nVidia 3090 reviews thread

[AtenRa]

Both VEGA 2 and RDNA 1 have

I'm curious to read about that. Do you have a link to where A100 is said to use a high density library at TSMC?

I dont have a direct article to link but there are two possibilities for the much higher A100 density over VEGA 20 and NAVI 10.

1. They use the HD Libraries
2. The design is giving priority to density vs higher clocks. But if you want higher density it is better to use the HD Libraries than using the HP Libraries. So we go to number 1 again.

So when we have two designs at 7nm HP process (VEGA 20 and NAVI 10) with density at 40Mt/mm2 that can reach 1800/1900MHz and then a chip (A100) with density of 65Mt/mm2 that only reach 1400MHz , we can understand there are fundamental differences either in the process or in the design. In this situation most of the data points to the process.

But even if it is the design of A100 that makes the density difference and not the Libraries, making it a low speed high density design, its is only logical that if they would port GA102 that can reach 1800-1900MHz, it would not have the same high density as A100 on the same process with the same Libraries but much lower and closer to VEGA 20 and NAVI 10 in order to keep the high speeds.

 

[JoeRambo]

GA102 seems to handle undervolt real well. @ ~0.8 volts chip seems to clock to ~1800 and that "1800" is just 3-5% less performance than typical power limited ~1800 operation. Except wattage is 100W less, clock is stable, temps are low and noise ear friendly easy on those cards overbuilt for north of 350W.

Seems like NV took too many pages from retarded operations AMD had over the years. Cards clocked way too high, voltage pushed too far, insane amount of FP32 resources not backed by enough raster to really shine in games.

 

[guidryp]

Everyone always seems to forget that A100 has 40 MB of L2 cache vs the 6 MB in GA102. That's going to inflate the transistor count of A100 as well.

And density. Cache tends to really pack them into a smaller area.

 

[Mopetar]

The day 1 supply chain stuff etc is just a curse of success. Very little they could sensibly do.

Well they could wait another month or two in order to stockpile inventory so that the initial supply is closer to the initial demand.

There are good reasons for them not to do that, so it's understandable why they would want to launch when they did, but let's not pretend that this is a problem with no possible solution.

Even if AMD has some real competition for the first time in half a decade the 3080 at $700 still offers amazing value, including the best bar none at 4K. NVidia would still be an attractive option for many gamers.

 

[Qwertilot]

They'd probably have to start the entire logistics operation months in advance, not just stockpile some extra chips. And how would people then react knowing there were boxed cards ready to go but not allowed to buy them in advance?

The real solution would lie in people behaving much more calmly and rationally but seemingly very little hope of that.

 

[Kenmitch]

Seems like NV took too many pages from retarded operations AMD had over the years. Cards clocked way too high, voltage pushed too far, insane amount of FP32 resources not backed by enough raster to really shine in games.

AMD did that to be more competitive with Nvidia's offerings. I guess we'll see soon enough the reasoning behind Nvidia's choices this go around....Of course we could all just blame Samsung and look the other way.

 

[PlanetJosh]

For those that will use air cooling on a 2080ti or above what's with some pre builts requiring you order liquid cooling installed before you can buy? It seems the Alienware/Dell site mandates you have liquid cooling on a pre built mid tower pc with a 2080ti. To help prevent returns on air cooled ones if they break down? Or something like that? For a 2060 it appears they allow the prebuilt to ship with just air cooling.

Anyway Cyberpower lets you order a number of fans for air cooling on a pre built with 2080ti or up. I'd prefer not to mess with liquid cooling for now. Or I should just build myself, cases come basically ready to pop in a mobo.

 

[MrTeal]

For those that will use air cooling on a 2080ti or above what's with some pre builts requiring you order liquid cooling installed before you can buy? It seems the Alienware/Dell site mandates you have liquid cooling on a pre built mid tower pc with a 2080ti. To help prevent returns on air cooled ones if they break down? Or something like that? For a 2060 it appears they allow the prebuilt to ship with just air cooling.

Anyway Cyberpower lets you order a number of fans for air cooling on a pre built with 2080ti or up. I'd prefer not to mess with liquid cooling for now. Or I should just build myself, cases come basically ready to pop in a mobo.

I can’t imagine it’s for reliability reasons, the pumps on an AIO are always going to be at higher risk of failure than a fan. I would guess it’s for thermals and acoustics in the cases they want to use.

 

[loki1944]

Is it me or had nvidia kinda botched this launch? 3070 only having 8GB in a late 2020/early 2021 card, a powerhouse 3080 but limited in the future by vram, and now a 3090 that draws so much power it may give a boost to the power industry.

If they make a profit, it's a win for them. To me everything is going to plan for Nvidia. You're also hard pressed to breach 10GB VRAM, even @4K.

 

[PlanetJosh]

Hmm that 8k tv above. Well that's one way to get more detail in games on a 55 inch screen. Maybe I'll try it on a single 3090 if the card is ever available again.

 

[linkgoron]

Is it me or had nvidia kinda botched this launch? 3070 only having 8GB in a late 2020/early 2021 card, a powerhouse 3080 but limited in the future by vram, and now a 3090 that draws so much power it may give a boost to the power industry.

I'm pretty sure that Ampere on 8nm didn't deliver what Nvidia expected. GA104 is the first 04 die that isn't faster than the previous 10/00/02 GPU in ages. It has the first 80 card that uses a big die since Kepler, and I'm not sure if Nvidia ever released a card that at stock used over 300 TBP (non-dual card, anyway), especially not a 80s card.

 

[EXCellR8]

Interesting stuff from a PCB designer: On Reddit

 

[Mopetar]

They'd probably have to start the entire logistics operation months in advance, not just stockpile some extra chips. And how would people then react knowing there were boxed cards ready to go but not allowed to buy them in advance?

The real solution would lie in people behaving much more calmly and rationally but seemingly very little hope of that.

So it's more feasible to change human nature than it is your product schedule?

Im reminded of an old saying: Wish in one hand and crap in the other. See which one fills up first.

Obviously no company can get every launch just right. There's too many unknowns that make it easy misjudge and having inventory you can't shift is far worse than not having enough. That said with the unprecedented levels of demand NVidia saw there's not a lot of good arguments for shipping early with almost no supply. In encourages the awful behavior that people complain about. NVidia even apologized. Do they strike you as a company that would apologize for their successes? Historically they're not even a company that likes to apologize for their mistakes. I think that says a lot about the current situation.

 

[guidryp]

That said with the unprecedented levels of demand NVidia saw there's not a lot of good arguments for shipping early with almost no supply.

All indications are that they had normal launch supply. Even if they had doubled that, it might have accomplished what? Lasting 2 minutes, instead of 1 minute?

Your previous post suggested stockpiling for 2 months, which is utterly absurd way to do business. Modern manufacturing aims for JIT, not spending money on Warehousing for months.

That doesn't get cards to customers faster, and in fact gets them slower.

 

[Mopetar]

Interesting stuff from a PCB designer: On Reddit

Although this is incredibly informative, I liked it better when I could just think of the POScaps as the crummy ones.

All indications are that they had normal launch supply. Even if they had doubled that, it might have accomplished what? Lasting 2 minutes, instead of 1 minute?

Your previous post suggested stockpiling for 2 months, which is utterly absurd way to do business. Modern manufacturing aims for JIT, not spending money on Warehousing for months.

That doesn't get cards to customers faster, and in fact gets them slower.

The levels may be normal or close to it, but that's just historical data. Whether they actually hit those numbers or not, I think it's fair to say that the demand has increased and grown the size of the market for new NVidia cards. The very first iPhone sold somewhere near 1.5 million units in its first two quarters of availability. Apple would be raked over the coals if they tried to only ship that many units in the first ~six months for their upcoming iPhone. I think the sales expectations for just the first quarter are a full order of magnitude (15 million) higher at the least now. I would imagine Apple stockpiles components or even completed devices in the lead up to that. You should go tell the most profitable company on the planet that they have an utterly absurd way of doing business.

As an interesting experiment, let's take your own argument and take it further. Shouldn't it have been even better for NVidia to launch two weeks earlier than they did? They'd have even fewer cards to stockpile and those cards would have made their way into the hands of customers even faster. Logically, isn't the best possible launch then effectively a paper launch where the very first card to roll of the production lines is packaged and shipped off to the lucky individual that managed to hit his or her F5 key faster than all of the other boys, girls, and bots? You can't get cards to customers faster than that!

But clearly NVidia has found the best way to do business. In fact, it's so good that they even felt the need to apologize for it. I think the last time they had to do that was with the GTX 970 when they similarly must have found the best way to deliver a 4 GB graphics card.

 

[guidryp]

The levels may be normal or close to it, but that's just historical data. Whether they actually hit those numbers or not, I think it's fair to say that the demand has increased and grown the size of the market for new NVidia cards. The very first iPhone sold somewhere near 1.5 million units in its first two quarters of availability. Apple would be raked over the coals if they tried to only ship that many units in the first ~six months for their upcoming iPhone. I think the sales expectations for just the first quarter are a full order of magnitude (15 million) higher at the least now. I would imagine Apple stockpiles components or even completed devices in the lead up to that. You should go tell the most profitable company on the planet that they have an utterly absurd way of doing business.

Poor argument given that overall GPU sales have been stagnating or declining for years:

 

[Markfw]

Poor argument given that overall GPU sales have been stagnating or declining for years:

That ends 5 years ago. What has it done in the last 5 years ?????

 

[Hitman928]

Poor argument given that overall GPU sales have been stagnating or declining for years:

A lot of that is because of APUs making it so that discrete cards are needed less and less. The market has shifted to APUs on the low end and by and large only what used to be called mid-range and above for discrete solutions. If you look at Nvidia's gaming revenue over the last 5 - 10 years, it's grown substantially due to being the clear market leader for a few years now but also because of this trend. There are more people buying $300+ GPUs than ever before and it's been trending that way for at least 3 - 4 years. I don't know how long it will sustain, but that's the most recent history.

 

[guidryp]

That ends 5 years ago. What has it done in the last 5 years ?????

Unfortunately, for some reason most people seemed sto stop updating graphs around then. If someone can find a better one, please do.

But this certainly isn't a case for GPU's are rapidly expanding new market.

Edit, here is one into 2019, which still shows overall declining trend:

 

[Hitman928]

That ends 5 years ago. What has it done in the last 5 years ?????

This graph goes through Q317 and shows AIB volume recovering in terms of volume.

But even more so, as I mentioned before, the market for AIB has shifted due to APUs becoming more and more effective at taking over the low end needs. So mid range+ add in cards are actually increasing significantly as their portion of the market and I believe the trend has only continued after 2016.

That's why Nvidia's gaming revenue has grown so much over the last 5 years or so.

And PC game sales have also experienced continued growth.

 

[guidryp]

This graph goes through Q317 and shows AIB volume recovering in terms of volume.

No doubt, since 2017 was peak Mining bubble, and you could hardly find a GPU in stock.

IIRC, they collapsed back in 2018.

 

[Hitman928]

No doubt, since 2017 was peak Mining bubble, and you could hardly find a GPU in stock.

IIRC, they collapsed back in 2018.

Check the rest of what I posted, the increase already started in 2016 and there was a bit of dip after the mining crash due to over saturation but once the next gen came out the upward trend continued.

 

[AtenRa]

That ends 5 years ago. What has it done in the last 5 years ?????

DT dGPUs are in a fall.
They bounce up a little bit in H1 2020

 

[deathBOB]

I don't see how total market numbers, which include huge numbers of low-end PCs, tell us anything about sales of high-end GPUs.

 

[guidryp]

DT dGPUs are in a fall.
They bounce up a little bit in H1 2020

Thanks, that makes the long term downward trend clear.

I don't see how total market numbers, which include huge numbers of low-end PCs, tell us anything about sales of high-end GPUs.

This whole diversion was begun when someone suggested having similar stock to the last launch was comparable to using first Year iPhone sales, as a basis for launching current iphones (which are 100x higher today).

This downward trend makes it abundantly clear that analogy/argument is complete nonsense. Discrete GPU sales are NOT a fresh new product space on a massive upward swing. In fact they are in long term decline.

I know a lot of people dislike the dominant position NVidia held in the market, and will use any reason to attack them, but sorry this one isn't a valid argument.