NVIDIA GeForce RTX Ampere Chips Feature Three Binning Tiers, Mostly Good Dies are Present

[mtcn77]

Because you're more qualified at choosing a foundry than NVIDIA, who has been doing this for decades?

Wait... Apple? I like the sound of where this is going. Here we go...

 

[BoboOOZ]

Interesting write up, but what is ignored here is that Samsung is not giving Nvidia all the chips, it's giving them only the good ones (which is an unusual deal, but hey, samsung must've wanted this badly) , because they have a special deal going on.

As such, it's impossible to evaluate how good are the yields at Samsung, because we don't know how many chips they throw away.

However, the one thing that we could infer is that most of the cards will make good overclockers/undervolters (which admitedly was already known/laked a long time ago) if the AIB or individuals are willing to cope with the high TDP's.

For those of us looking at these cards, this Samsung 8nm process is making me a bit nervous compared to TSMC 7nm...

Why is that, it seems 70% of the chips are more than OK? Those who are OK can still run as is written on the box, so this is a great deal for Nvidia.

 

[TheoneandonlyMrK]

so 30% bin 0= FE
60% bin 1= AIB stock speeds+ min. OC
10% bin2 = high OC version

???

What no the opposite from history top bin 0 goes to Fe everyone else gets what's left.

Soooooooo small run ,new node, big chip's 4 bins ,3 !good! 1 bad.

Sounding like a great paper launch this one.

 

[Grigor]

How is this news

 

[Xaled]

Bin 0 = over the announced price = overpriced 1600~1700$
Bin 1 = way overpriced 1700~1800$
Bin 2 = exaggeratedly overpriced ~2000$

 

[The Foldinator]

We desperately need a roll eyes emoji in the list here. lolol

Hater gonna hate man... HGH. Studies show. lololol

whut..sorry \EDIT/ aaaah got you

 

[EarthDog]

whut..sorry

the "like" list.

 

[robot zombie]

Water. It's still wet.

...but according to the internet, that in itself is a topic of intense debate! Its like, a microcosm of all speculative discussion online. Some say water is implicitly wet at all times. Others say wetness cannot be a property of water but instead only exists as an attribute of objects that have a sufficient amount of water on them.

Arent you glad we can have these illuminating conversations on this ever expanding highway of information and ideas? Stuff like this... we're redefining mankind as we speak. Discussions for the ages.

 

[SQLG0d]

"The production period was quite short..."

That is the only thing of consequence in the OP, and it isn't good for ANYONE - as opposed to the 10% statement affecting just those that want the "rare" top-tier cards. It would seem to indicate that Supply and Demand is gonna suck for consumers for longer than usual. I have no problem waiting though, as I am fortunate enough to be sitting pretty with a very nice gaming computer[(liquid-cooled 3900X/2080 Super (both of which have nicely undervolted and overclocked), fast NVMe drive, etc] and a bunch of other tech stuff that I got for free. I'm actually planning on selling everything but the computer: Vertigear gaming chair, Lexmark cx331 color laser multifunction machine, and Dell: Latitude 7400, Thunderbolt dock, and last but certainly not least a 34" Curved Ultrasharp monitor (which I successfully tested at an overclocked 75Hz). With the proceeds from that stuff I hope to be able to purchase a very nice monitor (I would love a Samsung G9 if they get the flickering fixed and start producing it again) and an RTX3080 . . . or something better should AMD impress with their release.

 

[Berfs1]

This is very little useful information.
- Which die are we talking about?
- What exactly do tiers represent - is Tier 0 something like Turing's non-oc cores that run at or slightly above spec?
- There is no comparison with binning results of some other die or some other manufacturing process.

Reread the article. It is VERY useful.

1. The article states Ampere chips, referring primarily to GA104 and GA102 (for now). It is unlikely they are talking about GA100 because that has already been out for a while and is not in this segment.
2. Tiers are already explained in the article, yes T0 is the ones that, at minimum, meet the specifications. Considering how silicon lottery plays, there will be some chips (actually a majority of these) that will exceed the binning spec for T0, but not enough to make it to T1, so yes, it is very likely you will have some overclocking leeway with T0, but don't expect something crazy from those. With the numbers given, you have a 3:6:1 ratio of T0:T1:T2. Which indirectly means, NVIDIA actually could have raised the specifications of the 30 series, because they have way more T1 chips than T0 chips, but they are playing it safe to get an extra 43% yield (30/70). What is likely to happen is, you will see more than half the cards (70% based on these numbers) will be able to overclock past NVIDIA's spec, and since there are a lot of T1 chips, it is likely those will be at MSRP +-10%. I expect a majority of T0 chips to be on the graphics cards that will be sold below MSRP levels. T2 GA104 chips may be reserved for the future GA104 graphics cards that enable more cores, because they are a top bin. In total, 60% will be T1, 30% will be T0, and 10% T2. When compared to T2, there are 3x as many T0 and 6x as many T1 chips. When compared to T0, there are 2x as many T1 and 0.33x as many T2 chips. When compared to T1, there are 0.5x as many T0, and 0.17x as many T2 chips. Make of that what you will.
3. Why would NVIDIA (or even Intel or AMD) want to tell you their binning results? They make the specifications for those chips, they don't want warranty claims, which if they do make public about their binning results, then people will start claiming warranty because it "doesn't meet the binned specifications". That is primarily why most companies do not tell you the binned specs, rather the specs of what the minimum in a set can handle. For AIBs, they put their own specs of what expect a majority of graphics cards in a set can handle. There is still some bad apples, and those can be RMA'd. A company like Silicon Lottery is the kind of company that will tell you the more exact binned specs, though they only do CPUs.

Bin 0 = over the announced price = overpriced 1600~1700$
Bin 1 = way overpriced 1700~1800$
Bin 2 = exaggeratedly overpriced ~2000$

I would expect something more like this:

1. T0 = 1300-1400$
2. T1 = 1500-1800$
3. T2 = 1900$+

Remember, NVIDIA's FE cards are likely to have T1 chips, as they have a lot of them. Sooooo, it is also likely that the bare minimum spec chips will be put in graphics cards priced under MSRP, by about 10%, in addition to a weaker but cheaper cooling solution (such as blower).

How is this news

It provides a rare insight into the binning process at the factory level, which is actually quite informative of expected overclocking performance.

 

[R-T-B]

Where could we have seen this? 3... 2... 1...

You guys act like the process node is going to make a difference vs AMD's historic incompetence with energy consumption. I doubt it will, frankly.

How is this news

Because it is? Some of us care?

Go educate yourself , read some reviews check some facts them come to talk me later about BS !

Because Ampere reviews are everywhere just now.

People, can we turn down the brand war just a bit? Please?

 

[DeathtoGnomes]

Water. It's still wet.

according to this read, its only 30% wet.

 

[saki630]

Bin 0: 144FPS
Bin 1: 149FPS
Bin 2: 155FPS
KingBin: 156.6FPS

 

[Vayra86]

Go educate yourself , read some reviews check some facts them come to talk me later about BS !

Non FEs also performed stellar? Yes, on air, and if they were the A chip, they were good on water too. I've never stated otherwise. Although the FE cards had a higher change of getting the A chip, especially considering that they cost as much as the cheapest AIB cards that were virtually guaranteed to be the non-a chips.
The non-a had a 280W power limit even with BIOS modding, while the A chips could go upto 380W with BIOS modding. That's another 100W of headroom for overclocking, which was huge for watercooling. If we're talking air, it's not relevant as you'll face thermal restraints way before power, but for those of us who put them under water the A chips were definitely worth the premium, and thus the FE cards were a good bet as they were cheap and virtually guaranteed to be the A chip.


Because they performed better out of the box as they had the superior A chip?

'Educate yourself' is not a source, boys. Try to find one. I challenge you to find an FE bench that scores higher than an equal open air AIB card at the same TDP.

Don't parrot marketing. Investigate. What you can do with mods and water is irrelevant. This is about what these cards are sold as. Since Kepler, I haven't seen a SINGLE FE / NVTTM card bench better than a non-FE. In fact, all FE's have throttled until Turing, and Turing's FE's had a markup and were the first open air versions and also carried higher TDPs than the equivalents from AIBs.

Even so, an MSI Trio still scores better than a Turing FE. Consistently.

Enjoy finding those sources for me. And note, a reviewer saying the same as you doesn't count. Numbers or its all nonsense.

Note that 1-2 FPS gaps don't count, because you've got that additional TDP. This does not prove you have a better bin, might even be contrary

And of course they are selling them with the same price, regardless of this "bin" thing ...

This procedure is well known, always has been like this.

Thank you. And there is no reason to start caring now. The overall performance spread is so low, this is marketing more than anything else. Create a 'need' where there isn't one, and look at us... (well, not me).

 

[mtcn77]

You guys act like the process node is going to make a difference vs AMD's historic incompetence with energy consumption. I doubt it will, frankly.

Thanks for the company, dude. I couldn't live with myself without some juicy green team red team beef, although I got no 'steak' in the discussion.
I don't post these things just because I don't like Samsung, it is to spite you wonderful gentlemen with incendiary comments. Because it is not about Samsung's process, but irrational teamster preferences.
I better take my opinion, elsewhere... nobody can see without their colored goggles here.

 

[londiste]

Reread the article. It is VERY useful.

1. The article states Ampere chips, referring primarily to GA104 and GA102 (for now). It is unlikely they are talking about GA100 because that has already been out for a while and is not in this segment.
2. Tiers are already explained in the article, yes T0 is the ones that, at minimum, meet the specifications. Considering how silicon lottery plays, there will be some chips (actually a majority of these) that will exceed the binning spec for T0, but not enough to make it to T1, so yes, it is very likely you will have some overclocking leeway with T0, but don't expect something crazy from those. With the numbers given, you have a 3:6:1 ratio of T0:T1:T2. Which indirectly means, NVIDIA actually could have raised the specifications of the 30 series, because they have way more T1 chips than T0 chips, but they are playing it safe to get an extra 43% yield (30/70). What is likely to happen is, you will see more than half the cards (70% based on these numbers) will be able to overclock past NVIDIA's spec, and since there are a lot of T1 chips, it is likely those will be at MSRP +-10%. I expect a majority of T0 chips to be on the graphics cards that will be sold below MSRP levels. T2 GA104 chips may be reserved for the future GA104 graphics cards that enable more cores, because they are a top bin. In total, 60% will be T1, 30% will be T0, and 10% T2. When compared to T2, there are 3x as many T0 and 6x as many T1 chips. When compared to T0, there are 2x as many T1 and 0.33x as many T2 chips. When compared to T1, there are 0.5x as many T0, and 0.17x as many T2 chips. Make of that what you will.
3. Why would NVIDIA (or even Intel or AMD) want to tell you their binning results? They make the specifications for those chips, they don't want warranty claims, which if they do make public about their binning results, then people will start claiming warranty because it "doesn't meet the binned specifications". That is primarily why most companies do not tell you the binned specs, rather the specs of what the minimum in a set can handle. For AIBs, they put their own specs of what expect a majority of graphics cards in a set can handle. There is still some bad apples, and those can be RMA'd. A company like Silicon Lottery is the kind of company that will tell you the more exact binned specs, though they only do CPUs.

- GA102 and GA104 are quite different, most importantly 628mm² vs 395mm² on a pretty unknown manufacturing process. Bunching them up together makes no sense. Or, alternatively if it makes sense then Samsung's 8N process is absolutely excellent if yields are similar for both dies, one being almost 60% bigger than the other and close enough to the reticle limit.
- OK, Good, Best is absolutely pointless division unless we know what each means and we do not. We do not have a single clue. Are they 2% apart? 10%? Are they binned for speed or efficiency? Are they binned for functional units? Something else? The rest is fairly baseless conjecture.
- Binned specifications have no relevance on product specifications and if they did reveal binned specs publicly they would be in full right to point at product specs and say - it achieves that. End of story.

 

[dink]

'Educate yourself' is not a source, boys. Try to find one. I challenge you to find an FE bench that scores higher than an equal open air AIB card at the same TDP.

Don't parrot marketing. Investigate. What you can do with mods and water is irrelevant. This is about what these cards are sold as. Since Kepler, I haven't seen a SINGLE FE / NVTTM card bench better than a non-FE. In fact, all FE's have throttled until Turing, and Turing's FE's had a markup and were the first open air versions and also carried higher TDPs than the equivalents from AIBs.

Even so, an MSI Trio still scores better than a Turing FE. Consistently.

Enjoy finding those sources for me. And note, a reviewer saying the same as you doesn't count. Numbers or its all nonsense.

Note that 1-2 FPS gaps don't count, because you've got that additional TDP. This does not prove you have a better bin, might even be contrary



Thank you. And there is no reason to start caring now. The overall performance spread is so low, this is marketing more than anything else. Create a 'need' where there isn't one, and look at us... (well, not me).

It has been common knowledge among extreme OC crowds and bench score ninjas that the FE series are the best at release. Later in the lifecycle it doesn't matter as much. If that trend continues is speculation, but why do you care anyways? You clearly aren't in those circles.

 

[Berfs1]

- GA102 and GA104 are quite different, most importantly 628mm² vs 395mm² on a pretty unknown manufacturing process. Bunching them up together makes no sense. Or, alternatively if it makes sense then Samsung's 8N process is absolutely excellent if yields are similar for both dies, one being almost 60% bigger than the other and close enough to the reticle limit.
- OK, Good, Best is absolutely pointless division unless we know what each means and we do not. We do not have a single clue. Are they 2% apart? 10%? Are they binned for speed or efficiency? Are they binned for functional units? Something else? The rest is fairly baseless conjecture.
- Binned specifications have no relevance on product specifications and if they did reveal binned specs publicly they would be in full right to point at product specs and say - it achieves that. End of story.

Dude, not even a lot of professional overclockers know exactly how well CPUs and GPUs are binned from the factory. TBH, you are expecting way too much. As for the manufacturing process, I don't know, it's kind of like techpowerup has it in their own GPU databases that they use Samsung 8nm, it's also kind of like NVIDIA actually stated that in their own 30 series announcements. At this point, I think you are just trolling, so I will stop replying to you.

 

[londiste]

Dude, not even a lot of professional overclockers know exactly how well CPUs and GPUs are binned from the factory. TBH, you are expecting way too much. As for the manufacturing process, I don't know, it's kind of like techpowerup has it in their own GPU databases that they use Samsung 8nm, it's also kind of like NVIDIA actually stated that in their own 30 series announcements. At this point, I think you are just trolling, so I will stop replying to you.

That was my entire point that this story yields very little useful information. If any.
I have no idea what you mean by talk about manufacturing process here.

Edit:
OK, I am note sure what or how exactly but you read my manufacturing process comment somehow completely wrong.

Samsung 8N or 8nm or even 10nm are very much unknown quantities when we are talking about big dies and desktop hardware. 8nm and 10nm have been used primarily by mobile SoCs. This means both low power and smaller dies (probably in the range of maybe 150mm²). There are no data points we know about nowhere near the 300W and 400/600mm² targets that are Ampere dies.

 

[R-T-B]

it is to spite you wonderful gentlemen with incendiary comments

My comment was not intended as incendiary, but thanks. I'm just looking at the numbers.

 

[mtcn77]

My comment was not intended as incendiary, but thanks. I'm just looking at the numbers.

Well, opinions are quantitative now?
You guys are giving me the nosebleeds.
I don't mind fans and giggles from time to time, but maybe we should leave facts do the speaking.
PS: I wasn't speaking for yourself, I was acting ironic as usual. You guys should know astroturfing is my fine honed art.

 

[R-T-B]

Well, opinions are quantitative now?

No, but watts are.

 

[mtcn77]

No, but watts are.

Funny you would go there. Samsung isn't your bread and butter foundry. Expect a few hiccups.

If we continue, I'll have to resort to Mortal Kombat jokes;
- "You are still trying to win?"
Let's not go there pal, I've never missed my common troll courtesy.

 

[R-T-B]

You assume I am trolling. I'm not. You'll end up playing with yourself.

I just can't see the watts on Samsung 7nm being worse than Turing, which is about what it'd need to matter I think.

 

[mtcn77]

I just can't see the watts on Samsung 7nm being worse than Turing

Hmm, never thought it that way. It is quite an uncommon way of thinking. Turing is neither the competition, nor made at the same foundry, but hold on to your beliefs, I guess...