TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

Twenty Five Years of Warehouse-Scale Computing

62 点作者 matt_d11 个月前

3 条评论

jeffrallen11 个月前
&gt; in late 2017, ML cycles surpassed non-ML cycles in the fleet for the first time<p>Wow, I was not expecting that. TPUs were only introduced 2 years before.
jauntywundrkind11 个月前
Really nice having this all written down, having some chronology &amp; epochs laid out to talk about.<p>Couple sections after that stirred me a bit. They&#x27;re kind of framed a bit as the hardware challenges, as data-center challenges, but I feel like this all applies double to the software, to what glues the system together.<p>&gt; <i>Technology islands and industry ecosystems</i><p>&gt; <i>The initial success of WSCs was driven by being different, out of necessity reinventing many of the then-conventional approaches to system design. However, as WSCs have scaled, and their adoption has increased via public cloud providers, a broad industry ecosystem supports WSC use cases, allowing &quot;build vs buy&quot; decisions.</i><p>&gt; <i>Custom designs work best when they target some unique needs of WSC workloads or systems that are currently not satisfied cost-effectively by existing solutions in the market (for example, the design of TPU accelerators for WSC machine learning workloads). For more mature markets, however, volume economics often reduce costs and increase velocity, favoring products built on top of industry standards (example, server form factors). Focus on building modular, composable, and interoperable architectures built on standardized interfaces; without this focus on composability and standards, you may end up on a &quot;tech island&quot; unique to yourself because one custom component forces all others to be custom too. In many ways, this is the hardware equivalent of the monoliths-vs-microservices tradeoff.</i><p>And,<p>&gt; <i>Optimizing the time variable of Moore’s law: agility, modularity, and interoperability</i><p>&gt; <i>The traditional formulation of Moore’s law (performance doubles every two years for the same cost) typically focuses on three variables: performance, cost, and time. As performance and cost improvements start slowing down, focusing on the time variable — the velocity of hardware development — can be a good way to optimize the &quot;area-under-the curve&quot; for continued improvements. Incremental smaller benefits, but at more finer granularities, when compounded, can still achieve exponential benefits.</i><p>&gt; <i>To achieve such agile, faster improvements, we need to build more modular hardware platforms with appropriate investment in interfaces, standards, etc. Chiplets in particular allow us to co-design in a multi-die system context, allowing cost advantages from die geometries, but also mix-and-match integration across heterogeneous IP blocks and different process technologies.</i><p>&gt; <i>The emergence of open source hardware, is another particularly exciting development in this context and enables a more collaborative ecosystem that hardware designers can build on — open-source IP blocks (e.g., Caliptra root of trust), verification and testing suites (e.g., CHIPS alliance, OpenCompute), and even open source tools&#x2F;PDKs (e.g., OpenRoad). Given how profound open source software has been to WSCs, the opportunity to have a similar impact with open source hardware is significant.</i><p>The discussion on roofshotting, on mild 1.3X-2x improvements, done repeatedly, and revisiting &amp; reapplying old successes I think dovetails with the discussion of modularity. Finding patterns that are broad &amp; reappliable across domains is a huge win. Kubernetes for example keeps getting compared to Docker Compose. But docker compose is good for assembling a set of containers. Where-as Kubernetes is a set of management&#x2F;manufacturing patterns for that happen to include containers. There&#x27;s platform modularity by scoping your systems layer bigger, by reusing the wins.<p>I am very hopeful we see a compatibility of datacenters start to emerge. CXL as a very fast fabric interconnect is exciting. Ultra Ethernet Consortium borrowing some RDMA style wins is promising. Hopefully we see industry players arise &amp; serve this market, make a competitive and rich supply side ecosystem that data center builders can keep extracting value from. Right now the market feels early &amp; a boutique interest; getting chiplets and interconnects back to bread and butter of chip making would help drive innovation upwards.
pulse711 个月前
With more and more cores and RAM soon we will not need warehouse-scale anymore... everything will be on a single server again - for almost all needs... one can get servers with more than 10TB of RAM and soon we will have more than 1000 cores...
评论 #40911341 未加载
评论 #40910774 未加载
评论 #40910610 未加载
评论 #40911611 未加载
评论 #40911963 未加载