NVIDIA GeForce RTX 5090 PCB Pictured, Massive GPU Die and 16-Chip Memory Configuration

[Visible Noise]

Sorry, but your goal is no to "fit", but to have a PCB with the approximate size of the heatsink above it.

More nonsense. Is your cpu heatsink the same size as your motherboard?

 

[nexxusty]

Oh boy, we got a qualified PCB designer here.

3090ti didnt melt.

Not if you plug it in right.

So it's simultaneously too cramped but also too big?

Why, so you can cut the bus size in half and kneecap the chips performance?

You heard him, this big GPU is too big, dont even think about it!


If you see complaining about poor quality games and hardware as "right wing incels", it may be time for you to go outside and touch grass.

Just because you buy geforce cards to run some commercial software and dont play games doesnt meant hats how other people use it. People need to accept that.

3090/3090Ti, different size/amount of memory IC's, same bus width.

Oof.

 

[igormp]

So close to making a comment without making yourself look like a fool. So close bro.

3090/3090Ti, different size/amount of memory IC's, same bus width.

Oof.

The 3090ti came out when 16Gb GDDR6X became available, allowing nvidia to forgo the clamshell design of the 3090 with 2 memory ICs per channel to just a single one with double the density, but still had 12 of those whereas the 3090 had 24.
For the 5090, with its 512-bit bus, given that each controller is 32-bit, you need a minimum of 16 memory chips, period. Currently on GDDR7 the smallest memory ICs available are 16Gb ones, there are no 8Gb ones.

So yeah, to use less chips you'd need a smaller memory bus, which would decrease the performance.

 

[nexxusty]

The 3090ti came out when 16Gb GDDR6X became available, allowing nvidia to forgo the clamshell design of the 3090 with 2 memory ICs per channel to just a single one with double the density, but still had 12 of those whereas the 3090 had 24.
For the 5090, with its 512-bit bus, given that each controller is 32-bit, you need a minimum of 16 memory chips, period. Currently on GDDR7 the smallest memory ICs available are 16Gb ones, there are no 8Gb ones.

So yeah, to use less chips you'd need a smaller memory bus, which would decrease the performance.

Agreed. However, your comment came across as a universal truth. I was pointing out how there are instances where this isn't true.

Good response though. Obviously you know what you're talking about about.

 

[igormp]

your comment came across as a universal truth.

It is not, since this depends on the memory controller design. GPUs usually use 32-bit controllers, so you can divide the bus size by 32 to get the amount of memory chips needed in the pcb (double that for clamshell designs that want to increase capacity, like your 3090 example).
You regular x86 desktop has 64-bit (or 2x32-bit that kinda work as a single unit) controllers, so for your regular consumer CPU with 128-bit (the so called "dual-channel"), you need a minimum of 2 sticks.
Apple's Mx lineup uses 16-bit controllers, and so on and so on.

I was pointing out how there are instances where this isn't true.

Yeah, but you ended up talking about a different thing. The comment you were replying to was correct still: decreasing the amount of memory modules for the 5090 would imply in a smaller bus, and thus lower performance.

 

[trsttte]

Turing had no such restriction

Turing had a true Titan card - https://www.techpowerup.com/gpu-specs/titan-rtx.c3311 - with professional features unlocked, with Ampere they brought back the x090 series instead and kept professional stuff locked.

Huge déjà vu of all the arguments from 4 years ago about how the 3090 was not a Titan.

 

[igormp]

Turing had a true Titan card - https://www.techpowerup.com/gpu-specs/titan-rtx.c3311 - with professional features unlocked, with Ampere they brought back the x090 series instead and kept professional stuff locked.

Huge déjà vu of all the arguments from 4 years ago about how the 3090 was not a Titan.

And still all Turing products with tensor cores, from the 2060 up to the 2080ti, had the same tensor fp16:fp32 rates as the titan rtx.

The only "locked" feature on Ampere and Ada was tensor fp16 with fp32 acc, as I had said before, which is not really a "professional" stuff given how other stuff is "unlocked", but it is a market segmentation tactic nonetheless.

Regular fp16 performance has the same rates in both the GeForce and in their professional lineup, that's the point I wanted to correct in your original statement.

 

[SkypredatorGaming]

Quite low quality design. The thermal density will be high - 600 watts in so small area will be tough to keep cool.

1. The PCB will melt;
2. The single power connector will melt;
3. Wrong PCB size;
4. Too many memory chips - this needs either 3 GB or 4 GB chips.

Overall, given the $3000-4000 price tag - it is a meh. Don't buy.

No just 2GB Micron GDDR7 chips, no 3 or 4GB, if there is room to add 32x 1GB modules, it will be far better for memory efficiency, but here is an image of Gigabyte Aorus Extreme Air Version RTX 5090

 

[londiste]

x090 are Titan-class things. Ultra high end stuff regardless of naming. The features have and will vary based on what is available and what makes sense for the manufacturer to sell. FP64 (or FP16) ratios were dropped not to cannibalize the expensive Quadro/Tesla cards which makes perfect sense from Nvidia's point of view. It still left performance, VRAM and other bits and pieces to somewhat justify the price in addition to bragging rights. Whether these things are worth the price is wholly up to buyer. There are still professional use cases for x090 cards where their price is a bargain and also looks like everyone underestimates the wealth and "want" of enthusiast gamer crowd who did and will buy those for gaming.

If you find these to be obscenely overpriced and useless, that is OK. Others will have a different opinion, and that is also OK.

 

[TokyoQuaSaR]

Is the large size supposed to be impressive?

Both the die size (744mm2) and the pin number are actually impressive. The die size makes it hard to produce with 0 defect, greatly diminishing the yield, and the number of IOs makes the PCB way more complex, with many layers to be able to route the 512b memory bus, so a more expensive PCB as well.

Regular fp16 performance has the same rates in both the GeForce and in their professional lineup, that's the point I wanted to correct in your original statement.

That's also what I pointed out in his original statement. I don't see what's so different between a Titan and a xx90 card.

 

[igormp]

FP64 (or FP16) ratios were dropped not to cannibalize the expensive Quadro/Tesla cards which makes perfect sense from Nvidia's point of view.

FP64 has been dropped in all products par the x100 chips after kepler. The latest consumer-facing product to have proper FP64 hardware was the Titan V, which used a V100 chip.
All other chips, be it geforce, quadro or tesla, lack proper FP64. You'll only be seeing those in the x100-based chips.
Regular FP16 has not been capped between products. Tensor FP16 with FP32 acc has, which is clearly a market segmentation tactic, as I had said before.
I do agree with your points.

That's also what I pointed out in his original statement. I don't see what's so different between a Titan and a xx90 card.

I guess there's a point in that the titan used to have no artificial limitations whatsoever, while the 3090/4090 did, even if those differences were mostly irrelevant, but were still differences nonetheless.
Still, complaining about this is kinda moot IMO, the GPU delivers the performance of the highest-end available product, has a price tag to match, and fits perfectly for the use of prosumers, like a halo product should be.

 

[TokyoQuaSaR]

I guess there's a point in that the titan used to have no artificial limitations whatsoever, while the 3090/4090 did, even if those differences were mostly irrelevant, but were still differences nonetheless.
Still, complaining about this is kinda moot IMO, the GPU delivers the performance of the highest-end available product, has a price tag to match, and fits perfectly for the use of prosumers, like a halo product should be.

Yes, that's exactly my thoughts. The only time a Titan was noticeably different is as you pointed out the Titan V but because of the architecture it was based on.