No full-die Ada to be released?

[magellan]

Is Nvidia already focused on their next gen products or will there be a full-die, consumer Ada GPU released? Or would the perf. difference between a 4090 and a 4090ti or Titan Ada be so slim as to be unnoticeable?

If Nvidia already has full-die Ada HPC/AI parts it shouldn't be too difficult to adapt them to the consumer market should it?

 

[EarthDog]

I thought they canceled it to work on more AI or something...

Depending on the clocks they could get and w/e power envelope....... I'd imagine it would be notably faster on rasterization. The 4090 is 16384/512/1276 (Shaders/TMUs/ROPs) while a full Ada should be around 18432/576/192. Stack that with the same clocks (ignore the dimming lights), and yeah, I bet it screams. But, core clocks, I'd guess, would be slower to save some power.

 

[Woomack]

I guess it would cost a lot, and it's not a product that gives a lot of money (high margin but very low sales amount). It would also cause a lot of problems with cooling. There were photos of expected coolers for such a high TDP, and all were quite ridiculous. Since there is no pressure from AMD, they focus on AI and the next-gen GPUs.

 

[magellan]

@Woomack, if Nvidia doesn't want to design a cooler for a consumer, halo, full-die, Ada part why not release it to the AIBs? I'm sure at least a few of the AIBs would want to try and manufacture/release a 4090Ti or 4090 Super product for more cash. It would still be cash in Nvidia's pocket and I'm sure they wouldn't be losing any profit on such a venture.

 

[Woomack]

Nvidia still has to deliver a reference design and still gives a warranty on GPUs. If AIBs fail something, then there is a high chance they will blame Nvidia. Looking only at coolers, they will be heavy and, if not designed well, may damage the PCB/GPU (directly or indirectly). Nvidia also has to invest in production, and opening it for a chip that won't sell well is not the best investment. Another thing is that RTX4000 is already old, considering how everything evolves in IT. I'm sure they calculated all the pros and cons, and that's why we haven't seen it yet (or we never will). In less than a year, we may see RTX5000, so let's hope that Nvidia will release all the high chips and won't delay "full" dies too much. We may see something from the top AMD series before the RTX5000 if I'm right. I could be wrong as all those news are mainly leaks, and even if confirmed, it doesn't mean there won't be any delays.
RTX4000 Super will have its premiere in a month, so maybe we will see something unexpected not much later. I mean, besides already announced GPUs.

 

[magellan]

It's too bad the full-die(?) Ada parts are all HPC/AI products. It would've been interesting to see what a full-die Ada part could've done under water, or even high end air.

 

[mackerel]

From the numbers EarthDog posted above, +12.5% more execution potential if you up the power and cooling to match, but it may be more held back more by VRAM speed unless they can fit faster on there.

I think the biggest reason we're not seeing it is that the 4090 already exists as the stupid high end model. If AMD were at all competitive for the top spot maybe they'd consider a higher model but that doesn't seem likely this cycle.

 

[magellan]

Not releasing a full-die part is unprecedented in Nvidia's history -- as far back as Kepler Nvidia always released a full-die part for the consumer/gamer market. Volta would be the one generation that didn't have a full-die consumer/gamer part, but only because no consumer/gamer parts were released at all for anyone.

From what I can tell from the TPU GPU DB, it looks like no full-die Ada was even released in the professional AI/HPC market either. Requiring 2 16-pin power connectors
would make powering the beast interesting, I guess you would have to buy a new PSU because utilizing two octopus 16-pin 12VHPWR cable adapters would require 8 8-pin PCIe cables.

 

[mackerel]

From what I can tell from the TPU GPU DB, it looks like no full-die Ada was even released in the professional AI/HPC market either.

I was going to say RTX 6000 Ada but that's "only" a 98.6% configuration, presumably for yield. Close enough? Does it really matter if it is full or not? You're paying for the perf you get.

 

[magellan]

So Ada is having yield issues and that's why no full-die Ada part? Was that also an issue w/Volta?

 

[EarthDog]

So Ada is having yield issues and that's why no full-die Ada part? Was that also an issue w/Volta?

No/not necessarily. Yields on high-end SKUs are typically less than low end skus. A lot more has to work right on the monster gpus.

 

[mackerel]

Making good silicon is not a perfect process. There will be defects along the way. I still have my numbers from when I tried estimating die costs in the past. There are a lot of assumptions since neither nvidia nor TSMC are going to give me the real numbers. This will be wrong, but may serve as an indicator. I should hopefully be similarly wrong on all the numbers I'm about to give so they're still relative.

AD102 I estimate at 56% defect free dies. That's not product yield since it doesn't consider binning, nor use of cut down offerings. With rounding, let's say RTX 6000 Ada is a 99% die, and a 4090 uses a 89% die. Use of a 99% die allows for a minor imperfection. Use of a 89% die would have much more flexibility in mapping out those defects, so the effective yield is much higher than the 56% defect free dies.

NAVI31 appears much better at 74% estimated defect free dies. Why so much better? It's about half the size of AD102. AMD instead shifts some risk into packaging as they have to pair that up with a bunch of MCDs. Nearest sized Nvidia offering is AD104 which is fractionally smaller, and I get 75% defect free dies.