思维纪要社 HTML 来自本地 Obsidian vault 的静态发布
主题 2026-09-21/topics/508.md 原文

#外资研报

图片

无图片

正文

详情页仍未获取到全文。原始列表正文可能被截断。

总体总结

主题正文

  1. 详情页仍未获取到全文。
  2. 原始列表正文可能被截断。

附件总结 Toward Recursive Self-Improvement- How GLM Built Its Own Inference Infrastructur.pdf

  1. In October 2025, we began researching how to strengthen its cybersecurity capabilities. Our
  2. production-grade inference service from scratch on a cluster of more than 100,000 Chinese-
  3. ultimately tripling end-to-end throughput relative to the initial baseline. Figure 1 shows the
  4. performance trajectory of GLM-5.3-Flash from its first successful run to production launch.
  5. test failed," "TTFT increased by 30%," or "output throughput dropped by 20%" after a change
  6. Transfer, following the call chain to the Python/C++ boundary. In DeepEP v1.2.1, the version
  7. execution time, from v0 to v1. The agent's division optimization then reduced v1's execution
  8. threefold throughput, and 100,000 accelerators, tell us that progress at this boundary will

UBS-爱尔眼科-300015-2026瑞银A股研讨会：短期经营边际改善；海外并购与高端术式升级构筑增长支撑-20260902.pdf

  1. 1IQGVJo4rXcUdX6MbPbRnPqQmOoQeRmMyRfQrRnRaQoPoQMYpMrRvPnPmQ

UBS-春秋航空-601021-2026瑞银A股研讨会：Q3国内线收益率承压-20260902.pdf

  1. 4FPFWGm2oUeWdX8ObP8OpNrRmOtNlOmMzQiNmMnR9PoPpRMYoNmMvPtRtR

UBS-德福科技-301511-快评：2026瑞银A股研讨会，HVLP加速放量，锂电箔盈利持续修复-20260902.pdf

  1. AHzBXHp5sY8YcWaQ8QaQtRnNmOoQeRpPyRlOoOnR9PmNoQwMoNqQNZpNtO

UBS-鼎胜新材-603876-H126业绩符合预期，钠电产品将按预期Q4出货-20260831.pdf

  1. 0JOEUIjBsYfXbZaQbP9PsQpPpNpRlOpPzQiNoOmP8OqQuNuOpNnPNZmQoP

UBS-恩捷股份-002812-2026瑞银A股研讨会：盈利改善趋势延续，行业供需格局有望优化-20260903.pdf

  1. 企业价值/息税折旧摊销前利润(UBS核心) x 20.4 33.3 16.0 11.6 8.8 7.4 6.0 5.9
  2. 户的风险情况、投资重心和投资角度（例如是覆盖整个市场还是专注于某个行业，是长期还是短期等），客户与瑞银全球研究和UBS Evidence Lab的整体
  3. 若您对特定公司的权证执行价格感兴趣，请发邮件至UBS Securities LLC, 11 Madison Avenue, New York, NY 10010, USA, Attention:
  4. （https://www.ubs.com/global/en/legalinfo2/privacy.html）和cookie通告（https://www.ubs.com/global/en/legal/privacy/users.html）处理其个人数据并使
  5. 接收本文件即表明接收方了解并同意上述预期目的，而且不会预计或相信这些信息构成对接收方的投资建议或旨
  6. 做出投资或财务决定前，本文件或这些信息的所有接收方都应采取措施来了解投资的风险和回报并听取其个人财务、法律、税务和其他专业顾问考虑到接
  7. 任在美国《证券交易法》（Securities Exchange Act）15B（即Municipal Advisor Rule）下定义的市政实体(Municipal Entity)的市政顾问(Municipal Advisor)
  8. PT UBS Sekuritas Indonesia的注册办公地址为Sequis Tower Level 22 unit 22-1, Jl. Jend. Sudirman, kav.71, SCBD lot 11B, Jakarta 12190, Indonesia，

PB：中国通胀回升受供给端带动 央行降息仍不可少-20260909.pdf

  1. This document is being provided for the exclusive use of MENG WANG at CHINA CITIC BANK. Not for redistribution.
  2. 温影响，能源价格⾛⾼，⼈⼯智能领域的⼤规模投资则推升了电⼦产品价格。
  3. This report may not be modified or altered in any way. The BLOOMBERG PROFESSIONAL service and BLOOMBERG Data are owned and distributed locally by Bloomberg Finance LP ("BFLP") and its subsidiaries in
  4. all jurisdictions other than Argentina, Bermuda, China, India, Japan and Korea (the ("BFLP Countries"). BFLP is a wholly-owned subsidiary of Bloomberg LP ("BLP"). BLP provides BFLP with all the global marketing
  5. and operational support and service for the Services and distributes the Services either directly or through a non-BFLP subsidiary in the BLP Countries. BFLP, BLP and their affiliates do not provide investment
  6. This report may not be modified or altered in any way. The BLOOMBERG PROFESSIONAL service and BLOOMBERG Data are owned and distributed locally by Bloomberg Finance LP ("BFLP") and its subsidiaries in
  7. all jurisdictions other than Argentina, Bermuda, China, India, Japan and Korea (the ("BFLP Countries"). BFLP is a wholly-owned subsidiary of Bloomberg LP ("BLP"). BLP provides BFLP with all the global marketing
  8. and operational support and service for the Services and distributes the Services either directly or through a non-BFLP subsidiary in the BLP Countries. BFLP, BLP and their affiliates do not provide investment

SemiAnalysis-Engram架构与DRAM-SSD卸载协同设计-20260918.pdf

  1. official InferenceX agentic inference serving results on Engram models like DeepSeekv4.1
  2. Removing Engram raised answer loss from 0.2848 to 0.3093 bits/token. Forcing the ablated
  3. model to use the original Engram-on expert choices made it worse still, at 0.3375 bits/token.
  4. Engram’s table is large, but each lookup is small. DeepSeek-V4.1-Flash requests 24 rows at
  5. Eventually when they did publicly release for “day 0” image support, performance-wise, it is
  6. currently up to 14.8x worse perf per dollar than H200 and up to 42x worse perf per dollar
  7. P90 interactivity. Near 125 tokens/s/user, DRAM delivers 121 million total tokens per dollar
  8. point has a DRAM alternative that delivers higher P90 interactivity and more total tokens

Semianalysis-内存领域-为何4层堆HBM胜出.pdf

  1. which results in 16 GB, 32 GB, and 48 GB for 4-hi, 8-hi, and 12-hi stacks, respectively. An
  2. even though 4-hi can deliver lower concurrency, the faster token/s/user wholly or partially
  3. throughput than 8-hi HBM. Then when the interactivity is doubled further to 213 tok/s/user,
  4. that 4-hi can deliver allows for more than double the interactivity. We also reiterate that in
  5. a present workload on a future system to show that 4-hi is sufficient. It will only be a year
  6. Second, if 4-hi accelerators become a significant part of the AI server accelerator installed
  7. design, and their calling for 4-hi reflects that future models are heading in a direction that
  8. get the same bandwidth from a 4-hi stack as an 8-hi or 12-hi stack, in simple terms the 4-hi

文档

Toward Recursive Self-Improvement- How GLM Built Its Own Inference Infrastructur.pdf

总结

  1. In October 2025, we began researching how to strengthen its cybersecurity capabilities. Our
  2. production-grade inference service from scratch on a cluster of more than 100,000 Chinese-
  3. ultimately tripling end-to-end throughput relative to the initial baseline. Figure 1 shows the
  4. performance trajectory of GLM-5.3-Flash from its first successful run to production launch.
  5. test failed," "TTFT increased by 30%," or "output throughput dropped by 20%" after a change
  6. Transfer, following the call chain to the Python/C++ boundary. In DeepEP v1.2.1, the version
  7. execution time, from v0 to v1. The agent's division optimization then reduced v1's execution
  8. threefold throughput, and 100,000 accelerators, tell us that progress at this boundary will

预览

Toward Recursive Self-Improvement- How GLM Built Its Own Inference Infrastructur.pdf

UBS-爱尔眼科-300015-2026瑞银A股研讨会：短期经营边际改善；海外并购与高端术式升级构筑增长支撑-20260902.pdf

总结

  1. 1IQGVJo4rXcUdX6MbPbRnPqQmOoQeRmMyRfQrRnRaQoPoQMYpMrRvPnPmQ

预览

UBS-爱尔眼科-300015-2026瑞银A股研讨会：短期经营边际改善；海外并购与高端术式升级构筑增长支撑-20260902.pdf

UBS-春秋航空-601021-2026瑞银A股研讨会：Q3国内线收益率承压-20260902.pdf

总结

  1. 4FPFWGm2oUeWdX8ObP8OpNrRmOtNlOmMzQiNmMnR9PoPpRMYoNmMvPtRtR

预览

UBS-春秋航空-601021-2026瑞银A股研讨会：Q3国内线收益率承压-20260902.pdf

UBS-德福科技-301511-快评：2026瑞银A股研讨会，HVLP加速放量，锂电箔盈利持续修复-20260902.pdf

总结

  1. AHzBXHp5sY8YcWaQ8QaQtRnNmOoQeRpPyRlOoOnR9PmNoQwMoNqQNZpNtO

预览

UBS-德福科技-301511-快评：2026瑞银A股研讨会，HVLP加速放量，锂电箔盈利持续修复-20260902.pdf

UBS-鼎胜新材-603876-H126业绩符合预期，钠电产品将按预期Q4出货-20260831.pdf

总结

  1. 0JOEUIjBsYfXbZaQbP9PsQpPpNpRlOpPzQiNoOmP8OqQuNuOpNnPNZmQoP

预览

UBS-鼎胜新材-603876-H126业绩符合预期，钠电产品将按预期Q4出货-20260831.pdf

UBS-恩捷股份-002812-2026瑞银A股研讨会：盈利改善趋势延续，行业供需格局有望优化-20260903.pdf

总结

  1. 企业价值/息税折旧摊销前利润(UBS核心) x 20.4 33.3 16.0 11.6 8.8 7.4 6.0 5.9
  2. 户的风险情况、投资重心和投资角度（例如是覆盖整个市场还是专注于某个行业，是长期还是短期等），客户与瑞银全球研究和UBS Evidence Lab的整体
  3. 若您对特定公司的权证执行价格感兴趣，请发邮件至UBS Securities LLC, 11 Madison Avenue, New York, NY 10010, USA, Attention:
  4. （https://www.ubs.com/global/en/legalinfo2/privacy.html）和cookie通告（https://www.ubs.com/global/en/legal/privacy/users.html）处理其个人数据并使
  5. 接收本文件即表明接收方了解并同意上述预期目的，而且不会预计或相信这些信息构成对接收方的投资建议或旨
  6. 做出投资或财务决定前，本文件或这些信息的所有接收方都应采取措施来了解投资的风险和回报并听取其个人财务、法律、税务和其他专业顾问考虑到接
  7. 任在美国《证券交易法》（Securities Exchange Act）15B（即Municipal Advisor Rule）下定义的市政实体(Municipal Entity)的市政顾问(Municipal Advisor)
  8. PT UBS Sekuritas Indonesia的注册办公地址为Sequis Tower Level 22 unit 22-1, Jl. Jend. Sudirman, kav.71, SCBD lot 11B, Jakarta 12190, Indonesia，

预览

UBS-恩捷股份-002812-2026瑞银A股研讨会：盈利改善趋势延续，行业供需格局有望优化-20260903.pdf

PB：中国通胀回升受供给端带动 央行降息仍不可少-20260909.pdf

总结

  1. This document is being provided for the exclusive use of MENG WANG at CHINA CITIC BANK. Not for redistribution.
  2. 温影响，能源价格⾛⾼，⼈⼯智能领域的⼤规模投资则推升了电⼦产品价格。
  3. This report may not be modified or altered in any way. The BLOOMBERG PROFESSIONAL service and BLOOMBERG Data are owned and distributed locally by Bloomberg Finance LP ("BFLP") and its subsidiaries in
  4. all jurisdictions other than Argentina, Bermuda, China, India, Japan and Korea (the ("BFLP Countries"). BFLP is a wholly-owned subsidiary of Bloomberg LP ("BLP"). BLP provides BFLP with all the global marketing
  5. and operational support and service for the Services and distributes the Services either directly or through a non-BFLP subsidiary in the BLP Countries. BFLP, BLP and their affiliates do not provide investment
  6. This report may not be modified or altered in any way. The BLOOMBERG PROFESSIONAL service and BLOOMBERG Data are owned and distributed locally by Bloomberg Finance LP ("BFLP") and its subsidiaries in
  7. all jurisdictions other than Argentina, Bermuda, China, India, Japan and Korea (the ("BFLP Countries"). BFLP is a wholly-owned subsidiary of Bloomberg LP ("BLP"). BLP provides BFLP with all the global marketing
  8. and operational support and service for the Services and distributes the Services either directly or through a non-BFLP subsidiary in the BLP Countries. BFLP, BLP and their affiliates do not provide investment

预览

PB：中国通胀回升受供给端带动 央行降息仍不可少-20260909.pdf

SemiAnalysis-Engram架构与DRAM-SSD卸载协同设计-20260918.pdf

总结

  1. official InferenceX agentic inference serving results on Engram models like DeepSeekv4.1
  2. Removing Engram raised answer loss from 0.2848 to 0.3093 bits/token. Forcing the ablated
  3. model to use the original Engram-on expert choices made it worse still, at 0.3375 bits/token.
  4. Engram’s table is large, but each lookup is small. DeepSeek-V4.1-Flash requests 24 rows at
  5. Eventually when they did publicly release for “day 0” image support, performance-wise, it is
  6. currently up to 14.8x worse perf per dollar than H200 and up to 42x worse perf per dollar
  7. P90 interactivity. Near 125 tokens/s/user, DRAM delivers 121 million total tokens per dollar
  8. point has a DRAM alternative that delivers higher P90 interactivity and more total tokens

预览

SemiAnalysis-Engram架构与DRAM-SSD卸载协同设计-20260918.pdf

Semianalysis-内存领域-为何4层堆HBM胜出.pdf

总结

  1. which results in 16 GB, 32 GB, and 48 GB for 4-hi, 8-hi, and 12-hi stacks, respectively. An
  2. even though 4-hi can deliver lower concurrency, the faster token/s/user wholly or partially
  3. throughput than 8-hi HBM. Then when the interactivity is doubled further to 213 tok/s/user,
  4. that 4-hi can deliver allows for more than double the interactivity. We also reiterate that in
  5. a present workload on a future system to show that 4-hi is sufficient. It will only be a year
  6. Second, if 4-hi accelerators become a significant part of the AI server accelerator installed
  7. design, and their calling for 4-hi reflects that future models are heading in a direction that
  8. get the same bandwidth from a 4-hi stack as an 8-hi or 12-hi stack, in simple terms the 4-hi

预览

Semianalysis-内存领域-为何4层堆HBM胜出.pdf

源文件：2026-09-21/topics/508.md
站内副本：2026-09-21/topics/508.md