{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,3,5]],"date-time":"2026-03-05T07:14:19Z","timestamp":1772694859877,"version":"3.50.1"},"reference-count":45,"publisher":"IEEE","license":[{"start":{"date-parts":[[2026,1,31]],"date-time":"2026-01-31T00:00:00Z","timestamp":1769817600000},"content-version":"stm-asf","delay-in-days":0,"URL":"https:\/\/doi.org\/10.15223\/policy-029"},{"start":{"date-parts":[[2026,1,31]],"date-time":"2026-01-31T00:00:00Z","timestamp":1769817600000},"content-version":"stm-asf","delay-in-days":0,"URL":"https:\/\/doi.org\/10.15223\/policy-037"}],"content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"published-print":{"date-parts":[[2026,1,31]]},"DOI":"10.1109\/hpca68181.2026.11408526","type":"proceedings-article","created":{"date-parts":[[2026,3,4]],"date-time":"2026-03-04T20:47:22Z","timestamp":1772657242000},"page":"1-19","source":"Crossref","is-referenced-by-count":0,"title":["Characterizing Cloud-Native LLM Inference at Bytedance and Exposing Optimization Challenges and Opportunities for Future AI Accelerators"],"prefix":"10.1109","author":[{"given":"Jingwei","family":"Cai","sequence":"first","affiliation":[{"name":"ByteDance Seed"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Dehao","family":"Kong","sequence":"additional","affiliation":[{"name":"ByteDance Seed"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Hantao","family":"Huang","sequence":"additional","affiliation":[{"name":"ByteDance Seed"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Zishan","family":"Jiang","sequence":"additional","affiliation":[{"name":"ByteDance Seed"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Zixuan","family":"Ma","sequence":"additional","affiliation":[{"name":"ByteDance Seed"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Qingyu","family":"Guo","sequence":"additional","affiliation":[{"name":"ByteDance Seed"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Zhenxing","family":"Zhang","sequence":"additional","affiliation":[{"name":"ByteDance Seed"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Guiming","family":"Shi","sequence":"additional","affiliation":[{"name":"Tsinghua University"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Mingyu","family":"Gao","sequence":"additional","affiliation":[{"name":"Tsinghua University"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Kaisheng","family":"Ma","sequence":"additional","affiliation":[{"name":"Tsinghua University"}],"role":[{"role":"author","vocabulary":"crossref"}]},{"given":"Minghui","family":"Yu","sequence":"additional","affiliation":[{"name":"ByteDance Seed"}],"role":[{"role":"author","vocabulary":"crossref"}]}],"member":"263","reference":[{"key":"ref1","doi-asserted-by":"publisher","DOI":"10.1145\/3470496.3527405"},{"key":"ref2","volume-title":"Gqa: Training generalized multi-query transformer models from multi-head checkpoints","author":"Ainslie","year":"2023"},{"key":"ref3","article-title":"Amd instinct ${ }^{\\text{TM}}$ mi350 series gpuss","volume-title":"AMD"},{"key":"ref4","article-title":"System card: Claude opus 4 & claude sonnet 4","volume-title":"Anthropic"},{"key":"ref5","article-title":"Inferentia2 architecture","volume-title":"AWS"},{"key":"ref6","volume-title":"Maia2","author":"Azure"},{"key":"ref7","doi-asserted-by":"publisher","DOI":"10.1145\/3579371.3589048"},{"key":"ref8","doi-asserted-by":"publisher","DOI":"10.1109\/HPCA57654.2024.00022"},{"key":"ref9","doi-asserted-by":"publisher","DOI":"10.1145\/3731569.3764808"},{"key":"ref10","doi-asserted-by":"publisher","DOI":"10.1145\/3695053.3731409"},{"key":"ref11","author":"Comanici","year":"2025","journal-title":"Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities"},{"key":"ref12","author":"Dao","year":"2023","journal-title":"Flashattention-2: Faster attention with better parallelism and work partitioning"},{"key":"ref13","doi-asserted-by":"publisher","DOI":"10.52202\/068431-1189"},{"key":"ref14","author":"Guo","year":"2025","journal-title":"Deepseek-r1: Incentivizing reasoning capability in 11 ms via reinforcement learning"},{"key":"ref15","author":"Liu","year":"2024","journal-title":"Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model"},{"key":"ref16","doi-asserted-by":"publisher","DOI":"10.1145\/3297858.3304014"},{"key":"ref17","volume-title":"Better & faster large language models via multi-token prediction","author":"Gloeckle","year":"2024"},{"key":"ref18","doi-asserted-by":"publisher","DOI":"10.1145\/3582016.3582018"},{"key":"ref19","article-title":"Characterization of large language model development in the datacenter","volume-title":"Proceedings of the 21st USENIX Symposium on Networked Systems Design and Implementation, ser. NSDI\u201924","author":"Hu","year":"2024"},{"key":"ref20","doi-asserted-by":"publisher","DOI":"10.1145\/3372780.3380846"},{"key":"ref21","author":"Jouppi","year":"2023","journal-title":"Tpu v4: An optically reconfigurable supercomputer for machine learning with hardware support for embeddings"},{"key":"ref22","doi-asserted-by":"publisher","DOI":"10.1145\/3079856.3080246"},{"key":"ref23","volume-title":"Transformers are rnns: Fast autoregressive transformers with linear attention","author":"Katharopoulos","year":"2020"},{"key":"ref24","volume-title":"Fast inference from transformers via speculative decoding","author":"Leviathan","year":"2023"},{"key":"ref25","author":"Li","year":"2025","journal-title":"Eagle-3: Scaling up inference acceleration of large language models via training-time test"},{"key":"ref26","volume-title":"Ub-mesh: a hierarchically localized nd-fullmesh datacenter network architecture","author":"Liao","year":"2025"},{"key":"ref27","doi-asserted-by":"publisher","DOI":"10.1109\/hcs52781.2021.9567153"},{"key":"ref28","article-title":"Understanding stragglers in large model training using what-if analysis","volume-title":"Proceedings of the 19th USENIX Conference on Operating Systems Design and Implementation, ser. OSDI \u201925","author":"Lin","year":"2025"},{"key":"ref29","doi-asserted-by":"publisher","DOI":"10.1145\/3696443.3708967"},{"key":"ref30","article-title":"Blackwell-series accelerators","volume-title":"NVDIA"},{"key":"ref31","article-title":"Rubin cpx","volume-title":"NVDIA"},{"key":"ref32","article-title":"Gpt-5 system card","volume-title":"OPENAI"},{"key":"ref33","doi-asserted-by":"publisher","DOI":"10.1109\/micro61859.2024.00100"},{"key":"ref34","doi-asserted-by":"publisher","DOI":"10.1109\/isca45697.2020.00045"},{"key":"ref35","author":"Shah","year":"2024","journal-title":"Flashattention-3: Fast and accurate attention with asynchrony and lowprecision"},{"key":"ref36","volume-title":"Step-3 is large yet affordable: Model-system co-design for cost-effective decoding","author":"Wang","year":"2025"},{"key":"ref37","doi-asserted-by":"publisher","DOI":"10.1109\/hcs55958.2022.9895534"},{"key":"ref38","volume-title":"Llama: Open and efficient foundation language models","author":"Touvron","year":"2023"},{"key":"ref39","doi-asserted-by":"publisher","DOI":"10.1109\/mm.2021.3061912"},{"key":"ref40","volume-title":"Tilelang: A composable tiled programming model for ai systems","author":"Wang","year":"2025"},{"key":"ref41","article-title":"Grok 4 model card","year":"2025","journal-title":"xAI"},{"key":"ref42","volume-title":"Servegen: Workload characterization and generation of large language model serving in production","author":"Xiang","year":"2025"},{"key":"ref43","volume-title":"Qwen3 technical report","author":"Yang","year":"2025"},{"key":"ref44","doi-asserted-by":"publisher","DOI":"10.1007\/s11390-024-4178-1"},{"key":"ref45","volume-title":"Megascaleinfer: Serving mixture-of-experts at scale with disaggregated expert parallelism","author":"Zhu","year":"2025"}],"event":{"name":"2026 IEEE International Symposium on High Performance Computer Architecture (HPCA)","location":"Sydney, Australia","start":{"date-parts":[[2026,1,31]]},"end":{"date-parts":[[2026,2,4]]}},"container-title":["2026 IEEE International Symposium on High Performance Computer Architecture (HPCA)"],"original-title":[],"link":[{"URL":"http:\/\/xplorestaging.ieee.org\/ielx8\/11408404\/11408433\/11408526.pdf?arnumber=11408526","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,3,5]],"date-time":"2026-03-05T06:35:55Z","timestamp":1772692555000},"score":1,"resource":{"primary":{"URL":"https:\/\/ieeexplore.ieee.org\/document\/11408526\/"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,1,31]]},"references-count":45,"URL":"https:\/\/doi.org\/10.1109\/hpca68181.2026.11408526","relation":{},"subject":[],"published":{"date-parts":[[2026,1,31]]}}}