Model Parallelism with Pipeline Parallelism

In Model Parallelism, the neural network is partitioned across multiple GPUs, with each GPU responsible for specific layers of the model. This strategy is particularly beneficial for large-scale models that surpass the memory limitations of a single GPU.

Conversely, Pipeline Parallelism involves dividing the model into consecutive stages, assigning each stage to a different GPU. This setup allows data to be processed in a pipeline fashion, akin to an assembly line, enabling simultaneous processing of multiple training samples. Without pipeline parallelism, each GPU would process its inputs sequentially from the complete dataset, while all other GPUs remain idle.

Our example neural network in Figure 8-3 consists of three hidden layers and an output layer. The first hidden layer is assigned to GPU A1, while the second and third hidden layers are assigned to GPU A2 and GPU B1, respectively. The output layer is placed on GPU B2. The training dataset is divided into four micro-batches and stored on the GPUs. These micro-batches are fed sequentially into the first hidden layer on GPU A1. 

Note 8-1. In this example, we use a small training dataset. However, if the dataset is too large to fit on a Continue reading

PP053: Rethinking Secure Network Access and Zero Trust With Bowtie (Sponsored)

What A Tangled OpenAI Web We CoreWeave

Wouldn’t it be funny if all of that money that Microsoft spent last year paying neocloud upstart CoreWeave was just to support ever-embiggening AI training workloads at OpenAI as it makes its GPT models smarter? …

What A Tangled OpenAI Web We CoreWeave was written by Timothy Prickett Morgan at The Next Platform.

HS097: Build A Strategy, Not Just A Project Plan

How Calico Network Security Works

In the rapidly evolving world of Kubernetes, network security remains one of the most challenging aspects for organizations. The shift to dynamic containerized environments brings challenges like inter-cluster communication, rapid scaling, and multi-cloud deployments. These challenges, compounded by tool sprawl and fragmented visibility, leave teams grappling with operational inefficiencies, misaligned priorities, and increasing vulnerabilities. Without a unified solution, organizations risk security breaches and compliance failures.

Calico reimagines Kubernetes security with a holistic, end-to-end approach that simplifies operations while strengthening defenses. By unifying key capabilities like ingress and egress gateways, microsegmentation, and real-time observability, Calico empowers teams to bridge the gaps between security, compliance, and operational efficiency. The result is a scalable, robust platform that addresses the unique demands of containerized environments without introducing unnecessary complexity. Let’s look at how Calico’s key network security capabilities make this possible.

Calico Ingress Gateway

The Calico Ingress Gateway is a Kubernetes-native solution, built on the Envoy Gateway, that serves as a centralized entry point for managing and securing incoming traffic to your clusters. Implementing the Kubernetes Gateway API specification, it replaces traditional ingress controllers with a more robust, scalable, and flexible architecture that is capable of more Continue reading

Data Deluge Pushes Financial Services Deeper Into AI

SPONSORED FEATURE: As an industry, financial services is accustomed to big numbers. …

Data Deluge Pushes Financial Services Deeper Into AI was written by Timothy Prickett Morgan at The Next Platform.

The Linux Bridge MTU Hell

It all started with an innocuous article describing the MTU basics. As the real purpose of the MTU is to prevent packet drops due to fixed-size receiver buffers, and I waste spend most of my time in virtual labs, I wanted to check how various virtual network devices react to incoming oversized packets.

As the first step, I created a simple netlab topology in which a single link had a slightly larger than usual MTU… and then all hell broke loose.

Europe Takes Another Whack At Homegrown Compute Engines

With RISC-V International, the body controlling the RISC-V instruction set, located in Switzerland for the past five years, RISC-V now has just as much right to call itself indigenous to Europe as does Arm Ltd, the British chip company that finds itself on the other side of the English Channel after the Brexit break up and that is still around 90 percent owned by Japanese conglomerate SoftBank. …

Europe Takes Another Whack At Homegrown Compute Engines was written by Timothy Prickett Morgan at The Next Platform.

Tech Bytes: How Internet Synthetic Transactions Boost App Performance Visibility (Sponsored)

Parallelism Strategies in Deep Learning

Introduction

Figure 8-1 depicts some of the model parameters that need to be stored in GPU memory: a) Weight matrices associated with connections to the preceding layer, b) Weighted sum (z), c) Activation values (y), d) Errors (E), e) Local gradients (local ∇), f) Gradients received from peer GPUs (remote ∇), g) Learning rates (LR), and h) Weight adjustment values (Δw).

In addition, the training and test datasets, along with the model code, must also be stored in GPU memory. However, a single GPU may not have enough memory to accommodate all these elements. To address this limitation, an appropriate parallelization strategy must be chosen to efficiently distribute computations across multiple GPUs.

This chapter introduces the most common strategies include data parallelism, model parallelism, pipeline parallelism, and tensor parallelism.

Figure 8-1: Overview of Neural Networks Parameters.

Data Parallelism

In data parallelization, each GPU has an identical copy of the complete model but processes different mini-batches of data. Gradients from all GPUs are averaged and synchronized before updating the model. This approach is effective when the model fits within a single GPU’s memory.

In Figure 8-2, the batch of training data is split into eight micro-batches. The first four micro-batches are Continue reading

NB517: TSMC Pledges $165 Billion for US Chip Manufacturing; Will the CHIPS Act Get Axed?

Global expansion in Generative AI: a year of growth, newcomers, and attacks

AI (Artificial Intelligence) is a broad concept encompassing machines that simulate or duplicate human cognitive tasks, with Machine Learning (ML) serving as its data-driven engine. Both have existed for decades but gained fresh momentum when Generative AI, AI models that can create text, images, audio, code, and video, surged in popularity following the release of OpenAI’s ChatGPT in late 2022. In this blog post, we examine the most popular Generative AI services and how they evolved throughout 2024 and early 2025. We also try to answer questions like how much traffic growth these Generative AI websites have experienced from Cloudflare’s perspective, how much of that traffic was malicious, and other insights.

To accomplish this, we use aggregated data from our 1.1.1.1 DNS resolver to measure the popularity of specific Generative AI services. We typically do this for our Year in Review and now also on the DNS domain rankings page of Cloudflare Radar, where we aggregate related domains for each service and identify sites that provide services to users. For overall traffic growth and attack trends, we rely on aggregated data from the cohort of Generative AI customers that use Cloudflare for performance (including Continue reading

Worth Reading: Traffic Steering into LSPs

You can use SR-MPLS, MPLS-TE, or an SDN controller to build virtual circuits (label-switched paths) across the network core. The controller can push the LSPs into network devices with PCEP, BGP-LU, or some sort of NETCONF/RESTCONF trickery.

Unfortunately, you’re only half done once you have installed the LSPs. You still have to persuade the network devices to use them. Welcome to the confusing world of traffic steering explored in the Loopback as a Service blog post by Dmytro Shypovalov.

Tailscale VPN – A Network Engineer’s Perspective

Before we get into what Tailscale is or how it compares to a traditional remote access VPN, let’s take a quick look at Tailscale in action. The main problem Tailscale solves is remote access to your internal workloads.

In my homelab, I have a server running Linux. When I’m on my home network, I can access it directly without any issues. But if I step outside and want to access the same server over the Internet, Tailscale makes that much easier and you can have it up and running in about 10 minutes for free.

Typically, you would set up some kind of a VPN, either running on a server or a dedicated firewall. Then, you’d install a VPN client on the devices and point them to the public IP of your VPN server or firewall. That’s exactly what I have at the moment.

Tailscale takes a completely different approach, and you don’t need any of that. I’m not saying one is better than the other, I’m just pointing this out for comparison. I’ve shared my thoughts on the pros and cons of each solution at the end of this post.

Tailscale Installation and Setup

Head over to the Tailscale Continue reading

From Python to Go 016. Advanced SSH Interaction With Network Devices.

Hello my friend,

This blog continues discussion of how to manage devices (network switches and routers, servers, virtual machines, etc) using SSH, which we started in previous blog. In this discussion we’ll cover advanced interaction with devices, which include multiple commands, change of contexts and validations. Let’s dive in.

If You Say “From Python to Go”, Shall I Learn Python?

Each programming language has its strong and weak capabilities. Golang, by virtue being a low level (or at least much lower level compared to Python) is very good, when performance and efficiency are paramount. However, you don’t need it for all applications. Python give quicker time to market, possibility to iteratively develop your code with Jupyter and vast ecosystem of existing libraries. Both programming languages are important and both of them play crucial role in IT and network infrastructure management. So if you are good with Python, learn Go (Golang) using our blog series.

And if you are not, or you want to have good intro to IT and network automation holistically, enroll to our training programms:

We offer the following training programs in network automation for you:

• Zero-to-Hero Network Automation Training
• High-scale automation with Nornir
• Ansible Automation Orchestration with Continue reading

Auto-expanding aliases in Zsh

To avoid needless typing, the fish shell features command abbreviations to expand some words after pressing space. We can emulate such a feature with Zsh:

# Definition of abbrev-alias for auto-expanding aliases
typeset -ga _vbe_abbrevations
abbrev-alias() {
    alias $1
    _vbe_abbrevations+=(${1%%\=*})
}
_vbe_zle-autoexpand() {
    local -a words; words=(${(z)LBUFFER})
    if (( ${​#_vbe_abbrevations[(r)${words[-1]}]} )); then
        zle _expand_alias
    fi
    zle magic-space
}
zle -N _vbe_zle-autoexpand
bindkey -M emacs " " _vbe_zle-autoexpand
bindkey -M isearch " " magic-space

# Correct common typos
(( $+commands[git] )) && abbrev-alias gti=git
(( $+commands[grep] )) && abbrev-alias grpe=grep
(( $+commands[sudo] )) && abbrev-alias suod=sudo
(( $+commands[ssh] )) && abbrev-alias shs=ssh

# Save a few keystrokes
(( $+commands[git] )) && abbrev-alias gls="git ls-files"
(( $+commands[ip] )) && {
  abbrev-alias ip6='ip -6'
  abbrev-alias ipb='ip -brief'
}

# Hard to remember options
(( $+commands[mtr] )) && abbrev-alias mtrr='mtr -wzbe'

Here is a demo where gls is expanded to git ls-files after pressing space:

I don’t auto-expand all aliases. I keep using regular aliases when slightly modifying the behavior of a command or for well-known abbreviations:

alias df='df -h'
alias du='du -h'
alias rm='rm -i'
alias mv='mv -i'
alias ll='ls -ltrhA'

Broadcom And Marvell Ride The Compute Engine Independence Wave

Nvidia sells the lion’s share of the parallel compute underpinning AI training, and it has a very large – and probably dominant – share of AI inference. …

Broadcom And Marvell Ride The Compute Engine Independence Wave was written by Timothy Prickett Morgan at The Next Platform.

Choosing Manual or Auto-Instrumentation for Mobile Observability

As applications run in production, you’ll need to find out what’s happening. You might want to know if you’re overloading the hardware, moving to the wrong feature in an A/B test or, on the mobile side, even such simple contingencies as whether the battery is running out. Developing an app to send information about itself means adding instrumentation. Apps can send such telemetry as OpenTelemetry (OTel) project. An added benefit is, if most applications in a given language need to observe the same types of operations and workflows, developers building on the OTel standard can identify and build

HN771: Is The CCIE Still Worth It In 2025?

TNO019: A Look At NetOps Inside a Colocation Facility