Question Why don't CPUs have an odd number of cores ?

[FlameTail]

There are 2,4,6,8,10,12,14...32..128 core CPUs. But all these are EVEN numbers. CPUs with odd-number of cores are almost non-existent. WHY ?

1. Making an actual CPU die with an odd number of cores.

In this scenario, I know the answer to the question.

It's possible to actually make a die with an odd number of physical cores. However this is challenging, because having an odd number of cores kills the symmetry and makes designing the CPU really complicated.

The Xenon ( not Intel Xeon ) processor is a rare example with 3 actual cores physically present in the die.

https://www.quora.com/Why-do-CPUs-always-have-an-even-number-of-cores-4-6-8-etc -Quora: Why do CPUs always have an even number of cores (4, 6, 8, etc.)?

2. Binning a CPU with even-number of cores to get an odd-number of cores.

Example:

An AMD Zen 3 CCD has 8 cores. If one core is defective, AMD can disable that core and sell it as 7 core CPU. However they do not do so. Instead they will probably disable 2 cores and sell it as 6 core CPU. WHY?

(a) Is there any physical/technical/software/hardware limitation that prevents them from doing this ?

(b) If NOT, what is the reason binned CPUs don't have an odd number of cores ?

The Phenom processors used to have a Triple-Core variant, which was basically obtained by disabling one core of a 4-core die. Nonetheless, we haven't seen such a CPU in recent years.

So (a) and (b) are my actual questions and i am looking forward to an answer to these.

 

[amd6502]

AMD used to make plenty of odd core CPU's during the golden Phenom era.

During Piledriver generations this wouldn't have made sense because the "CCX" (loosely speaking) consisted of two modules, with two cores per module. The mobile focused products (aka as "APU"'s) came with one "CCX", and the Opteron and FX desktop products with two CCXs. No way to break that up with satisfactory symmetry unless one had to make a single core product out of these.

I'm not sure why they never offered 3c/6t for Zen products like Picasso; it would've been more preferrable to me to a 2c/4t or SMT disabled 4c. Perhaps olde ways, or stubborn odd superstitions in their engineering or marketing departments.

As for their Zen1 and Zen2 desktop dual CCX products, those had to be paired up with symmetry (unless completely disabling one CCX).

I expect die salvage 1c/2t will ultimately be offered for their Raven2/Dali/Pollock---perfect for chromebook and emmc-ultracheap laptops.

I think there is no technical reason not to offer odd numbered cores for Zen3. The only odd numbers that would make sense (other than in multi chiplet arrangements) would be 1 or 3. The difference between 2c vs 3c is significant, while the difference between 5c vs 4c or 6c wouldn't be very significant (and the difference of 7c vs 8c or 6c even less significant). So if you were on an AMD engineering team proposing 7c die salvage, the marketing department would probably have a word with you.

 

[NostaSeronx]

So a module is a core is it?

A module is a processor. The cores are defined by control unit, instruction bus, datapath, memory bus.

The control unit is fully replicated in BD-XV. => two retire/control units per module
The instruction bus is fully replicated in BD-XV. => two scheduler bus units per module
The datapath is fully replicated in BD-XV. => two ALU/AGU/PRF arrays per module
The memory bus is fully replicated in BD-XV. => two paths to LD/ST per module
Everything else is processor resources.

Thus there is two cores in a Bulldozer Compute Module a.k.a. a Bulldozer Processor.

this website we are on reported the legal outcome of this argument right here

Legal outcome != actual definition by engineers/architects.

 

[amd6502]

A module is a processor. The cores are defined by control unit, instruction bus, datapath, memory bus.

The control unit is fully replicated in BD-XV. => two retire/control units per module
The instruction bus is fully replicated in BD-XV. => two scheduler bus units per module
The datapath is fully replicated in BD-XV. => two ALU/AGU/PRF arrays per module
The memory bus is fully replicated in BD-XV. => two paths to LD/ST per module
Everything else is processor resources.

Thus there is two cores in a Bulldozer Compute Module a.k.a. a Bulldozer Processor.Legal outcome != actual definition by engineers/architects.

I agree, it's no debate. 2c per module. Well put details nosta.

I don't know much about it, but it always seemed to me like a nuissance suit that resulted in a settlement. Which means it wasn't even determined whether such 2c modules were two cores from a legal standpoint.

 

[Mopetar]

With most chips having 8-cores by design it means that there will be plenty of ways to bin those dies. If you have 2, 4, 6, and 8 core variants and three different performance bins for each that's 12 different products already. Creating additional bins for 1, 3, 5, and 7 now mean 24 products which starts to crowd things in terms of pricing and means fewer parts in many of the existing bins.

Even if they could do it, it's really not worth the hassle and makes it more difficult to differentiate performance tiers.

 

[igor_kavinski]

If they have 7 working cores on a die and they disable one perfectly working core just to have even number of cores, that's wasteful. Evil. Unconscionable. It's like burying a baby alive.

 

[igor_kavinski]

Hopefully future CPUs will gain the ability to unlock such dormant cores. For a price. At least on Intel's side.

 

[Thunder 57]

Hopefully future CPUs will gain the ability to unlock such dormant cores. For a price. At least on Intel's side.

They tried something like that with their upgrade cards. It failed spectacularly. Honestly it seemed like a good idea. But people were appalled when they learned about artificial segmentation.

Think about it. Save a few bucks on your initial purchase. When the times comes where you want more performance, pay a bit more to unlock say a 6 core to an 8 core.

 

[IntelUser2000]

If they have 7 working cores on a die and they disable one perfectly working core just to have even number of cores, that's wasteful. Evil. Unconscionable. It's like burying a baby alive.

It seems and may be wasteful but you have to understand from their perspective.

The cost of silicon itself is very little. A 100mm2 die may be in the $5-10 range. And a package adds $5 on top of that. And that's for the entire die, not sections of it.

What's expensive is the R&D expenses, building up fabs, keeping lights on, and marketing.

Say you have 100K chips, and you have 40K 2 core dies, 30K 3 core dies, and 30K 4 core dies. But the market demands that they want 50K 2 core dies and 20K 3 core dies. Then the 3 cores will be disabled to meet the demands of the 2 core market. Yes you can go and start selling the 3 core for the same price as 2 core, but you lose the revenue, because you'll nullify the price difference. So you just disable them, since it costs them very little.

@Thunder 57 Yes because it is a shady marketing practice. It's one thing buying a product set like that from the factory, it's another to sell you a perfectly working chip disabled only to have you pay to enable them.

 

[Thunder 57]

It seems and may be wasteful but you have to understand from their perspective.

The cost of silicon itself is very little. A 100mm2 die may be in the $5-10 range. And a package adds $5 on top of that. And that's for the entire die, not sections of it.

What's expensive is the R&D expenses, building up fabs, keeping lights on, and marketing.

Say you have 100K chips, and you have 40K 2 core dies, 30K 3 core dies, and 30K 4 core dies. But the market demands that they want 50K 2 core dies and 20K 3 core dies. Then the 3 cores will be disabled to meet the demands of the 2 core market. Yes you can go and start selling the 3 core for the same price as 2 core, but you lose the revenue, because you'll nullify the price difference. So you just disable them, since it costs them very little.

@Thunder 57 Yes because it is a shady marketing practice. It's one thing buying a product set like that from the factory, it's another to sell you a perfectly working chip disabled only to have you pay to enable them.

They do that all the time except there is no way to enable them. It's not just Intel. Have you heard about paying monthly fees for heated seats in cars? Everyone is trying to sell subscriptions these days.

 

[igor_kavinski]

Well, they could put 3-core chips in the 2-core boxes and let people win silicon lottery