Reported by TechPowerUp on 2026-10-09. Summarized and analyzed by XAIPART Editorial.
What happened
At the Linux Plumbers Conference (LPC) 2026 in Prague on , AMD engineer Bharata Bhasker Rao presented the IBS Memory Profiler, a hardware feature of the upcoming Zen 6 processorsS1.
Instruction-Based Sampling (IBS) is AMD's existing profiling hardware, and Zen 6 adds a separate, lighter instance that only samples data memory accessesS1. Each sample delivers the kernel the physical address of the access and indicates whether the data came from DRAM or external memory such as CXLS1. This is relevant on tiered-memory servers, where Linux must locate frequently used ("hot") pages on slower CXL memory and migrate them to DRAMS1. AMD's EPYC 9006 "Venice" chips support memory expansion via CXL 3.1S1.
Linux currently identifies those pages by scanning page tables and trapping page faults, and AMD proposes feeding the profiler's samples into pghot, an in-development kernel subsystem for hot page trackingS1. On a test system with 256 logical CPUs and 256 GB of CXL memory, XSBench ran at 4.57x over a no-tiering baseline versus 1.18x for the current method, and Graph500 hit 3.17x versus 2.34xS1. Redis showed no gain, and NAS BT landed slightly behind; per-folio promotion cost fell from 26.1 to 15.8 microseconds in an overhead test, though that benchmark finished about 5% slowerS1. These are AMD's own figures, the slides list no virtualization support for the profiler, and EPYC Venice is in production with cloud deployments expected later this yearS1.
| Metric | Reported value | Source |
|---|---|---|
| LPC 2026 presentation date | 2026-10-05, Prague | [S1] |
| Test system | 256 logical CPUs, 256 GB CXL memory | [S1] |
| XSBench speedup (pghot vs current) | 4.57x vs 1.18x over no-tiering | [S1] |
| Graph500 speedup (pghot vs current) | 3.17x vs 2.34x | [S1] |
| Per-folio promotion cost | 26.1 to 15.8 microseconds | [S1] |
| Overhead benchmark delta | about 5% slower | [S1] |
| EPYC Venice CXL support | CXL 3.1 | [S1] |
| Virtualization support for profiler | Not disclosed (slides list none) | [S1] |
Background timeline
- - AMD engineer Bharata Bhasker Rao presents the Zen 6 IBS Memory Profiler at LPC 2026 in Prague [S1].
- - TechPowerUp publishes its summary of the LPC talk [S1].
- No prior milestones disclosed by the source.
XAIPART take
Technical impact
The Zen 6 IBS Memory Profiler is a hardware sampler, not an OS feature, so designers evaluating EPYC 9006 "Venice" platforms must treat it as silicon-tied capability: any benefit only appears on Zen 6 parts and depends on a kernel that consumes the samplesS1. The lack of disclosed virtualization support means hypervisor-mediated tiering on AMD EPYC Venice is unlikely to use this path soon, pushing VM workloads back to software-only page scanning or to alternatives such as Intel's MLDOM/2D-style VM memory tagging. XAIPART analysis
Market impact
For procurement, the immediate effect is on CXL 3.1 tiered-memory configurations built around EPYC Venice: buyers should expect first-generation pghot integration to be uneven, since AMD's own numbers show workload-dependent results from a 4.57x XSBench gain down to no gain on RedisS1. Cloud deployments are expected later this year, so lead-time pressure on EPYC 9006 "Venice" parts and CXL 3.1 expansion modules is plausible but not yet quantified by the source. XAIPART analysis
Action
Engineers designing tiered-memory EPYC Venice servers should plan for a mixed-outcome rollout: validate pghot-driven promotion on representative workloads (HPC, key-value, NAS, graph) before committing to large CXL 3.1 memory-expansion purchases, and budget for kernel-side work since the pghot patches are still in developmentS1. Track upstream pghot patch status and confirm with AMD whether virtualization support will arrive in a follow-up silicon or firmware revision. XAIPART analysis
XAIPART Forecast
XAIPART expects that by , if upstream pghot patches for the Zen 6 IBS Memory Profiler are not merged into a Linux kernel release that EPYC Venice cloud deployments can run, the headline 4.57x XSBench gain reported on AMD's 256 logical CPU / 256 GB CXL test system will not translate into broadly available production speedups for tiered-memory cloud servers. XAIPART analysis
Facts are as reported by the cited source. The design-in, sourcing and forward-looking statements above are XAIPART's analysis, not the source's claims. Uncertainty: all benchmark figures are AMD's own, virtualization support is absent from the disclosed slides, and the source gives no timeline for upstream Linux kernel merging of pghot or for AMD disclosing VM-mode operation.
Corrections: none at publication. This brief will be updated with a timestamp if the source corrects its report.
대체품 및 수명 주기
| 부품 | 관계 | 수명 주기 | 근거 |
|---|---|---|---|
| — | 등록된 대체품 없음 데이터 필요 | — | 요구 사항 match_type 어떤 전임자 라벨이 표시되기 전에 |
"전임자" 및 "대체품" 라벨은 기록된 일치 유형, 수명 주기 상태 또는 출처 텍스트에서만 부여됩니다. 일치가 기록되지 않은 경우 라벨이 표시되지 않습니다. 당사 자체 등록 상품은 시장 기준이 아닙니다.
경쟁 구도
- 21763 XAIPART의 이 카테고리 활성 등록 상품
- 88 등록된 제조사
- 57 최근 30일간 뉴스 아카이브에 언급된 제조사
수치는 XAIPART 카탈로그만 반영하며 시장 점유율이 아닙니다.
영향받는 부품
카탈로그 일치 없음. 이 보도는 현재 XAIPART에 등록된 어떤 부품과도 매핑되지 않습니다. 일치를 추론하지 않으며 자리 표시자 목록도 표시하지 않습니다.
이 내용의 맥락 위치
- amd-epyc-9006-venice-700-8-core-to-14-904-256-core-what-sp7-sp8-buyers-need
- astera-labs-broadens-aries-7-pcie-7-connectivity-line-for-multi-rack-ai-clusters
- what-amd-s-zen-6-ibs-memory-profiler-means-for-cxl-tiered-memory-server-designs
- amd-gpu-price-hikes-of-up-to-104-hit-china-first-international-board-partners-already-notified
- ibuypower-s-amd-ai-halo-mini-pc-what-the-ryzen-ai-max-pro-495-means-for-compact-workstation-sourcing
- amd-ryzen-ai-7-hx-455-leak-signals-a-cut-back-hx-470-with-z2-extreme-class-radeon-890m-igpu
- AMD가 모든 31개 Zen 6 EPYC 9006 SKU의 1Ku 가격을 공개했습니다. SP7은 2026년 4분기 출하, SP8은 2027년 상반기 출하 예정이며, TDP는 Default CPU Power로 대체됩니다. trade
- Astera Labs가 Aries 7 포트폴리오에 리타이머, 리드라이버, 케이블 모듈, 광 드라이버 및 TA를 추가하며, OCP Global Summit 데모에 앞서 2026년 4분기 샘플링이 예정되어 있습니다. product launch trade
- AMD disclosed a Zen 6 hardware sampler aimed at CXL tiered-memory hot-page tracking; benchmark deltas are mixed and patches remain in development. trade
- AMD has raised GPU vendor pricing in China by up to $104 on the RX 9070, citing memory and manufacturing costs. International increases are expected to follow. price change trade
- iBUYPOWER is shipping a Ryzen AI Max+ PRO 495-based mini workstation from $3999 in Q4 2026. Read what changes for design-in and procurement. product launch trade
- A Weibo leak surfaces an 8-core Ryzen AI 7 HX 455 APU with 16 CU Radeon graphics matching the Radeon 890M and 24 MB of L3 cache, pointing at a thinner HX 470 sibling. trade
- Colorful's E-ATX X870E Ultra-S V14 AM5 board ships in China at 1,899 CNY with 16+2+1 80 A VRMs, five M.2 slots and Wi-Fi 7. product launch trade
- A leaked OpenBenchmarking entry for an AMD Gainsborough engineering sample shows 6 cores, 24 GB LPDDR5 at 9600 MT/s, and Furmark parity with today's Steam Deck at lower resolution. trade
체인은 게시된 브리프로만 구성됩니다 — 브리프가 없는 노드는 그려지지 않습니다.
차원별 비용 압력
| Dimension | FRED series ID | Unit | Lag | As of |
|---|---|---|---|---|
| Input cost | PCOPPUSDM, PALUMUSDM, PNICKUSDM | USD/mt | 96 d | |
| Supply/demand | PCU334413334413, IPG3344S | index(1982=100) / index(2017=100) | 46 d | |
| FX & energy | DEXCHUS, DCOILBRENTEU, DHHNGSP | rate / USD/bbl / USD/MMBtu | 6 d |
No free authoritative interface for: gold · silver · rare earth · polysilicon · lithium carbonate · market share — listed rather than approximated.
이 콘텐츠의 제작 방식
프로세스 증거 — 수집, 출처 온전성, 모델 및 게이트
| Source fetched | TechPowerUp · www.techpowerup.com · 2,834 chars · fetch status success |
|---|---|
| Source sanity (S1) | ok — passed before any model call (no s1_ block recorded) |
| Model / prompt | MiniMax-M3 · prompt Blog Agent - News Summary (v2.8-final) · 6,966 tokens · 21.5s · run completed |
| Gate | verdict hold · 6/6 applicable checks passed · spec v2.9 · C10 ratio 0.113788 · C10 rounds 0 |
| Time contract | 4 <time datetime> values in this brief — 0 violations (+08:00 or date-only) |
위의 모든 행은 파이프라인에서 생성되며 수작업으로 작성되지 않습니다.
댓글
첫 번째 댓글을 남겨 보십시오.
🔐 먼저 로그인 댓글을 게시하려면
계정이 없으십니까? 회원가입