NVidia Ada Speculation, Rumours and Discussion

[DegustatoR]

It will be changed at least 5 times till launch

These are apparently TBPs as well so rated TDPs should be ~30W lower? Which would mean that 4080 would be below 3080 in power consumption...

There's also this though:

https://twitter.com/x/status/1556979094666874882

Which makes me wonder if this one above is the actual 4080 while that 320W GA103 card will end up being 4070.

Anyway, all these reports on Nvidia tests of various configurations are fairly useless.

 

[TopSpoiler]

Looking at these power values, it looks like Nvidia is starting to consider using AD102 for 4080 and AD103 for 4070.

 

[Deleted member 2197]

https://twitter.com/x/status/1561942397680254976

NVIDIA GeForce RTX 4080 rumored to feature 23 Gbps 16GB GDDR6X memory and 340W total board power - VideoCardz.com

NVIDIA RTX 4080 rumored specs updated with faster memory Another week, another update to RTX 40 specs. NVIDIA is working around the clock to prepare their next-gen GPU series for launch later this year. This also means that the company is cycling through multiple cards specs while they evaluate...

videocardz.com

 

[techuse]

https://twitter.com/x/status/1561942397680254976

Not that much of an increase in cuda cores over a 3080. What are these expected to clock at and is the shader architecture expected to see big changes other than the rumored cache improvements?

 

[Jawed]

Not that much of an increase in cuda cores over a 3080.

Looks extremely suspicious, 9728 ALU lanes versus 3080's 8704.

 

[Seanspeed]

For a while now, AD103 has only ever been rumored to be 10752 FP32 as a full chip, same as GA102(even from that big Nvidia 'hack' leak). And Kopite himself has been saying 9728 for the 4080 for a while now, so that's not exactly new. It'd certainly leave the door open for a 4080Ti.

Clocks would certainly be much higher than a 3080. Wouldn't be surprised if it was around a 2.7-2.8Ghz actual boost clock.

I'd guess something roughly like 25-30% better than a 3090, in turn being like 40% better than a 3080.

 

[PSman1700]

Nice with large performance improvements in pure raster still. Going to be intresting to see RT/ML performance and finally pricings for future gpus.

 

[Deleted member 2197]

Leaked OEM documents claim GeForce RTX 4090 with 24GB memory is already in production - VideoCardz.com

NVIDIA GeForce RTX 4090 with 24GB GPU and AD102 GPU is already in production There might be no pictures of the next-gen NVIDIA GPUs yet, but there are documents claiming it is already manufactured. Over at Baidu (as spotted by @harukaze5719) there is a set of images posted by users suggesting...

videocardz.com

 

[trinibwoy]

Nice with large performance improvements in pure raster still. Going to be intresting to see RT/ML performance and finally pricings for future gpus.

Those RT MIMD cores will appreciate the clock boost. If the caches can keep up it could be 40% faster than the 3080.

 

[DegustatoR]

Not that much of an increase in cuda cores over a 3080. What are these expected to clock at and is the shader architecture expected to see big changes other than the rumored cache improvements?

2.8 official boost with actual being closer to 3Ghz is the latest rumoured figure.

 

[nnunn]

"Going to be interesting to see RT/ML performance... "

Indeed! Also wondering if the new cache subsystems will add Hopper-like async transfers between (caches, shared.mem) and gpu main mem. If not, bandwidth-limited compute kernels might do better with that full-fat 912.4 GB/s bandwidth of el cheapo 3080/12. Running 4 x 3080/12 at (cpu base clock, +400 on mem)... easy way to get 4 TB/s total bandwidth on a single 1600W PSU.

 

[vola]

apparently the reveal will be at GTC next month on September 20

In Gaming, our partners and ecosystem are responding to a sudden slowdown in consumer demand and correcting channel inventory. Still, the fundamentals of gaming are strong. We’ll get through this over the next few months and go into next year with our new architecture. I look forward to telling you more about it at GTC next month.

I look forward to next month’s GTC conference, where we will share new advances of RTX reinventing 3D graphics and gaming.

NVIDIA CEO - Jensen Huang (Q2 2023 Earnings Call)

 

[pjbliverpool]

This wording worries me greatly from a pricing perspective:

And it (Ampere) remains the best GPUs in the world, and it will be very successful for some time. However, we do have exciting new next-generation coming and it’s going to be layered on top of that.

That sounds to me like Ampere will be sticking around and Ada will be priced above it. I hope I'm being overly pessimistic there.

 

[DegustatoR]

That sounds to me like Ampere will be sticking around and Ada will be priced above it. I hope I'm being overly pessimistic there.

It's going to be a year with both co-existing on the market. Nothing new for a generation switch.

 

[Malo]

Looks like we'll probably see 4090 limited release Q4 then start releasing more SKUs Q1 2023.

 

[Remij]

This wording worries me greatly from a pricing perspective:



That sounds to me like Ampere will be sticking around and Ada will be priced above it. I hope I'm being overly pessimistic there.

Yep.. I said as much back in early August.

We'll see. I'm skeptical.

I think they will bump it up and they'll say "You want cheaper GPUs? We've got some for you.. the 30 series.

Anyway I hope you're right.

 

[troyan]

Samsung's 8nm should be up to 4x cheaper than TSMC's 4nm process. So Ampere will be used for low-end to (lower) mid-range for a while.

 

[Kaotik]

Samsung's 8nm should be up to 4x cheaper than TSMC's 4nm process. So Ampere will be used for low-end to (lower) mid-range for a while.

People really should stop talking about nm when talking about processes, they have nothing to do with anything anymore (and while manufacturers still use them, the actual process names don't)
For example, TSMC doesn't have 4 nm process, they have N5, N5P, N4, 4N (nvidia wants own name for their specific config for whatever reason) etc which belong to their "5nm class processes" but not one of them has anything 5nm "sized"

 

[troyan]

Hm, TSMC claims they have a "4nm" process:

TSMC’s 5nm (N5) Fin Field-Effect Transistor (FinFET) technology successfully entered volume production in the second quarter of 2020 and experienced a strong ramp in the second half of 2020.
[...]
In addition, TSMC plans to launch 4nm (N4) technology, an enhanced version of N5 technology. N4 provides further enhancement in performance, power and density for the next wave of N5 products. The development of N4 technology is on schedule with good progress, and volume production is expected to start in 2022.[/Quot€]

https://www.tsmc.com/english/dedicatedFoundry/technology/logic/l_5nm

 

[Kaotik]

Hm, TSMC claims they have a "4nm" process:

Like I said, they still talk about it, but even process names don't refer to nanometers anymore (it's N5 instead of 5nm etc, Intel 7 instead of 10nm and so on)
The little relation "nanometers" still had with reality was lost for good when we moved to finfets. The nm is supposed to refer to smallest feature size, gate length in transistor. There isn't too detailed information on all processes, but 14nm has enough details for 3 manufacturers. Depending on manufacturer "14nm gate length" is actually 18 to 30 nm