.16441v1 Announce Type: new
Abstract: Accurate proton dosimetry requires consistent stopping-power data for detector-to-water conversion and Monte Carlo radiation transport. We quantify the sensitivity of water-to-silicon stopping-power ratios, converted dose-to-water distributions, and Bragg-peak metrics to the stopping-power dataset for a 67.5 MeV pristine proton beam in water. PHITS simulations used SRIM-2013, PSTAR/NIST, ATIMA, and TDDFT-Penn stopping powers up to 10 MeV, together with a common SBETHE-based extension at higher energies. The water-to-silicon stopping-power ratio showed relative entrance-to-distal variations of 17.40%$-$22.10%. Conversion of normalized PTW silicon-diode percentage-depth-ionization data yielded relative percentage-depth-dose curves with a common maximum at 36.86 mm, consistent with the experimental Bragg-peak depth of (36.81 $\pm$ 0.15) mm. The converted curves were approximately 3% below the reference throughout the entrance and plateau regions, while deviations of up to 2.1% occurred near the Bragg peak. Direct PHITS calculations predicted Bragg-peak depths of 36.575$-$36.675 mm; only the TDDFT-Penn result lay within the quoted experimental uncertainty. These results show that stopping-power treatment affects the magnitude of the converted dose more strongly than the Bragg-peak position does, and that it should be harmonized or explicitly included in uncertainty budgets for proton-beam calibration, commissioning, quality assurance, and Monte Carlo validation.
toXiv_bot_toot