Amazon EC2 的 Trn1 正式開放使用

AWS 自家研發晶片的 trn1.* 上線了:「Amazon EC2 Trn1 Instances for High-Performance Model Training are Now Available」。

先前三家雲端的廠商只有 Google Cloud PlatformTPU 可以 train & evaluate,現在 AWS 推出 AWS Trainium,補上 train 這塊的產品。其中官方宣稱可以比 GPU 架構少 50% 的計算成本:

Trainium-based EC2 Trn1 instances solve this challenge by delivering faster time-to-train while offering up to 50% cost-to-train savings over comparable GPU-based instances.

然後 PyTorchTensorFlow 都有支援:

The Neuron plugin natively integrates with popular ML frameworks, such as PyTorch and TensorFlow.

另外用 neuron-ls 可以看到 Neuron 裝置的資訊,不過沒看懂為什麼要 mask 掉 private ip 的資訊:

大型的 cluster 會使用 Amazon FSx for Lustre 整合提供服務:

For large-scale model training, Trn1 instances integrate with Amazon FSx for Lustre high-performance storage and are deployed in EC2 UltraClusters. EC2 UltraClusters are hyperscale clusters interconnected with a non-blocking petabit-scale network.

但第一波開放的區域有點少,只有萬年美東一區 us-east-1 與美西二區 us-west-2

You can launch Trn1 instances today in the AWS US East (N. Virginia) and US West (Oregon) Regions as On-Demand, Reserved, and Spot Instances or as part of a Savings Plan.

us-east-1trn1.2xlarge 的價錢是 US$1.34375/hr,但沒有實際跑過比較好像沒辦法評估到底行不行...


Prerender 從 AWS 搬回傳統機房的成本節省

Hacker News 上看到「We reduced our server costs by moving away from AWS (」這篇,原文在「How we reduced our annual server costs by 80% — from $1M to $200k — by moving away from AWS」這邊。

偶而會看到這類的報導,這次是 Prerender 這家的服務,從本來在 AWS 上的 $1m/y 降到 $200k/y (這邊都是用美金在計算)。

但好像沒提到第一次投資購買硬體花了多少錢,不過就以前的經驗上來說,把每個月非人力的 OPEX 加上 CAPEX 的各種攤提,大概會是雲端的 1/3 到 1/2 的費用。

近年來 k8s 以及各種架構愈來愈完整,很多技術也都收斂,慢慢變成業界標準了,不需要自己土砲搞一堆東西而導致產生可觀的維護成本。 上面搜尋 可以看到 Prerender 選擇的工具,像是在地端上面應該是用 k8s,然後有用 Sentry 以及 Redash

這個跟當年算過的經驗類似,startup 可以先上雲端把服務做起來,雲端的擴充性對於初期的幫助會很大 (先 scale up 撐住,再改寫成可以 scale out 的版本),而當量又大到一定程度時就反過來先問 discount,如果不滿意的話就可以規劃搬回地端。

Prerender 的量看起來超過臨界值不少,搬到地端省成本應該是蠻理所當然的...

Kagi 公佈了收費三個月後的進展

Kagi 公佈了收費三個月後的進展 (可以參考「Kagi 開始收費了」這篇):「Kagi status update: First three months」。

搜尋的部份 (Kagi 這個產品線),目前有 2600 個付費使用者,以 US$10/mo 的費用來算大概是 US$26K/mo 的收入:

Kagi search is currently serving ~2,600 paid customers. We have seen steady growth since the launch 3 months ago. Note, this is with zero marketing and fully relying on word of mouth. We prefer to keep things this way for now, as we are still developing the product towards our vision of a user-centric web search experience.

後面在講財務狀況也是類似的數字 (幾乎都是 Kagi 的付費收入):

Between Kagi and Orion, we are currently generating around $26,500 USD in monthly recurring revenue, which incidentally about exactly covers our current API and infrastructure costs.

這個收入差不多 cover 目前的 infrastructure 部份,但還有薪資與其他的 operating cost 大約在 US$100K/mo 這個數量級,看起來還有很大的距離:

Between Kagi and Orion, we are currently generating around $26,500 USD in monthly recurring revenue, which incidentally about exactly covers our current API and infrastructure costs.

That means that salaries and all other operating costs (order of magnitude of $100K USD/month) remain a challenge and are still paid out of the founders’ pocket (Kagi remains completely bootstrapped).

然後要大概是目前十倍的付費數量才會打平 (25K 個使用者):

We are planning to reach sustainability at around 25,000 users mark, by further improving the product, introducing new offerings and pricing changes. With the product metrics being as good as they are, we should be able to reach this as our visibility increases.

比較好一點的消息是 churn rate 很低:

Product stickiness is also very high, with churn being lower than 3%.

然後提到每個使用者大約 27 次查詢 (包括 free tier),有些 user 大約在 100 次,peak 是 400 次:

We are currently serving around 70,000 queries a day or around ~27 queries/day/user (this includes free users which are about 10% of total users). There is a lot of variance in use though, with some users regularly searching >100 times a day. Every time we see a search count go >400 times in day we are happy to be an important part of someone’s search experience.

我看了一下自己的用量,看起來偏高一些,但沒到他說的每天平均 100 次:

然後提到了推出新方案的計畫,包括 Teams Plan & Family Plan,而目前在跑的方案會被分類到 Individual Plans。

另外比較重要的是 Individual Plans 有漲價的計畫。新的方案預定分成三個層級,主要是增加了一個 Kagi Starter 的版本:

  • Kagi Unlimited - $19/mo or $180/year ($15/mo) or $288/biennial ($12/mo) - Original Kagi experience, with unlimited searches
  • Kagi Starter ($5/mo; 200 searches) - For casual users who make less than 200 searches per month
  • Free basic - 50 free searches that reset every month


If such change to Individual plans is to occur, we plan to grandfather-in all early adopters (meaning all current and future paid customers, up until this change) allowing them to keep their existing subscription price as long as they don’t cancel it.


Slack 宣佈漲價

Slack 宣佈了漲價的計畫:「Everything you need to know about our first price change and plan updates」。

從今年九月開始漲價,Pro 的月費從 US$8 漲到 US$8.75 (+9.375%),年費從 US$80 漲到 US$87 (+8.75%):

On September 1, 2022, monthly Pro plans will increase from $8 USD to $8.75 USD per user per month, and annual Pro plans will increase from $6.67 USD to $7.25 USD per user per month (pricing for other countries is available here).

現有用戶有一次用原價 renew 一年的機會:

Current customers on an annual Pro plan can lock in the existing annual rate of $6.67 USD per user per month for another year by early-renewing the annual Pro plan before September 1.


Amazon EC2 有 Mac (M1) 機種可以租用了

2020 年年底的時候 AWS 推出用 Mac mini 配合搭建出 Mac (Intel) 機種:「Amazon EC2 推出 Mac Instance」,當初有計畫在 2021 年推出 M1 的版本:

Apple M1 Chip – EC2 Mac instances with the Apple M1 chip are already in the works, and planned for 2021.

不過就沒什麼意外的 delay 了,這次則是推出了 M1 的版本:「New – Amazon EC2 M1 Mac Instances」。

依照說明看起來還是 Mac mini,掛上 AWS Nitro System

EC2 Mac instances are dedicated Mac mini computers attached through Thunderbolt to the AWS Nitro System, which lets the Mac mini appear and behave like another EC2 instance.

然後跟 Intel 版本一樣,因為是掛進 Dedicated Hosts 的計價方式,雖然是以秒計費,但還是設定最低 24 小時的租用時間限制:

Amazon EC2 Mac instances are available as Dedicated Hosts through both On Demand and Savings Plans pricing models. The Dedicated Host is the unit of billing for EC2 Mac instances. Billing is per second, with a 24-hour minimum allocation period for the Dedicated Host to comply with the Apple macOS Software License Agreement. At the end of the 24-hour minimum allocation period, the host can be released at any time with no further commitment.

Intel 版本代號是 mac1,只有一種機種 mac1.metal,M1 版本代號是 mac2,也只有一種機種 mac2.metal

以最經典的美東一區 us-east-1 來看,mac1.metal 的 on-demand 價錢是 US$1.083/hour,mac2.metal 則是 US$0.65/hour,差不多是 60% 的價錢,便宜不少,大概是反應在硬體攤提與電費成本上了。

另外目前大家用 M1 的經驗來看,Rostta 2 未必會比原生的機器慢多少,雖然 mac1.metal 是 12 cores,mac2.metal 是 8 core,但以雲上面一定要用 Mac 跑的應用來說,馬上想的到的還是綁在 Apple 環境裡 CI 類的應用?

目前看起來主要的問題還是 24 小時的最小計費單位讓彈性低不少...

AWS 在同一區不同 AZ 頻寬費用的特別地方

剛好在處理 AWS 同一個 region 下不同 AZ 之間的傳輸費用,跟帳單互相比對,查了以後才發現跟想像中不一樣,這邊以 EC2 為例子,可以參考「Amazon EC2 On-Demand Pricing」這頁裡面的說明。

從 Internet 端進 AWS 的流量是不計費的:

Data Transfer IN To Amazon EC2 From Internet
All data transfer in $0.00 per GB

但從 AZ 進到另外一個 AZ 時,in 與 out 都要收費:

Data transferred "in" to and "out" from Amazon EC2, Amazon RDS, Amazon Redshift, Amazon DynamoDB Accelerator (DAX), and Amazon ElastiCache instances, Elastic Network Interfaces or VPC Peering connections across Availability Zones in the same AWS Region is charged at $0.01/GB in each direction.

所以直接用 US$0.01/GB 的計算是不夠的,得用 US$0.02/GB 來計算。

同樣的,如果是 Public IP 與 Elastic IP 也都是雙向收費,跨 VPC 也是雙向收費,所以都要用 US$0.02/GB 來算:

IPv4: Data transferred “in” to and “out” from public or Elastic IPv4 address is charged at $0.01/GB in each direction.
IPv6: Data transferred “in” to and “out” from an IPv6 address in a different VPC is charged at $0.01/GB in each direction.

Kagi 開始收費了

在「Kagi search and Orion browser enter public beta」這邊可以看到 public beta 與收費的消息:

We decided to start charging for Kagi search while in beta status because the cost of beta usage has gone up dramatically and we are not able to sustain it. Also, we want to get some kind of a financial “airworthiness” signal and see how we stand with our positioning and outlook for sustainability.

收費的費用是 US$10/mo,剛剛先把信用卡掛上去了...

DigitalOcean 全面漲價

清 RSS reader 的時候才翻到的,DigitalOcean 宣佈漲價:「New DigitalOcean Pricing」,在 Hacker News 上有對另外一篇「公告」也有些討論:「New $4 Droplet and updated pricing」、「DigitalOcean: New $4 Droplet and updated pricing (」。

如果就比例來說,VPS 裡漲最多的就是一般性的產品線 Basic Droplets 這組,直接漲了 20%,其他的翻了一下舊的 pricing 頁面,看起來大多在 20% 以下,除了 IPv4 的費用 (這個倒是知道發生什麼事情)。

VPS 的 $5/mo 1GB 競爭算是很激烈的,算是頂標,這次 DigitalOcean 主動漲價,沒意外應該是會有不少人開始規劃搬家?接下來應該看看另外兩家的動作...

看了一下另外兩家最近的情況,Linode 在二月被 Akamai 買走之後目前主要是推出了 Managed Databases 這個服務,看起來速度慢下來不少。(不過 Linode 推新產品的速度本來就不快)

Vultr 還是照慣例丟了一堆東西出來,在新的據點開機房,以及一些產品線...

話說回來,查資料的時候發現 DigitalOcean 有 581 人 (2020/12),Linode 則是 200+,Vultr 則是 75 (2022,engineering only),看起來 DigitalOcean 的組織頗肥的啊...

Amazon SESv2 的 Deliverability Dashboard

其實是看到「Amazon SES V2 now supports email size of up to 40MB for inbound and outbound emails by default」這篇才注意到寄信的 Amazon SES 服務有了 SESv2,原文主要是講放寬信件的大小限制:

With this launch, the default message size limit in Amazon SES V2 increases from 10MB for email sending and 30MB for email receiving, to 40MB for both sending and receiving .

不過我跑去「Amazon SES pricing」看的時候意外翻到這個貴貴的東西:

The Deliverability Dashboard (via the SES API V2) is available for a fixed price of USD $1,250 per month. This charge includes reputation monitoring for up to five domains and 25 predictive email placement tests.

然後我試著去找 Deliverability Dashboard 是什麼,卻沒有專文介紹?(還是我找錯關鍵字...)

倒是在 2018 年的時候 Amazon Pinpoint 有個公告提到 Deliverability Dashboard,價錢也是 US$1,250/mo:「Amazon Pinpoint Announces a New Email Deliverability Dashboard to Help Customers Reach their Users' Inboxes」。

本來以為是 Amazon Pinpoint 的服務轉移掛到 SESv2 下,但看「Amazon Pinpoint Pricing」這邊,好像還是在啊...

雖然用不太到,但還是一頭霧水 XDDD

Amazon EC2 推出 c6a 的機器

Amazon EC2 以新的 AMD 架構 (雖然也推出一陣子了) 的機器推出 c 系列的機種,代號為 c6a:「New – Amazon EC2 C6a Instances Powered By 3rd Gen AMD EPYC Processors for Compute-Intensive Workloads」。

價位上與 c5a 相比便宜一點點,是真的一點點:在 us-east-1c5a.24xlarge 是 US$3.696/hr,而 c6a.24xlarge 是 US$3.672,差 0.65% 左右... (千分之六點五 XD)

所以宣稱的 15% 基本上都是從 CPU 效能提昇貢獻的:

Up to 15 percent improvement in compute price performance.

然後機器可以提供的範圍比較大台,c5a 最大到 c5a.24xlarge,而 c6a 支援了 c6a.32xlargec6a.48xlarge


C6a instances are available today in three AWS Regions: US East (N. Virginia), US West (Oregon), and EU (Ireland). As usual with EC2, you pay for what you use. For more information, see the EC2 pricing page.

另外這次推出後,EC2 的機種超過 500 種了,主要是靠排列組合弄出來的:

PS – With the launch of C6a instances there are now officially more than 500 Amazon EC2 instances for customers to choose from!