Hacker News Reader

Comment by bmenrigh | Hacker News Reader

Semiconductor manufacturing, especially using EUV, is a staggeringly power-hungry process.

The EUV light production requires more than a megawatt by itself because of all the losses involved.

No doubt this is why they only report on embodied carbon since it’s the one metric they can win on.

The thing is that anything electrical energy intense is going to become less CO2 producing every year as energy systems transition. Its the use of Iron and other necessary fuel usage that will become the CO2 intense items over the next 5 to 10 years. The energy usage wont be the issue and CO2e for production of any item is going to drop quite drastically.

Iron and steel can be made without fossil fuels. It's just more expensive.

Is anyone using EUV for SSDs? I don't think it's economical for NAND manufacturing. I'm not sure if SSD controllers are using EUV yet; they usually aren't on leading-edge process nodes.

Not all energy is carbon intensive. Using a lot of power is fine as long as you get the power in a sustainable way. That isn't necessarily true for a lot of current production. But that's changing rapidly. Especially in the parts of the world where this stuff is produced.

Agree if that power is truly sustainable.

OTOH, if you have solar panels and drive power to the grid, you can sell "credits" to large power users who will use them to "offset" their less-green power sources to appear to get their power in sustainable ways / be carbon-neutral. Kind of an "I'm using dirtier power, but I teamed up with someone who's using fully sustainable power, so it all comes out in the wash!"

It will certainly become more fine so long as the grid is transitioning to renewable energy sources which almost all countries are doing. In the next decade the intensity of CO2e for electrical energy usage should be considerably lower than it is now and CO2e from necessary fuel usage will begin to dominate in industrial processes. Iron I think will be a big contributor to this since at the moment electrical production is not viable.

China the largest producer of iron and steel has started restricting coal based for electric arc based furnaces for steel making so iron and coal carbon use will probably come down as well.

https://energyandcleanair.org/publication/turning-point-chin...

Its just going to take a bit longer for Iron and steel production to transition, they likely wont bother until we see the full reduction in electrical price that comes with renewables. They do seem to have a process it just needs some refinement. The main grid transition is happening first and faster.

This is part of the issue with Seagate's assessment, its changing rapidly right now and in 5 years time Taiwan will be on mostly renewable energy and suddenly all of Seagate's iron usage will make them more expensive, then in 10 years time they are using electrically produced iron and its not an issue anymore.

The "not all" argument is very lazy of you. All these production plants are on the grid and they use carbon energy. How much is "not all" ?

Absolutely not.

The poster has highlighted a major issue that is often glossed over in terms of actual carbon footprint. It's like saying a Chinese coal plant is the same as a Scottish hydro dam otherwise which is very, very wrong.

Yes nothing is perfect, but running a dam in an area with little water silt for many many years is different in carbon footprint to a coal plant run at 80% efficiency from strip mined coal.

If you are part of a large grid, power is fairly fungible (within that grid).

Nope.

Again look at Spain, not all power sources are equal and just because it's been connected to a network doesn't magically mean all of the load is then spread between all sinks.

If your high school science teacher told you so, they were wrong. It's time to talk about proper grown up electronics in this discussion.

The “time to talk about grown up…” is a bit much.

No, electronics at scale isn't the same as a 9v and filament bulb. There's real effects which aren't overly observable in a small setup that need to be addressed.

It's like pretending all of compiling is a direct translation to a lower language all the way down. It's not you make changes, adapt and optimise for the language and system you're in not just write Fortran in assembly

I was including the weasel world 'fairly' in front of 'fungible' exactly because I foresaw such a discussion.

To be slightly more precise and wordy in what I wanted to express:

The original comment said:

> Not all energy is carbon intensive. Using a lot of power is fine as long as you get the power in a sustainable way.

And that's true. You can buy 'green electricity' from the grid. But that mostly just means that the people who don't care get allocated a larger fraction of 'non-green' electricity.

(That works, until you lose enough of that buffer of people who don't care where they electricity is coming from, that you have more demand for green electricity than is available on the grid. Then you actually need to spin up more green electricity or raise prices for that 'colour'.)

Yes, there's different demand levels over time (throughout the day, and throughout the year etc), and different demanders need different reliability levels. And all the power sources have different profiles for when they are available, and how easy (or hard or impossible) it is to turn them on or off on short notice.

> If your high school science teacher told you so, they were wrong. It's time to talk about proper grown up electronics in this discussion.

Haha, right.

I think the interaction of "thing is cheap" and "thing is tiny" break all of our heuristics for "thing is difficult and energy intensive to produce".

But a spinning harddrive contains electronics too ...

A HDD has 1 controller chip and that controller is relatively simple compared to the controller chip on an SSD, its certainly a lot smaller and on a less advanced process because it has to deal with ~200MB/s not 15000MB/s. On top of the controller chip a 2280 NVMe SSD will have 4 or 8 memory chips as well, which are on a recent process and have many many layers stacked together. Then the SSD will also often have a DRAM chip as well. An SSD uses substantially more silicon than a HDD does.

Not disputing your broader point necessarily, but don't hard drives also include memory for cache while dram cache has become less and less common on SSDs? (to point where it is now uncommon on new SSD devices)

Yes they do. A consumer HDD from Seagate seems to have about 512MB whereas an SSD from Samsung seems to use 2GB.

HDDs use much simpler electronics than SSDs. The controller in a hard drive mostly just moves the read/write head and spins the motor. It doesn’t need to do much processing. These chips are built on older manufacturing processes that are cheaper and less energy intensive.

SSDs though need a fast processor to manage the flash memory for wear leveling, error correction, and keeping track of where everything is written etc. This requires more advanced chips, built on newer process nodes, which take a lot more energy and resources to make.

And SSDs also need extra chips like DRAM and obviously the flash memory itself.

Is the controller really going to be that much simpler on a HDD? On HDD you still have things like command queuing, caching, checksumming and bad block remapping just as with SSD. Sure you don't have the flash translation layer, but I'd expect processing of the analog signal going to/from the heads to be a processing headache too. I wouldn't be surprised if they have to do temperature compensation and factory and/or runtime calibration of the analog path. Of course much of that will be done in fixed-function custom hardware blocks, but almost certainly so would a SSD for eg. ECC calculations.

https://spritesmods.com/?art=hddhack&page=3 has some data points on a 10+ year old HDD - 3 ARM9 feroceon cores + Cortex-M3 core. Doesn't sound too simple.

HDD controller calculates and checks ECC either and en/decodes the user's 0/1-bits to the 0/1-bits actually to be written on the platter but at much smaller speed (0.3 GB/s vs 14 GB/s SSD max). HDD has an extra servo controller. HDD has an extra 512 MB - 1 GB DRAM.

WD HDDs 20 GB+ have some FLASH inside for metadata offload.

DRAM is optional on NVME SSD's which can use host memory instead. DRAM-less SATA SSD's sucked.

Reply on news.ycombinator.com