“Binned” CPUs/GPUs

E

EnderW

[H]F Junkie
Joined
Sep 25, 2003
Messages
11,250
I’ve always thought of and seen “binning” in the context of sorting out higher performing samples (eg. 13900KS is a binned 13900K) but lately I’ve seen it used when referring to stripped down versions, especially Apple SOCs (M2 with only 8 GPU cores instead of the full 10).

Curious if anyone has thoughts on which is correct or is it one of those things that doesn’t have a clear consensus?
 
It just depends on how well yields go and what market segments the manufacturer is trying to capture.

There used to be a lot more variety in what clock speeds parts could reach, but with only one, two, four cores, any flaws had the potential to completely disable a die.

Now, clock speeds are limited by power and cooling, and flaws only mean a part of a die needs to be turned off.

In general, if yields are good, they will bin the low-end parts to make a budget product. But if yields are bad, they will bin the high-end parts to make a premium product.

They always plan for both outcomes, which is why there's a lot of speculation about code names and potential product lines. And there's always going to be some binning, because there's always going to be bad parts, parts that are just fine, and really awesome parts (they call these "golden samples").
 
I mean, you test the part and then put it in the good bin or the bad bin; sometimes, the literal trash bin. Either way, they've been binned.
 
Binning... Let's say we have some buckets, like 11 of them. And you grab each silicon die and put it in the corresponding bucket.
Some CPUs go to 11.
 
I think of binning as just testing the product. In your example, both are accurate examples of binning. The 13900K that boosts higher than the normal 13900k is going to be "binned" as a 13900KS. The same with the Apple M1. The chip is tested and if it meets the specs of the 10 core, it gets binned accordingly. If two of the cores are "defective" or don't meet the rated speeds, it gets binned as the lower model.

Gone are the days when a chip manufacturer left performance on the table for massive overclocking gains by just "binning" all chips to a low common denominator and allowing the end user to play the silicon lottery to find ones that would clock high on low voltage.
 
Now that core count (GPU and CPU) is a main factor between skus, it became more and more about how much of the silicon work-enabled than how well with the same amount of silicon working it does.

When we talk binning it will often be how many xtx achieved to be an xtx instead of an xt, how many CPU achieved to be a 5950x instead of a 5900x and so on, more so than how well a good 5950x or 5600x does versus a bad one.

Even if the difference between the same skus can still be significant:


But as company got better and better to moneytise fully the best silicon, it is less a factor than in the past when enough P4 3.0 ghz were being made versus the demand and you could end up with P4 2.4 ghz that was good enough to be a p4 3 ghz in reality and that was a big deal.
 
You must log in or register to reply here.
Back
Top