528: Port Automation, Indoor Air Quality, AMD's MI325X GPU vs Nvidia, From Copilot to Pilot Era, Digital Advertising, Starlink, and In Bruges
"a one-time cost that pays dividends forever after"
Uncertainty is where things happen.
–Oliver Burkeman
🫁🚭🧼🧹Indoor air quality is something few people think about, but over the years, I’ve seen a bunch of studies about the impact of fresh air on cognition and about the effects of particulate matter (PM) on cardiovascular health and even cancer rates (PM from diesel exhaust is particularly bad for that), as well as on allergies, asthma, etc.
So it’s probably worth optimizing a bit, especially because most of what can be done consists of one-time changes that pay dividends for a long time.
I don’t pretend to have a perfect setup, but here are a few of the things I do to mitigate risks and increase air quality in my home:
My house has an air exchanger with heat recovery for ventilation. It constantly pumps fresh air from the outside through the HVAC system and exhausts indoor air. This is especially important with newer well-insulated buildings — which is great for energy efficiency, but being more airtight can create problems with air quality. During the cold months when windows are kept closed and there’s reduced circulation, this becomes even more critical.
Back in Edition #429, I told you about a small Levoit 300S air purifier I bought for our master bedroom. At the time, the forest fires were making air quality TERRIBLE around here, and I started researching HEPA systems. I’ve been pleased enough with it that I bought the much larger Levoit 600s for our living room (got it on sale, of course!). It has enough capacity (CFM) to do a pretty good job of filtering the air for much of the rest of the house.
Whenever I fry something in a skillet, I see it detect the PM—like when I was cooking chicken recently:
Note on the graph that it takes a few hours for PM to come back down, and that’s with the purifier!
If you’re more of a DIY person, there’s a fairly inexpensive way to create a great air purifier with a box fan and some large HVAC filters. The idea is that these may not filter as well as expensive HEPA filters, but so much more volume of air goes through them that they end up cleaning the air as well or better (think of it as 25% of 5000 cubic meters vs 99% of 500 cubic meters).
🏦 💰 Business & Investing 💳 💴
🤖🏗️🚢⚓️ The Wonders of Port Automation 🚚📦📦
While I empathize greatly with the individual workers affected by automation at ports — a story that has occurred in other industries as well, and yes, I’ve seen The Wire S2 — the benefits to the rest of society of having higher throughput ports and lower shipping costs cannot be ignored.
Brian Potter has a great primer that covers nearly everything you’ve ever wanted to know about port automation.
Here are some highlights:
Since 2020, the World Bank has released a Container Port Performance Index, which ranks ports around the world based on how long vessels stay in port. Since a ship waiting in port isn’t making money by transporting cargo, shipowners want them in and out of ports as quickly as possible.
Major American ports routinely rank near the bottom of this list. Los Angeles, the largest port in the U.S. by container volume, ranked #375 in 2023, and Long Beach (the second-largest) ranked #373. Savannah ranked #395, and Seattle ranked #360. Of the five largest American container ports, only New York-New Jersey cracked the top 100 (at #92).
What’s the US’ ranking on that index?
Worst in the world 😬
Meanwhile, China does much better, with Shanghai’s port ranking #1.
The U.S. does better on bulk cargo and tanker handling, but only manages to be average or somewhat below average in cargo loading/unloading rates and ship waiting times.
Yay for “average”, I guess..
The basics of port automation
When we talk about port automation, we’re really talking about a collection of different tasks, which can be automated to different degrees. We can broadly break container handling automation into three different groups: ship- or quay-side tasks, yard tasks, and land-side tasks.
The first are tasks required to move containers from the ship onto the shore, the second are tasks required to move containers around the storage yard, and the third are tasks required to move containers out of the port via land-based transport.
I recommend the whole deep dive.
🤖 AMD Unveils Instinct MI325X Data Center GPUs to Challenge Nvidia: Can AMD Close the Gap? 🐜
One more salvo in the GPU wars, as AMD tries to get some of Nvidia’s milkshake:
AMD Instinct MI325X accelerators deliver industry-leading memory capacity and bandwidth, with 256GB of HBM3E supporting 6.0TB/s offering 1.8X more capacity and 1.3x more bandwidth than the H200. The AMD Instinct MI325X also offers 1.3X greater peak theoretical FP16 and FP8 compute performance compared to H2001.
The MI325X will be fabbed with a 4-nanometer process from TSMC. Power consumption is expected to be 1,000 watts.
The MIX325 will have 153 BILLION transistors!
This compares to 80 billion transistors for the Nvidia H200. The previous generation A100 had about 54 billion, and the next generation Blackwell has 208 billion (but that’s split across two chips that are fused).
This leadership memory and compute can provide up to 1.3X the inference performance on Mistral 7B at FP16, 1.2X the inference performance on Llama 3.1 70B at FP83 and 1.4X the inference performance on Mixtral 8x7B at FP16 of the H2004.
Inference is nice, but I wish they also provided training performance on the latest Llama 🦙. The fact that they don’t makes me suspect they’re behind Nvidia there 🤔
AMD Instinct MI325X accelerators are currently on track for production shipments in Q4 2024 and are expected to have widespread system availability from a broad set of platform providers, including Dell Technologies, Eviden, Gigabyte, Hewlett Packard Enterprise, Lenovo, Supermicro and others starting in Q1 2025.
At first glance, this seems impressive and highly competitive with the H200, especially if pricing is aggressive..
BUT
Nvidia is already forecasting that it will ship “several billion dollars” of Blackwell B200 GPUs in Q4, and will keep ramping up from there. So AMD is still a step behind Jensen.
Their Blackwell competitor, the MI355X is only planned to come out in the second half of 2025. By then Blackwell should be ramped up, and if all goes well, we should be hearing about Nvidia’s Rubin generation, which is scheduled for the beginning of 2026..
Keep reading with a 7-day free trial
Subscribe to Liberty’s Highlights to keep reading this post and get 7 days of free access to the full post archives.