Discussion Variance in CPU quality and how it affects reviews and user experience

[coercitiv]

Der8auer tested multiple R5 7600 CPUs and found up to 50% difference in efficiency while gaming.

From the start everyone should know this is not related to AMD, but rather a normal occurrence in all CPUs, and the variance is obviously bigger in budget SKUs. In the past, the same youtuber did a similar test with the more premium 10900K and found a 20% difference between the samples. One can easily see why looking at other budget SKUs could have produced the same 40-50% delta we see with the 7600.

A few snapshots from the video bellow.

Spoiler: Power and Clocks

Spoiler: Voltage and Clocks

Spoiler: Power and Temps

And finally the gaming efficiency comparison with the big numbers to impress people into reading more on the subject.

Spoiler: Gaming efficiency

 

[Kocicak]

You can see that the manufacturers should remove the lemons and also golden samples from the samples sent out to reviewers. Both extremes would have a negative impact to the manufacturer.

I have not watched the video, I am not sure if I am not saying the same thing as was in the video.

 

[gdansk]

If you watch the video he pretty much covers it all

It is inevitable with modern manufacturing and expect catch-all SKUs to have more variance.
The chips are still within spec but need more voltage to reach it.
Do not go by a single review source, check the big picture across many reviews.


Perhaps be skeptical of review samples and focus more on reviews that bought their own parts.

 

[Panino Manino]

"Never buy a product at launch".

Does he check and compare the manufacturing date to see if there's any improvement with time or still random?

 

[VirtualLarry]

expect catch-all SKUs to have more variance.

This. I expect that the Ryzen 5 7600 SKU "absorbs" all of the Zen4 silicon that fails being a higher-end chip, the exception being CCXs that only have 4-5 or less good cores, those are probably being stock-piled for a theoretic 7300 4C/8T SKU that may or may not occur in the future, like the Ryzen 3100(X) that was so mysterious and hard to find early on in the Zen2 days.

 

[eek2121]

"Never buy a product at launch".

Does he check and compare the manufacturing date to see if there's any improvement with time or still random?

This, IMO, is something that should be investigated. Does quality get better over time or does it continue to vary?

 

[gdansk]

"Never buy a product at launch".

Does he check and compare the manufacturing date to see if there's any improvement with time or still random?

I forget the details of the video but isn't it more of a disclaimer that manufacturers can influence launch day coverage with golden samples? Not that the chips necessarily get better over time.

It would be nice to see a graph of these ordered by manufactured date.

 

[GaiaHunter]

Nothing particularly new. Anyone that ever had multiples of the same product will know this. Started ever since we got boost clocks and variable voltages. I had 3 6700k all with different voltages to hit the same 4.5ghz clocks. Also don't expect all AMD CPUs to run at -30mV (my 7700x does) or all Intel CPUs to run DDR5 8000.

 

[GaiaHunter]

I forget the details of the video but isn't it more of a disclaimer that manufacturers can influence launch day coverage with golden samples? Not that the chips necessarily get better over time.

It would be nice to see a graph of these ordered by manufactured date.

The overall quality of the chips should increase if we are talking about a new process node or if there are respins of the chips. If the process is mature the quality of the chips should not differ noticeable.

 

[GaiaHunter]

dp...

 

[Panino Manino]

I forget the details of the video but isn't it more of a disclaimer that manufacturers can influence launch day coverage with golden samples? Not that the chips necessarily get better over time.

It would be nice to see a graph of these ordered by manufactured date.

Hum... shouldn't these companies always try to provide "golden samples" for launch review? Or as I could call it, "target samples". Because we know yields improve with time, some times so much that consumers may end with chips that perform above specs. Didn't we see cases like that?
Thinking about this I don't see the problem.
What would be a problem would be reviewers receiving subpar launch chips making consumers years later believe that the chip performs much worse than reality.

 

[eek2121]

The overall quality of the chips should increase if we are talking about a new process node or if there are respins of the chips. If the process is mature the quality of the chips should not differ noticeable.

This is actually untrue. Tooling wears out over time, fyi. Cannot comment on specific designs since that actually depends on the nodes and design in question. That is why I made my earlier post.

All of the major silicon companies are dealing with this.

 

[Hulk]

This is actually untrue. Tooling wears out over time, fyi. Cannot comment on specific designs since that actually depends on the nodes and design in question. That is why I made my earlier post.

All of the major silicon companies are dealing with this.

Huh? So yields get worse over time?

 

[JoeRambo]

I can't comment about tooling, but the biggest enemy of enthusiast is Intel binning. For example when Intel starts binning for those KS sku chips, obviuosly the rest of the line suffers. Same with "refreshes" that reuse same dies with different names.
The variance is huge between the chips as well. I had my paws on 3 13900KS and the gap between the best and worst ones was ridiculous.

Obviously this all gets amplified by motherboard vendors who are hell bent to avoid returns and stick to default configurations that overvolt...

 

[Kocicak]

Huh? So yields get worse over time?

Why not? It is only about timeframe. If you have say 4 year run of one technology and two improvements in it, you have (at least) three local peaks of yields, and also periods of stagnation and decline. You can imagine that say 10 month period of manufacture of one product can fit whole in the stagnation/decline period.

 

[Hulk]

Why not? It is only about timeframe. If you have say 4 year run of one technology and two improvements in it, you have (at least) three local peaks of yields, and also periods of stagnation and decline. You can imagine that say 10 month period of manufacture of one product can fit whole in the stagnation/decline period.

I has always thought as the engineers tweaked and figured out a particular process the yields got better over time not worse.

 

[Kocicak]

I has always thought as the engineers tweaked and figured out a particular process the yields got better over time not worse.

I do not believe that any continual "knob tweaking" and improvements are possible, if any notable process improvement is implemented, it itself needs some time to get to work to its full potential and after that no improvement is possible untill the next one. Not to mention that any change can bring unforseen problems and in short term worsen things instead of improving them.

So improving yields is valid in long term or in the phases of getting things working to their full potential, but not in every moment of the process life.

And there is a lot of older processes in use, which are only maintained and repaired to restore some standard performance and no improvements are happening at all.

 

[Mopetar]

It's not that unusual. Even the top bins will have some variation. I forget the name of the website, but years ago they would buy a lot of the top-end i7 CPUs, delid and lap them and then OC them and resell the CPUs based on how far they could be pushed. It wasn't unusual to see them have three or four new bins (in 100 MHz increments) for the chips they were selling.

I don't think they ever posted any of the results as far as power use or thermals went, but they did show the breakdown of how many of the CPUs they bought could hit each of the different OC bins.

Intel and AMD could add additional bins if they really wanted to, but it makes it more difficult for consumers to choose a product (it's a bit like trying to look through everything on NetFlix and being unable to decide what to watch) and if you don't get the pricing right, you can wind up with some lopsided sales and parts that no one wants because they're perceived as a bad value.

Edit: It was Silicon Lottery that I was thinking of. They still have the website up (even though they shut down) and lost the historical data for all the CPUs they sold, which is interesting to look through.

 

[moinmoin]

It was Silicon Lottery that I was thinking of. They still have the website up (even though they shut down) and lost the historical data for all the CPUs they sold, which is interesting to look through.

It's unfortunate (but understandable) that they shut down. I feel the increasing disparity in efficiency would warrant some further binning like Silicon Lottery did for frequency. As @gdansk said catch-all SKUs are bound to have more variance, and due to having fewer SKUs AMDs catch-all SKUs have even bigger variance.

 

[GaiaHunter]

Yields obviously get better overtime. Eventually they plateau.
If the tools get bad they are replaced/fixed.

Now AMD chiplets are so small that almost every chiplet on wafer will be functional.And those chips serve the entire product stack from desktop to servers.

So yeah, the 7600 are going to be absolute "trash".

 

[A///]

This was of course the end result of amd not wanting to develop mainstream quad cores with smt as a budget option and in turn only develop apus which iirc are lacking right now. I haven't watched myself but going by some of the posts particuarily by larry and gayahunter it's universally understood the the 7600x is the worst of the pack which beckons the questioning of why it isn't priced as such if its using trash graded silicon. It's very unfair for the end customer to pay a premium price over the competitor last i checked only to get a low end product. whether the competitor product is binned that far down or not is not something I know but perhaps others may.

this means max product for amd even on a die they might have thrown away in the past. possibly. if not using chiplets.

This, IMO, is something that should be investigated. Does quality get better over time or does it continue to vary?

improvements or additions come with steppings do they not?

 

[TheELF]

it's universally understood the the 7600x is the worst of the pack which beckons the questioning of why it isn't priced as such if its using trash graded silicon.

It's not trash graded, all of the CPUs are at least on the standard that is stated for them.
You can just get lucky and get a much better chip or one that is just on the standard.
Silicon lottery is real for every single CPU out there.
You pay to get the standard, anything above that is luck of the draw.

It's not like the 3000 series where they would promise boost clocks that most of the CPUs couldn't hit.

 

[A///]

It's not trash graded, all of the CPUs are at least on the standard that is stated for them.
You can just get lucky and get a much better chip or one that is just on the standard.
Silicon lottery is real for every single CPU out there.
You pay to get the standard, anything above that is luck of the draw.

It's not like the 3000 series where they would promise boost clocks that most of the CPUs couldn't hit.

Silicon lottery would only matter within the specific model grade would it not? we've established that the binning goes from epyc down to the 7600x. they would be the lowest quality silicon but still viable for amd to use and ship off. the general variance among that line can't be too great where silicon lottery may play a factor. my personal opinion from long ago still stands imo, if you're buying into zen 4 then or now and not wanting a faster model or the 7800x3d you're wasting money unless you plan on to upgrade to zen 5. the intel models at that range offered more, although I haven't kept up with my pricing models and the the 7950x is possibly on offer in combo for less than 600 these days invalidating intel's low end attempt to squeeze into amd's zone.

 

[Mopetar]

So yeah, the 7600 are going to be absolute "trash".

It's not a guarantee of trash, just that you can wind up with something that didn't qualify as a 7700 but is otherwise fine or something that barely qualifies as a 7600 and is only being put in that bin because there's no 7500 or 7400 bins or won't be for some time.

Quality distributions tend to be pretty close to normal, so your chances of getting a garbage chip are likely no higher than 15% and likely to be far lower than that. Some binning is done for economic reasons so you might get something that's effectively as good as a 7700X, but has had some cores sealed off.

 

[A///]

Not keen on those figures myself. AMD's economics are sound with little to no waste. the blessings of which will become more profound to their bottom dollar in the coming years as further node shrinks come to fruition. Pat "Guns Ablazing Gunslinger" Gelsinger needs to get on with his plan, get his head out of his tuckus and strop prancing around with the bs and put the pedal to the metal.

He's got his foot on the accelerator at 50% at the moment and needs to go wot to get the most out of his tenure before any seeds of doubt become sewn and intel finds itself in another circling mess like they did with the two prior bumbling idiots who may have signed off on alder lake or raptor but committed their own sinful deeds of bull.