For years now, I do not buy/create assemble a new computer, because I am totally overwhelmed by the options available to me.
If we agree there is ‘The Paradox of Choice’, it seems to make sense to have a much more limited choice between CPU models from a consumer point of view. For example, have for each year an entry, business and a pro model, add extreme for gamer and have each of these models have a version with a beefy integrated CPU.
But it seems also a good idea for the manufacturers: They have to design, test and build each of their models, create advertisement etc., like configuring their assembly lines alone costs money. Further, compilers have to generate code for a specific architecture, which means that all my software I didn’t compile myself ends up using an instruction set of the lowest common CPU, not utilizing whatever I bought fully.
Apple (not a fan ;-)) shows IMHO how it is done with their Apple Silicon: Basically even I understand which CPU choice would be the right one for me. The Steam Deck is IMHO another success story: As reference hardware I know easily if I can play a game, and it is easy to know if my hardware is faster than a Steam Deck. Compare that to games with hardware requirements like ‘AMD TI 5800 8GB RAM’ (made up model) which makes my life miserable.
What I am looking for is fact based knowledge:
- Why does it make (commercial) sense for AMD/Intel to create so many models?
- What are their incentives?
- What would happen, if they would reduce the amount of different CPUs they offer? (Is there historical knowledge?)
To a certain extent, processors that only partially work can be sold as a different model. For example, if AMD makes a bunch of 8 core processors, they might be sold as a number of different models based on what, if anything, is wrong with them. The best ones where everything works will be sold as the highest model. Then the ones that can’t achieve the highest clocks, but all the cores and the iGPU works will be sold as the next highest model. Then the ones where a couple cores are bad will be sold as 6 core models, and so on.
They’re made with, essentially, toggle jumpers that can be “cut” to disable different parts of the CPU (I think this is in the chip’s internal firmware and not an actual hardware cut). So two CPUs with the same SKU might have different internal cores that are enabled/disabled. The SKU is basically just a guarantee of how many cores it has, what frequency they will run at, same for iGPU cores, IO capability, and cache.
Obviously, this only applies to CPUs with compatible die sizes. That’s determined by the package it’s ultimately going to be mounted on.
It’s called binning, and it helps achieve higher yields by allowing you to sell more of the “defective” units. The more variations they have, the more they can get for each chip on average. Like if they have a SKU with 8 cores and 32 PCIe lanes, and a SKU with 6 cores and 24 PCIe lanes, then a chip where all 8 cores work, but only 30 of the PCIe lanes work, would have to be sold as the cheaper 6 core SKU. Adding another SKU with 8 cores and only 24 PCIe lanes, would let them price that same chip higher.
https://www.techspot.com/article/2039-chip-binning/
One of the coolest applications of this I’ve seen is the CPU AMD released recently which was literally a PS5 CPU with the graphics disabled (because it didn’t work, or didn’t work well enough to put in a PS5).
Thanks, great write-up!
I had no idea this is how lower end CPUs worked
Why does it make (commercial) sense for AMD/Intel to create so many models?
Because there is demand for various types of systems. And on top of that, if you make a chip with 8 cores and two are defective… just sell a 6 core chip instead of throwing it away.
What are their incentives?
Money
What would happen, if they would reduce the amount of different CPUs they offer? (Is there historical knowledge?)
They would lose customers to competitors in that space. When AMD didn’t make EPYC chips, all servers were Intel Xeon.
-
Binning. By detecting defects and working around them by making a design that can function without some portions of the processor and disabling a defective part of a processor, and selling it as a lower grade, one can have a higher rate of salable processors.
-
Splitting up the market to permit use of price discrimination. If I know that people who want error-correcting memory are less price sensitive than those that are, I have a separate line of processors and motherboards that costs much more and avoid offering error correcting memory on my cheaper line. This converts consumer surplus into producer surplus.
Binning is an important part of the answer. They could have only 5 CPU models, but they’d waste silicon that way.
-
Some of it comes from “binning” of chips. Despite our technology, processor manufacturing is kind of a gamble, the number of transistors a chip will have will be somewhat random, and the performance will vary. They will then sort and separate these processors by speed, “binning”.
That’s why you see CPU models that are nearly identical to each other but vary slightly in speed.
Plus, companies love money. They will make a product for every possible conceivable market. Say if somebody doesn’t want to spend 200$, but can afford something greater than 150$, there will be a CPU for that gap.
Then different workloads require different types of CPUS. Single core applications need high clock speed, protein folding and needs many, slower cores and servers need processors that prioritize stability.
I don’t know a lot of the specifics, but to provide a couple points that help put things into context:
In regards to manufacturing different more-similar models, this is often the result of binning. A bunch of CPUs will be made at once through the same process, but as CPU production is an incredibly precise process, theres still quite a bit of variety between the results. You can think of it the same as separating grades of produce on a farm. The best stuff becomes the higher-tier, more expensive models, the average stuff get an average price, and the worst stuff becomes a budget option. This provides 3 separate models for no manufacturing resources.
Similarly, some lower end parts, in some cases, can even come from defective higher-tier parts. I believe AMD released a model like that a year or two ago, which is an easy way to recoup costs and avoid waste.
Also notable is that the range of products and prices also allows for targeting a wider range of customers. For example, Ryzen 3 would be those who just want something that works, but won’t spend much, Ryzen 5 is for average people who just want a good processor but don’t want to spend much, Ryzen 7 is enthusiasts, and Ryzen 9 is people doing specialized commercial work or enthusiasts with more money than sense. All of these are not just different products, but also allow the company to charge different margins - low end won’t spend much whereas high-end is willing to spend money and often expects a monitary return from their purchase. Most consoles and pre-builts are more towards the low-end, but theres still penty of demand for those higher-end parts for better performance or for complex work that needs it.
The biggest pet peeve I have with modern hardware versions is that in a laptop, the same hardware performs similair to a lower version of a desktop variant.
Don’t call it a 3070TI if it performs like a 3060TI. Just call it a 3060 TI from the start.