Nvidia 3080/3090 shortage expected to last until 2021

[Bombastinator]

38 minutes ago, Rohith_Kumar_Sp said:

eh... at this point i rather wait for 3080Ti next year. 

I suspect a 3080ti will only happen if AMD beats the 3080

[Rohith_Kumar_Sp]

1 minute ago, Bombastinator said:

I suspect a 3080ti will only happen if AMD beats the 3080

which i hope it does. wishful thinking. 

[Bombastinator]

9 minutes ago, Rohith_Kumar_Sp said:

which i hope it does. wishful thinking. 

What I want to know is if a PC will be able to play games written for consoles, and if so what will be required to do it.  Faster than that I kind of don’t care.

[RejZoR]

14 minutes ago, Bombastinator said:

What I want to know is if a PC will be able to play games written for consoles, and if so what will be required to do it.  Faster than that I kind of don’t care.

Xbox would be the most direct comparison since it's running Windows. You need to calculate in some losses since games are not written and optimize specifically like they are on consoles.

[AnonymousGuy]

2 hours ago, porina said:

I love how some think predicting the future is easy. It is not.

 

On the "why not ramp up production" part, again that is not like turning on the tap. Companies would have to forecast their manufacturing early on. Even without the GPU, nvidia could provide partners with design info so they know what components they need and order it in advance. To get components in volume at a good price, you typically end up with relatively long lead times, potentially months. You don't want to order excess components over your need, because if you don't use them, you either bin them or have to store it, which is a cost either way. It may be possible to get it quicker if you pay more, but that means either you make less profit or increased user pricing. And this isn't just for one component, you have to do it for every single used component on the design. Problems getting a single one means you can't complete. Alternates may be available in some situations. They will have an approved parts list so they may have alternatives to look at, but doesn't resolve the rest of the problems.

Fun fact, the current cycle time (aka the time to fab a chip start to finish) is up beyond 14 weeks now.  Product planning starts around 3 years before launch.

 

I think off the top of my head thery'se also getting <about 75 die per wafer and for sure there isn't 100% yield.

 

And there's also probably issues with GDDR6X supply (HBM I know has supply issues) and total number of wafers you can shove through the fab on this leading node process.

[porina]

1 hour ago, AnonymousGuy said:

I think off the top of my head thery'se also getting <about 75 die per wafer and for sure there isn't 100% yield.

I just played with a yield calculator. Of course, I don't have all the exact numbers to put in, but we can make some assumptions as follows:

12 inch wafer is pretty much a given

GA102 die size of 26.8x23.4mm

Defect rate is the big unknown value. I'm assuming 0.1/cm² because that's the ball park area TSMC 7nm and 5nm is sitting at. I don't have any other reference value to compare with.

 

With optimum layout for number of dies you can squeeze 88 on there, with an estimated 55% yield. Now, 55% yield might sound horrific but remember the 3080 uses a cut down configuration of the die. GA102 has 84 SMs, of which the 3080 uses only 68, or about 81%. Not all the die is just the cores, there will be non-core stuff like the memory controller and IO as well, but if they map out defects that happen in the unused cores the actual yield would be much improved. Looking up a die shot, without attempting to measure area, I'd eyeball the core vs non-core to be about equal. So if a defect happens, it'll be roughly 50/50 if it is in an area that could be easily mapped out, and might put the "3080 grade" yield somewhere above 75%. This assumes a single defect hits a die. The average rate is well below 1 per die area, but by chance you could get two or more per die. If the defect rate is higher than 0.15/cm² then it starts to get more interesting as that's when on average you'd expect one defect per die.

 

The above simply assumes a work or not work level. It doesn't go into clock yields at all, which may be a separate factor that can't be predicted. Presumably they can also bin the "good" SMs and only enable the fastest ones for a given product offering. That doesn't leave much wiggle room for the 3090 since that only has one disabled SM. I also understand that dies in the middle of the wafer tend to be better performing than ones on the edges due to the optical processes used. 

 

 

[AnonymousGuy]

7 minutes ago, porina said:

The above simply assumes a work or not work level. It doesn't go into clock yields at all, which may be a separate factor that can't be predicted. Presumably they can also bin the "good" SMs and only enable the fastest ones for a given product offering. That doesn't leave much wiggle room for the 3090 since that only has one disabled SM. I also understand that dies in the middle of the wafer tend to be better performing than ones on the edges due to the optical processes used. 

They probably targeted the process and set final clocks based on some 3 sigma so they're not tossing die due to lack of performance.  They also fabbed redundancy into the chip (like ECC for caches, repairable rows and columns, etc) so minor defects aren't instant toss outs.  

 

But yeah the bottom line is these die are f-ing huge and expensive for how much effort it takes to make them.  A chip this big Intel would be selling for $5000 with E7/Platinum Xeons, not $800.  Nvidia would probably rather get to the literal moneyshot of just selling a ton of 3070s where the revenue per wafer is significantly better, and they also can go full lumberjack cutting it down to 3060s and 3050Ti's and mobile parts and crap for defects.  So it's a balance for them of "we'd rather make the cheaper, more profitable stuff, but our halo products are attention getters".

[Valentyn]

2 hours ago, Rohith_Kumar_Sp said:

which i hope it does. wishful thinking. 

If somehow it does, and I couldn't snag a 3090 before then; I'll take what AMD has.

My workstation need a massive upgrade. Zen 3, and whatever top GPU i can snag.

[ewitte]

I'm not sure what this measures (current tests or total) but I was on userbenchmark the other day and noticed that 3080 numbers on there were already approaching those that had a 2080TI.  If that is total there actually may have been a decent amount of cards sold just the demand is so high. Also IMO demand is high because such a large % are going to bots and scalpers.

[Loote]

4 minutes ago, ewitte said:

I'm not sure what this measures (current tests or total) but I was on userbenchmark the other day and noticed that 3080 numbers on there were already approaching those that had a 2080TI.  If that is total there actually may have been a decent amount of cards sold just the demand is so high.

It most likely means that 3080s get tested more often nowadays, 2080ti has 186k tests and 3080 a bit over 3k and I believe 90% of tests happen right after purchase anyway.

regarding 3080Ti, imo 3090 being 10-15% faster than 3080 makes the Ti very unlikely, or pointless. I'd sooner expect Ampere+ with some fixes or maybe on TSMC 7nm if they need improvements, those could be supers or whatever.

[Awesome84]

I wonder how AMD is going to handle this.

[Maticks]

My best friend is convicened that if he checks his oven the 3080 will arrive. it worked for Jensen.

I told him he isn't doing it right, it will only appear if you are wearing the correct leather jacket.

[AlTech]

5 hours ago, poochyena said:

this, this is exactly what they said happened.

MLID was saying there was less than 20K cards available worldwide.

 

That's not underestimation. That's Nvidia being incomeptent.

[AlTech]

7 hours ago, Random_Person1234 said:

Summary

Nvidia's CEO expects the shortage of RTX 3080's and 3090's to last until early 2021. Jensen also said that the issue was with demand and not supply.

I'm calling BS.

 

Nvidia tried to make supply small so that they could bait and switch jack up the prices and blame it on AIBs.

 

That plan has backfired already, the sentiment of Ampere is significantly worse than it was on the announcement date.

[BuckGup]

Firstly - Nvidia is full of shit as usual

Secondly - I wonder if adding in all the extra bits like RT cores and TPUs makes it much harder to produce so the yields are lower. I'm not an expert but I feel pushing 28 billion transistors in such a small area is not an easy task.

[Guest]

38 minutes ago, AluminiumTech said:

Nvidia tried to make supply small so that they could bait and switch jack up the prices and blame it on AIBs.

Why are people in the industry, aka AIBs, saying they had normal ammount of supply at launch. Even AIBs say that demand was higher or comparable than Black Friday demand which was not expected to be the case.

[BuckGup]

47 minutes ago, Awesome84 said:

I wonder how AMD is going to handle this.

By releasing a card for $50 more than the RTX3080, which they'll drop the price in 4 months, that outperforms it in everything except ray tracing. Which every PCGamerMasterRace will only care about and say the card is useless lol. Honestly if this is true all AMD has to have is a card that trades blows with the RTX3080 and be in stock

[BuckGup]

1 minute ago, Medicate said:

Why are people in the industry, aka AIBs, saying they had normal ammount of supply at launch. Even AIBs say that demand was higher or comparable than Black Friday demand which was not expected to be the case.

I think either they are underestimating how many are being bought for mining. As of now the 3090 is the most profitable card to mine on with a close second with the RTX3080. Dethroning the previous RTX Titan and Tesla V100s. Cards that cost 5X that of the 3090. 

[derr12]

Also jensen;

 

"I can't pleasure my wife sufficiently, but it isn't what i can supply, its what she demands"

 

"tried the devils inch and everything!"

[Helpful Tech Witch]

6 hours ago, RejZoR said:

Xbox would be the most direct comparison since it's running Windows. You need to calculate in some losses since games are not written and optimize specifically like they are on consoles.

isnt the xbox using basically a consumer cpu and a consumer CPU? the only thing I can think of is that the xbox would use ddr6. 

[JZStudios]

8 hours ago, porina said:

I love how some think predicting the future is easy. It is not.

 

On the "why not ramp up production" part, again that is not like turning on the tap. Companies would have to forecast their manufacturing early on. Even without the GPU, nvidia could provide partners with design info so they know what components they need and order it in advance. To get components in volume at a good price, you typically end up with relatively long lead times, potentially months. You don't want to order excess components over your need, because if you don't use them, you either bin them or have to store it, which is a cost either way. It may be possible to get it quicker if you pay more, but that means either you make less profit or increased user pricing. And this isn't just for one component, you have to do it for every single used component on the design. Problems getting a single one means you can't complete. Alternates may be available in some situations. They will have an approved parts list so they may have alternatives to look at, but doesn't resolve the rest of the problems.

This all may be true, but it seems like every launch they only make roughly 150 cards globally and then be like "We weren't expecting anyone to buy it, this is totally unprecedented. Except for the last 4 launches that we've done that turned into the same shit show of supply drought. I mean overwhelming demand."

[RejZoR]

1 hour ago, HelpfulTechWizard said:

isnt the xbox using basically a consumer cpu and a consumer CPU? the only thing I can think of is that the xbox would use ddr6. 

The thing is, there are two versions of Xbox and games are coded specifically to work great on that. PC's have 700 trillion variants. And that causes performance issues and there is no way devs can make any use of speedy SSD as data streaming coz most PC's still run on mechanical HDD's despite the fact small SSD's sell for peanuts.

[Helpful Tech Witch]

Just now, RejZoR said:

peanuts.

eh, you can get 2tb hdd for the same price of a 500gb sata ssd, so its still worth it for 4x the size for a small speed loss.

[MadyTehWolfie]

12 hours ago, Bombastinator said:

If there are people waiting for the 3080 and 3090 it’s not demand by definition. 

It's high demand low quantity supplied. 

[RejZoR]

4 minutes ago, HelpfulTechWizard said:

eh, you can get 2tb hdd for the same price of a 500gb sata ssd, so its still worth it for 4x the size for a small speed loss.

"small speed loss"

 

On what planet are you living where any HDD is even virtually close to any SSD? The speed difference is MASSIVE. There is just no comparison. And I'm telling you from personal experience where I always had top of the line HDD's in laptops and desktop. Now I have SSD's in everything and there is just no comparison. Not even the fastest 7200 RPM HDD's can come anywhere close to SSD's. Maybe the worst DRAM-less SSD's. And even that only at write speeds. With reads they should be still much faster.