Hacker News Reader

Comment by bee_rider | Hacker News Reader

I'd assume the opposite actually.

Chips always have to be binned. But previously a chip would have to be binned down to it's worst component I guess -- if they had a chip with great CPUs but the GPUs were a little wonky, and they didn't have an appropriate processor line for that combo, they'd have to bin the whole thing down to low-tier. Now they can instead match up the good CPUs and the good CPUs.

Plus they'll be able to satisfy some of their lust for SKUs by mixing and matching tiles, rather than making a bazillion slightly bins.

View on HN

> lust for SKUs

That's so Intel. All 37 variants of the Intel Core i9 CPU: [1]

[1] https://www.intel.com/content/www/us/en/products/details/pro...

Core i9 isn't a useful thing to complain about SKUs for. There's probably been thousands of Pentium products at this point.

What you want to look at is how many SKUs for an architecture, like say Alder Lake for Desktop[1]. Do we really need 5 to 7 SKUs at 16, 12, 6, or 4 cores, but only 2 SKUs at 10 cores, and 4 SKUs at 2 cores?

[1] https://ark.intel.com/content/www/us/en/ark/products/codenam...

2 cores seems like a niche product at this point, actually 4 SKUs for that is more than I'd expect.

Only 2 SKUs at 10 cores seems a little weird, I wonder if they are 12 core parts with some cores disabled or something like that.

Edit: Note it is just the i5 [...]K's from Q4 '21 that have 10 cores. It isn't that surprising that the early enthusiast parts are a little weird, right?

There are four Alder Lake dies in production, two for desktop (8+8 and 6+0 P/E-cores) and two for mobile (6+8 and 2+8 cores). So the 10-core desktop parts are probably the most cut-down variant of the full 8+8 die, and all the lower core count desktop parts are likely made from the 6-core die that omits the E-cores entirely.

Whether you look at just the desktop segment (29 SKUs from two dies) or the whole family (94 SKUs from four dies), it's an incredibly overcomplicated product stack.

Oh wow, I didn't realize at all that the "10 cores" in this case were actually 6p 4 e.

> , I wonder if they are 12 core parts with some cores disabled or something like that.

If demand for that SKU is higher than for the 12 core SKU and margin justifies it, then yes. It happens with every vendor, specially when node been out for some time - yields are much better and there aren't this many down-binned chips to handle the demand.

My understanding is that binning is mostly about core count and core operating frequency.

Intel has a hard-on for segmenting their product lines. To the point they "launch" new products like the i9, which is just what an i7 used to be. They also deliberately cripple products, like selling SKUs with VT-x disabled. Not to mention them keeping ECC memory out of the entire desktop market for basically all of history.

> They also deliberately cripple products, like selling SKUs with VT-x disabled

Ha, that has been a long time, no? I remember buying a computer with a Q8300 in 2008 or so and there I had to assure a specific revision.

I'm sure part of that is just marketing segmentation, but significant binning does take place. If VT-x is disabled, that may indicate that there were some defects in the location that those circuits inhabit.

Stuff like virtualization features and Hyperthreading are responsible for such a tiny fraction of extra die area and so deeply entwined with critical functionality that it's extremely unlikely for a given part to need to be binned on the basis of whether that feature works or not. Intel would never amass enough broken-HT but otherwise good chips to provide the impetus for an extra SKU.

I don't know any more than the general public, but I would assume VT-X and VT-D is actually just microcode.

Reply on news.ycombinator.com