How The Ubuntu Linux Performance Has Evolved For SiFive RISC-V Over The Last Four Years

SiFive recently sent over their new HiFive Premier P550 developer board and as part of that fresh RISC-V CPU testing I've also been re-testing the prior SiFive HiFive Unmatched developer board from 2020~2021 for reference. Out of curiosity, I've carried out some tests using the HiFive Unmatched to look at how the Ubuntu Linus RISC-V performance has evolved from Ubuntu 21.04 when I first tested that RISC-V quad-core developer board, Ubuntu 22.04 LTS, and now the latest Ubuntu 24.04 LTS packages.

SiFive HiFive Unmatched

In advance of the HiFive Premier P550 review in the coming days, this article today is looking at that historical RISC-V Linux performance perspective across Ubuntu 21.04 / 22.04 LTS / 24.04 LTS on that mature HiFive Unmatched RISC-V developer board. As a reminder the HiFive Unmatched features four RISC-V 64-bit cores, 16GB of RAM, and was running with a Samsung 980 NVMe SSD and a Radeon HD 6770 graphics card.

SiFive HiFive Unmatched desktop

The SiFive Freedom U740 SoC is what powers the HiFive Unmatched developer board.

SiFive HiFive with RISC-V SoC

Ubuntu 21.04 as a reminder was on Linux 5.11 and using GCC 10.3, Ubuntu 22.04 moved to Linux 5.15 and GCC 11.2 and other software updates, and now Ubuntu 24.04 LTS is on Linux 6.8 with GCC 13.3 and other updated software packages. The same hardware platform was used for all testing with simply looking at how the Ubuntu Linux performance has evolved for this SiFive RISC-V platform going back to Ubuntu 21.04 from early 2021.

Ubuntu 24.04 SiFive RISC-V HiFive Unmatched Benchmarks

Following those comparison points, at the end of the article are the Ubuntu 24.04 LTS benchmarks on the HiFive Unmatched up against the Raspberry Pi 400 and Raspberry Pi 500 ARM systems for putting it into perspective how that mature RISC-V platform now compares. Again, in the next week or so will be our review and benchmarks of their new HiFive Premier P550 developer board.

SiFive HiFive Unmatched and HiFive Premier P550

Coremark benchmark with settings of CoreMark Size 666, Iterations Per Second. Ubuntu 24.04 LTS was the fastest.

The SiFive RISC-V performance on the HiFive Unmatched developer board has continued evolving with newer versions of Ubuntu Linux thanks to the updated kernel, GCC compiler, and other software featuring better RISC-V architecture support.

OpenSSL benchmark with settings of Algorithm: SHA256. Ubuntu 22.04 LTS was the fastest.
OpenSSL benchmark with settings of Algorithm: RSA4096. Ubuntu 24.04 LTS was the fastest.
OpenSSL benchmark with settings of Algorithm: RSA4096. Ubuntu 24.04 LTS was the fastest.

For some workloads the bulk of the software optimization uplift came from Ubuntu 22.04 LTS with little gain in some areas on Ubuntu 24.04.

Stress-NG benchmark with settings of Test: Atomic. Ubuntu 24.04 LTS was the fastest.
Stress-NG benchmark with settings of Test: Crypto. Ubuntu 24.04 LTS was the fastest.
Stress-NG benchmark with settings of Test: Malloc. Ubuntu 22.04 LTS was the fastest.
Stress-NG benchmark with settings of Test: CPU Cache. Ubuntu 24.04 LTS was the fastest.
Stress-NG benchmark with settings of Test: CPU Stress. Ubuntu 22.04 LTS was the fastest.
Stress-NG benchmark with settings of Test: Semaphores. Ubuntu 24.04 LTS was the fastest.
Stress-NG benchmark with settings of Test: Vector Math. Ubuntu 24.04 LTS was the fastest.
Stress-NG benchmark with settings of Test: Context Switching. Ubuntu 22.04 LTS was the fastest.
Stress-NG benchmark with settings of Test: System V Message Passing. Ubuntu 24.04 LTS was the fastest.

There have been a number of RISC-V Linux kernel optimizations to land over the past four years.

Monkey Audio Encoding benchmark with settings of WAV To APE. Ubuntu 24.04 LTS was the fastest.
FLAC Audio Encoding benchmark with settings of WAV To FLAC. Ubuntu 24.04 LTS was the fastest.
WavPack Audio Encoding benchmark with settings of WAV To WavPack. Ubuntu 24.04 LTS was the fastest.
Smallpt benchmark with settings of Global Illumination Renderer; 128 Samples. Ubuntu 24.04 LTS was the fastest.

The newer GCC (and LLVM/Clang) compiler releases can also help in producing more efficient RISC-V binaries.

eSpeak-NG Speech Engine benchmark with settings of Text-To-Speech Synthesis. Ubuntu 24.04 LTS was the fastest.
SecureMark benchmark with settings of Benchmark: SecureMark-TLS. Ubuntu 24.04 LTS was the fastest.
Dolfyn benchmark with settings of Computational Fluid Dynamics. Ubuntu 24.04 LTS was the fastest.
WebP Image Encode benchmark with settings of Encode Settings: Default. Ubuntu 22.04 LTS was the fastest.
WebP Image Encode benchmark with settings of Encode Settings: Quality 100. Ubuntu 22.04 LTS was the fastest.
WebP Image Encode benchmark with settings of Encode Settings: Quality 100, Highest Compression. Ubuntu 22.04 LTS was the fastest.

In only a few areas was Ubuntu 24.04 LTS running slower than Ubuntu 22.04 LTS on this SiFive HiFive Unmatched RISC-V developer board.

LZ4 Compression benchmark with settings of Compression Level: 1, Compression Speed. Ubuntu 24.04 LTS was the fastest.
LZ4 Compression benchmark with settings of Compression Level: 1, Decompression Speed. Ubuntu 24.04 LTS was the fastest.
LZ4 Compression benchmark with settings of Compression Level: 3, Decompression Speed. Ubuntu 24.04 LTS was the fastest.
LZ4 Compression benchmark with settings of Compression Level: 9, Decompression Speed. Ubuntu 24.04 LTS was the fastest.
Zstd Compression benchmark with settings of Compression Level: 3, Compression Speed. Ubuntu 24.04 LTS was the fastest.
Zstd Compression benchmark with settings of Compression Level: 19, Compression Speed. Ubuntu 22.04 LTS was the fastest.
Zstd Compression benchmark with settings of Compression Level: 19, Decompression Speed. Ubuntu 22.04 LTS was the fastest.
GraphicsMagick benchmark with settings of Operation: Swirl. Ubuntu 22.04 LTS was the fastest.
7-Zip Compression benchmark with settings of Compress Speed Test. Ubuntu 22.04 LTS was the fastest.
PyBench benchmark with settings of Total For Average Test Times. Ubuntu 24.04 LTS was the fastest.
PHPBench benchmark with settings of PHP Benchmark Suite. Ubuntu 24.04 LTS was the fastest.

There tended to be some nice performance improvements across a variety of workloads with succeeding Ubuntu Linux releases on this SiFive RISC-V developer board.

Geometric Mean Of All Test Results benchmark with settings of Result Composite, Ubuntu 24.04 SiFive RISC-V HiFive Unmatched Benchmarks. Ubuntu 24.04 LTS was the fastest.

Going from Ubuntu 21.04 to 22.04 LTS was around 19% better performance while with Ubuntu 24.04 LTS it's up to 23% faster than where the SiFive HiFive Unmatched RISC-V with Freedom U740 SoC started out four years ago. All the benchmarks conducted can be found here.

HiFive Unmatched vs. Raspberry Pi 500 vs. Raspberry Pi 400 Benchmarks

While the HiFive Unleashed is nearly a quarter faster with the newest Ubuntu Linux LTS release than where it started out when the RISC-V developer board was first shipping, software optimizations alone aren't enough. The HiFive Unmatched with its aging RISC-V SoC still significantly trails ARM performance as measured with the Raspberry Pi 400 and Raspberry Pi 500 ARM keyboard computers... It will be interesting to see where the HiFive Premier P550 fits into the picture in the days ahead.

libavif avifenc benchmark with settings of Encoder Speed: 6. Raspberry Pi 500 was the fastest.
C-Blosc benchmark with settings of Test: blosclz shuffle, Buffer Size: 8MB. Raspberry Pi 500 was the fastest.
Timed FFmpeg Compilation benchmark with settings of Time To Compile. Raspberry Pi 500 was the fastest.
Timed PHP Compilation benchmark with settings of Time To Compile. Raspberry Pi 500 was the fastest.
BYTE Unix Benchmark benchmark with settings of Computational Test: Dhrystone 2. Raspberry Pi 500 was the fastest.
BYTE Unix Benchmark benchmark with settings of Computational Test: Whetstone Double. Raspberry Pi 500 was the fastest.
C-Ray benchmark with settings of Resolution: 1080p, Rays Per Pixel: 16. Raspberry Pi 500 was the fastest.
C-Ray benchmark with settings of Resolution: 4K, Rays Per Pixel: 16. Raspberry Pi 500 was the fastest.
CacheBench benchmark with settings of Test: Read. Raspberry Pi 500 was the fastest.
CacheBench benchmark with settings of Test: Write. Raspberry Pi 500 was the fastest.
Apache Cassandra benchmark with settings of Test: Writes. Raspberry Pi 500 was the fastest.
7-Zip Compression benchmark with settings of Test: Compression Rating. Raspberry Pi 500 was the fastest.
DaCapo Benchmark benchmark with settings of Java Test: Avrora AVR Simulation Framework. Raspberry Pi 500 was the fastest.
DaCapo Benchmark benchmark with settings of Java Test: Batik SVG Toolkit. Raspberry Pi 500 was the fastest.
DaCapo Benchmark benchmark with settings of Java Test: PMD Source Code Analyzer. Raspberry Pi 500 was the fastest.

Even on Ubuntu 24.04 LTS, the HiFive Unmatched aging RISC-V board is much slower than the ARM Cortex performance found with recent Raspberry Pi single board computers.

DaCapo Benchmark benchmark with settings of Java Test: Apache Lucene Search Index. Raspberry Pi 500 was the fastest.
DaCapo Benchmark benchmark with settings of Java Test: H2 Database Engine. Raspberry Pi 500 was the fastest.
Google Draco benchmark with settings of Model: Church Facade. Raspberry Pi 500 was the fastest.
Opus Codec Encoding benchmark with settings of WAV To Opus Encode. Raspberry Pi 500 was the fastest.
WavPack Audio Encoding benchmark with settings of WAV To WavPack. Raspberry Pi 500 was the fastest.
GraphicsMagick benchmark with settings of Operation: Noise-Gaussian. Raspberry Pi 500 was the fastest.
GraphicsMagick benchmark with settings of Operation: Resizing. Raspberry Pi 500 was the fastest.
GROMACS benchmark with settings of Implementation: MPI CPU, Input: water_GMX50_bare. Raspberry Pi 500 was the fastest.
Java SciMark benchmark with settings of Computational Test: Monte Carlo. Raspberry Pi 500 was the fastest.
PyPerformance benchmark with settings of Benchmark: async_tree_io. Raspberry Pi 500 was the fastest.
PyPerformance benchmark with settings of Benchmark: pathlib. Raspberry Pi 500 was the fastest.
Renaissance benchmark with settings of Test: Akka Unbalanced Cobwebbed Tree. Raspberry Pi 500 was the fastest.
Renaissance benchmark with settings of Test: Finagle HTTP Requests. Raspberry Pi 500 was the fastest.
WebP Image Encode benchmark with settings of Encode Settings: Quality 100, Lossless. Raspberry Pi 500 was the fastest.

Those wanting to see more of these fresh HiFive Unmatched vs. Raspberry Pi benchmarks can find them here.

Geometric Mean Of All Test Results benchmark with settings of Result Composite, HiFive Unmatched vs. Raspberry Pi 500 vs. Raspberry Pi 400 Benchmarks. Raspberry Pi 500 was the fastest.

It will be interesting to show where the HiFive Premier P550 fits into the equation on Ubuntu 24.04 LTS with the upcoming review and benchmarks, stay tuned.

Read More

苹果 CEO 库克深夜官宣 Air 新品本周见,有望为 M4 MacBook Air / iPad Air

IT之家 3 月 4 日消息,苹果 CEO 库克在 X 平台发文,宣布本周将迎来 Air 新品,参考先前古尔曼发文,这款新品最有可能是搭载 M4 芯片的 MacBook Air 机型,不过也有可能是搭载 M4 芯片的 iPad Air。

根据古尔曼此前的报道,M4 MacBook Air 的发布窗口期一直锁定在 3 月。随着 iPhone 16e 的发布告一段落,本周苹果的焦点将转向 Mac 系列产品。

新款 MacBook Air 将推出 13 英寸和 15 英寸两种尺寸,均搭载苹果已发布的 M4 芯片。该芯片配备 10 核 CPU 和 10 核 GPU。根据 Geekbench 6 的测试结果,M4 芯片在多核 CPU 性能方面比 M3 芯片快达 25%。去年 10 月,苹果已经为 MacBook Pro、iMac 和 Mac mini 更新了 M4 芯片。同时,苹果还将 MacBook Air 全系的最低 RAM 内存配置从 8GB 提升至 16GB。

IT之家获悉,新款 MacBook Air 有望在摄像头方面进行升级,配备 1200 万像素摄像头,并支持“居中显示”功能。该功能可在视频通话中自动追踪用户,使其始终保持在画面中央。此外,摄像头还将支持“桌面视图”,能够同时显示用户和桌面的俯视图。目前,“居中显示”功能已在 FaceTime 和 Zoom 等视频通话应用中可用,此前仅限于较新的 iPad 型号和连接苹果外接 Studio Display 的 Mac 机型。而最新款的 MacBook Pro 和 iMac 已经首次无需依赖 Studio Display 即可支持该功能。

新款 MacBook Air 机型预计将配备三个雷电 4 接口,而目前的机型仅配备两个雷电 3 接口。此外,新机型在开启盖子时仍可支持连接两台外部显示器。苹果最新款的 MacBook Pro 和 iMac 提供了纳米纹理显示屏配置选项,这种显示屏通过特殊涂层更好地散射光线,从而进一步减少屏幕反光。虽然尚不确定苹果是否会将这一配置选项扩展至新款 MacBook Air,但为了与 MacBook Pro 保持差异化,苹果可能会选择将其保留为 MacBook Pro 的专属功能。

目前尚未有消息称新款 MacBook Air 将在外观设计上有重大变化,苹果上一次对 MacBook Air 进行重新设计是在 2022 年。

Read More

三星Galaxy S25 Edge亮相MWC:只能看 禁止上手

三星Galaxy S25 Edge惊喜亮相MWC2025展会,三星在展会上陈列了数台Galaxy S25 Edge样机,但是禁止媒体上手体验,只允许拍照观察。该机主打超薄设计,它将对标9月份登场的iPhone 17 Air,后者也是一台超薄机型,目前三星未公布Galaxy S25 Edge的具体厚度,新品会在4月份正式发布。

三星Galaxy S25 Edge亮相MWC:只能看 禁止上手

据爆料,Galaxy S25 Edge采用6.7英寸中置挖孔直屏,搭载高通骁龙8至尊版for Galaxy移动平台,后置2亿像素主摄,电池是4000毫安时,其厚度可能是6.4mm,这是S25系列中最薄机型。

工业设计方面,与Galaxy S25相比,Galaxy S25 Edge最大不同在于其配备了双摄像头系统,且双摄和闪光灯都被椭圆形DECO包裹,这一独特设计提高了手机辨识度。

目前Galaxy S25 Edge已通过国家质量认证,认证信息显示,该机支持25W有线快充,与S25标准版相同。

三星Galaxy S25 Edge亮相MWC:只能看 禁止上手

三星Galaxy S25 Edge亮相MWC:只能看 禁止上手

三星Galaxy S25 Edge亮相MWC:只能看 禁止上手

Read More

Stability AI 与 Arm 合作优化音频生成模型 人工智能公司 Stability AI 与芯片制造商ARM合作,将 Stability 的 Stable Audio Open 引入运行 Arm 芯片的移动设备...

Stability AI 与 Arm 合作优化音频生成模型

人工智能公司 Stability AI 与芯片制造商ARM合作,将 Stability 的 Stable Audio Open 引入运行 Arm 芯片的移动设备。虽然许多人工智能应用可以生成音频,但大多数依赖于云处理,这意味着它们不能离线使用。此外部分音频生成模型是在受版权保护的内容上进行训练的,这构成了知识产权风险。Stability 声称 Stable Audio Open 的训练集完全由免版税的音频和歌曲组成。在 Arm 芯片上运行的 Stable Audio Open 将于本周在巴塞罗那举行的世界移动通信大会上进行演示,可以根据文本描述生成声音。在 Armv9 CPU 上生成单个 11 秒的音频样本大约需要 8 秒。

—— Techcrunch

Read More

据称美国将停止针对俄罗斯的进攻性网络行动

据报道,在特朗普政府努力向莫斯科让步以结束乌克兰战争之际,美国已暂停针对俄罗斯的进攻性网络行动。据《记录》报道,美国国防部长皮特-黑格塞斯(Pete Hegseth)授权下令停止美国针对俄罗斯发起的黑客行动。

新指南影响美国网络司令部(U.S. Cyber Command)开展的行动,该司令部是美国国防部的一个部门,主要负责网络空间的黑客攻击和行动,但不适用于国家安全局开展的间谍行动。

6076599594_2f1271780c_b.jpg

《纽约时报》和《华盛顿邮报》证实了这一命令。

报道称,该命令是在周五美国总统唐纳德-特朗普、副总统JD-万斯和乌克兰总统沃洛德梅尔-泽连斯基在椭圆形办公室举行会议之前下达的。《纽约时报》称,这一指示是吸引俄罗斯总统弗拉基米尔-普京就该国正在进行的乌克兰战争进行会谈的更广泛努力的一部分。

美国网络司令部和五角大楼对此没有回应,但一名高级国防官员告诉《记录》,该司令部不"讨论网络情报、计划或行动"。

《卫报》也报道称,特朗普政府已表示不再将俄罗斯黑客视为网络安全威胁,据说还命令美国网络安全机构 CISA 不再报告俄罗斯的威胁。 该报援引了一份最近的备忘录,其中列出了 CISA 的新优先事项,包括中国面临的威胁和保护本地系统,但该备忘录并未提及俄罗斯。 据报道,CISA 的员工接到口头通知,要求他们暂停任何有关俄罗斯网络威胁的工作。

负责监管CISA的国土安全部在一份声明中否认了这一报道。国土安全部助理部长特里西娅-麦克劳克林(Tricia McLaughlin)表示,这份备忘录并非来自特朗普政府,"CISA仍致力于应对美国关键基础设施面临的所有网络威胁,包括来自俄罗斯的威胁,我们在这方面的姿态和优先级没有任何变化。"

据报道,就在美国情报界称俄罗斯对美国构成"持久网络威胁"几个月后,特朗普政府的对俄政策发生了转变。 美国在其年度威胁评估中警告说,俄罗斯对外情报局(SVR)继续以美国政府机构和关键基础设施(如水下电缆和工业控制系统)为目标。

在过去几年中,美国政府也对大量俄罗斯网络犯罪分子采取了行动。 美国政府还对大量俄罗斯网络犯罪分子采取了行动:成功破坏了与俄罗斯有关联的勒索软件组织的基础设施,收回了支付给俄罗斯黑客的数百万美元赎金,对一些最臭名昭著的俄罗斯黑客进行了起诉和制裁。

Read More

[旅行 APP 产品诞生日记] 5day/100days

IterX:

Day5️⃣

进度不够,先用其他画虎烂的随笔凑下字数

今天看到 bluesky (原来 twitter 创始团队和工程师出来搞的一个去中心化社交网络项目)仅 13 人的团队,前端只有 1 个人,各端共用一套代码。要知道他们的服务是有31,951,353的用户(平均每秒增加 0.6 个用户),目前估值在 7 亿美元的水平

LLMs 时代改变了很多,最先席卷的就是开发者群体,现在基本大部分开发者都在开发过程中使用了 AI 辅助开发,或多或少都提高了效率并扩宽了能力边界,现在行业内很多例子都表明,小团队的效率得到了极大的提升,这是好消息,但是同时也有 Bad news 😟

数据表明,开发人员的工作机会在急剧收窄,我们看看 FRED 的数据(主要针对美国和全球的经济数据)

Xnip2025-02-21_10-41-19.png

基于 Indeed (一个招聘网站)的数据,美国本土找软件开发工程师的工作岗位在疫情后持续上涨(当时也是零利率时期),但是在 2022 年 5 月后就开始一路下降,目前一直保持在低位

现在很多公司包括 Big Tech 也收拢 HC (甚至裁员),一部分原因是联邦利率还没下来(今年大概率也没什么下调的空间),另外也是在消化疫情后那段时期招聘的 HC (对冲了一下),

Xnip2025-02-21_10-53-27.png

可以看到 FED 开始加息后,招聘岗位就跟着开始下降了(越大的公司越是资本游戏,在个体看来一点利率在大规模资金下差别很大)

对比一下总的工作岗位和软件开发工程师的对比

Xnip2025-02-21_11-59-41.png

可以看到是有一定差别的,总的工作机会走势比较平缓,下降也是降的少点。不过这一把交叉的区域拿出来,可以看出,正因为前面招聘多了,现在招聘少了,如果平衡一下,走势应该会接近一点

不过数据也是有一定的偏差的,比如这边是基于 Indeed 得出的,就好比现在 Boss 直聘给出一份数据报告就业情况,也只是基于其平台拥有的数据来评估,并不能完全代表大环境,尤其是一些以 AI 为主的 startup ,有比较大的可能没有覆盖在里面,不过以小见大,主要还是看一个趋势

不知道接下去几年会再发生什么大变革,个体和企业在其中总是在一定程度上的不可抗力,能做的只能提升自己,让自己更硬一点,跟上时代,期望别被时代丢下,被市场抛弃。仅此而已。回想为什么不断有初创崛起,可能也是每个世代都有人从这个环境长大,更能感受到这个世代的人需要什么,加上一下交叉技能和契机,就造就了新的 idea 。

今天和好几个人 1v1 聊了一下,有同行创业者,也有产品经理,交流交换了一下思想,也能感受到其他主创们的一些想法,有一些已经有产品的,能从产品中窥见一些背后的想法和意图,我觉得还蛮好的,就是和某些竞品聊会有一点怪,哈哈。还挺多人会提到护城河或者门槛这个东西的,我觉得这也合理,比较符合直觉,总会有人投身进来的,不过我还是保持比较乐观的心态,市场充满竞争本身也从一个侧面展现出潜力,不过旅行这块盘子有多大的市值还是看深耕的领域,目前比较多的盈利都是走携程这类 OTA 方向。看各家创始人各自的眼界和本领了,好产品和对的商业模式。不过我觉得不一定每份事业都是奔着大盘子去的,有一些小而美的产品和小而美的团队,也是现下挺多人追求的东西。期待后续更多大佬们不吝赐教😎

言归正传,今天做了主行程页的一部分,比较粗,还需要再打磨一下,有两种色调风格我比较喜欢的,一种是偏冷淡色系的,可以让人很平静,一种是色块对冲的,有一种很 Cool 的感觉。我正在考虑到时候可以让用户切换主题。

Xnip2025-02-22_00-36-46.png

Xnip2025-02-22_00-36-52.png

汇总目前情况:

  1. 原型设计:4%,有想法了,正和 2️⃣一起推进
  2. UI/UX 设计:4%,在做了在做了
  3. 后端( Go )开发:10%,手拿把掐,目前暂停这块工作,要先处理 2️⃣和 4️⃣
  4. 客户端( flutter )进度:0%,坐等 1️⃣&2️⃣完成

接下去几天会重点忙 1️⃣和 2️⃣

如果你认为你符合以下条件,欢迎来聊:

  1. 能坚持
  2. 有梦想
  3. 有兴趣

Read More