What is difference between i3 i5 i7 ?

[Someone]

But it is still true that i7 > i5. It's just not the whole story.

Hang on, its true but it isnt? Most people would call that not-true.

Its a marketing generalisation, not a truth.

[brucehoult]

I think you are failing to take into account that on these very small processes now it costs many millions of dollars to make a different mask set, not to mention probably a lot more than that for layout and verification.

An Alder Lake 8P+8E chip is 215 mm^2. That's about 300 dies from a 300mm wafer. If the cost is similar to TSMC then that's around $15000 for the processed wafer, or $50 per chip.

They are also making a 6P+0E die that is 163 mm^2. So that's about 400 per wafer or $37.50 per chip, a $12.50 savings.

If it costs $50m all up to make a different die for that (which is probably on the low side) then they have to sell 4 million of the cheaper chips to break even, vs just letting that silicon go to waste by disabling cores on the bigger chip.

There are about 90 million x86 PCs sold a year, between Intel and AMD.

I don't think they can afford to make a custom die for every core configuration they sell.

[coppice]

I think you are failing to take into account that on these very small processes now it costs many millions of dollars to make a different mask set, not to mention probably a lot more than that for layout and verification.

I think most engineers who are not involved in silicon development would be astonished to see some actual figures for the amount of dark silicon shipped today in devices from small MCUs to the largest and most complex. Just making a mask set costs a staggering amount for a fine geometry process, ignoring the large development and proving costs for a new die. A  lot of reduced devices are launched by turning off parts of a more complex device, to fit a market opportunity at an acceptable price. The basis is usually that if the device is a hit, a custom die will be produced to optimise price and profit. However, when it comes to a decision for a hit product there are usually other ways to use the available engineering resources that have greater appeal. Massive volumes can be shipped for highly unoptimised hit products.

[SiliconWizard]

(...)
I don't think they can afford to make a custom die for every core configuration they sell.

Of course not. And it's not just about the NRE and immediate cost of wafers (even though it's a huge driver in itself.) It's also about maintenance, testability, traceability, you name it. Different dies would have to be treated as completely different products. The added cost can be gigantic.

Somehow, many people tend to feel "scammed" when being sold a product (albeit at a lower price point!) they know could do much better and have been "artificially" dumbed down. But that's just the way the semiconductor industry can work at all.

And, as is also common with semiconductors, the dies may be the same, but some may have been tested fully working only for a subset of features, and sold as such. That's better for everyone than just throwing them away.

[coppice]

And, as is also common with semiconductors, the dies may be the same, but some may have been tested fully working only for a subset of features, and sold as such. That's better for everyone than just throwing them away.

You see people talking a lot about testing and accepting die to work around defects, but its a much smaller activity than you might think outside the leading edge processes, where the yields are still being aggressively worked on. The great majority of dark silicon being sold is perfectly good silicon that has just been turned off for commercial reasons. Its just not worth trying to separate out most wafers into bunches of dies which pass various test specs. The testing might be reduced when turning a wafer into a reduced functionality chip, but they will all either pass or fail just one test spec and move forward on that basis.

[TheBay]

After a certain generation it's mostly marketing. For example I'm typing this on a device with an i7 CPU but this is a Dual core i7, one would think it would be faster than an i3 or i5 of this generation but it is not.

Quickest way I find to gauge performance between models is to check the passmark score, yes it doesn't cover all bases but if there is a massive difference is scores it's pretty obvious what the difference is.

I'm in the process of trying to buy a new laptop and the 11th/12th generation Intel nomenclature is just all over the place. Look out for that last letter on the end as that can make a huge difference.

[Berni]

You see people talking a lot about testing and accepting die to work around defects, but its a much smaller activity than you might think outside the leading edge processes, where the yields are still being aggressively worked on. The great majority of dark silicon being sold is perfectly good silicon that has just been turned off for commercial reasons. Its just not worth trying to separate out most wafers into bunches of dies which pass various test specs. The testing might be reduced when turning a wafer into a reduced functionality chip, but they will all either pass or fail just one test spec and move forward on that basis.

Simpler chips can also take different testing routes.

For example MCUs that go for automotive qualification and high temp range might go trough extra testing to satisfy the spec. Even more so for the mil spec stuff (Where you can buy a 7805 regulator in a fancy ceramic package for like 500$ per chip from TI).

Then you have things like voltage reference chips that get sorted into grades by performance (absolute tolerance or temp co.) where the only difference is a single letter after the name.

There are also chips with multiple channels that get artificially handicapped down. Things like ADCs, DACs, Digital Pots...etc where from the datasheet you can clearly see the 8ch, 4 ch, 2ch and 1ch chips are the same die with identical internal registers, they just didn't provide you the pins to hook it up. Often they do give you a smaller chip package to save space, but there are chips out there that use the same package for the 4ch and 1ch version, except that 3 pins are labeled as NC on the 1ch version. You also find families of DAC chips that appear to be exact same chip including the internal registers, but the only difference is that one is 14bit, next one is 12bit next one is 10bit and last one is 8bit.

Some of this is market differentiation so that you can sell the same chip as low spec to cheepskates and at the same time charge a lot the people who are willing to pay for the higher spec. But yield can be a factor with precision analog chips even if they are not pushing the latest and greatest nanometers. Not that the process would produce a chip with faulty transistors, but instead transistors that don't meet the tightly controlled characteristics needed for the chip to be precise

[Mechatrommer]

someone is want to playing clever than Intel's internal reps? ironically i found it...

https://www.intel.com/content/www/us/en/processors/processor-numbers.html

i never care until this thread, i mean what do you care? and what do you get even if you know the absolute truth about it? like recently, i bought a 2nd hand i7-4790K just because its the fastest CPU that i can put in my other existing PC, and i'm happy with its used price of $100+. if i'm nuts about i9 nomenclature, i would have buy an entirely new CPU and motherboard and new RAM all that from scratch that can cost me another thousands $... and if i'm really nuts, i'm just google the fastest Intel or Ryzen around, with the largest cache and cores anything that i care about that i think will be beneficial for me... if i can afford that, such as Core i9-12900KS for a good $700+ CPU only... anyway selecting the right bang/buck CPU (PC HW combo) for our own need nowadays is not as easy as before, we have Intel and Ryzen (let alone Radeon and NVIDIA and MoBo brands) with many namings each, like many candies with different colors in a bin, eyeballs can spin try to google each spec, but luckily we have websites that can make comparisons, such as just type how many cores, minimum speed etc, and the list will appears... even the motherboards list that can support them vice versa... ymmv.

[Someone]

Had a few spare moments to scrape the data from the geekbench multicore benchmarks. Anything with letters attached to the model number are alpha blended.

[Doctorandus_P]

I've stopped long ago at putting any value on marketing speak.
I also do not care much about small differences.

When it was time for me to buy a new PC, I mostly just used the numbers from: https://www.cpubenchmark.net as an indication.
I especially found the XY scatter diagram useful:
https://www.cpubenchmark.net/cpu_value_available.html#xy_scatter_graph

AMD overall has a better price to performance ratio.
I finally settled on a Ryzen 5600G. It has a quite decent performance for a relatively low price, and because it has integrated graphics (which are quite adequate for use with my 4k monitor) the price to performance ratio is a lot better for me. (Graphics is ignored in those benchmarks). I also bought my PC in December 2021 and in that time video cards were overpriced by a factor of 3 or more, if you could even buy them at all.