NVidia Ada Speculation, Rumours and Discussion

Well for one there's the fact that GH100 isn't a videochip as it lacks most of graphics h/w.

 

Whitepaper says: “Only two TPCs in both the SXM5 and PCIe H100 GPUs are graphics-capable (that is, they can run vertex, geometry, and pixel shaders”.

It doesn’t say anything about the hardware though. Hopper still has a full complement of texture units and Nvidia is still using “GPC” and “TPC” to describe the architecture. I wonder if the rasterizers, ROPs and setup hardware are actually missing.

 

H100 can still run compute workloads, which require texture filtering, so texture units are still there.
I guess rasterizer and ROPs might only exist in the graphics-capable GPC after nvidia decoupled ROPs and MCs.

 

Asking the programmers, do DPX (dynamic programming) introduced in Hopper, will be beneficial on gaming GPU Lovelace ?
https://blogs.nvidia.com/blog/2022/...s-dynamic-programming-using-dpx-instructions/
I was thinking of using DPX to avoid shader stalls, optimize warp occupancy, or get better branching performance

 

There's no public available documentations of these DPX instructions right now (at least I can't find any).
But since these are about dynamic programming I'd guess that they are probably about finding/indexing something in a vector, since a lot of dynamic programming algorithms are about finding a specific value or a minimum/maximum value in an array.
These instructions might be useful in some game logics (e.g. path finding), but of course we'll only find out after the documentations are public.

 

What technology that only Hopper had could be integrated into Ada? I think Hopper doesn't have such technologies can benefit gaming architecture. It's not even MCM.

 

We don't really know that. Nvidia hasn't said anything about the changes in Hopper which would benefit gaming so far.
But I doubt that this tweet has anything in common with reality TBH.
Unless AD102 is MCM and that technology is what was "integrated" from GH202. But I don't see how that's "Hopper's new technology" really.

 

For gaming:
New SM scheduler (TMA accelerator in Hopper terminology ie more efficient asynchronous execution)
DPX instructions
Thread block cluster
The biggest boost: 2xFP + INT pipeline (still not sure about this one)

And not in Hopper but 100% sure, 3rd gen RT cores that are much improved (compressed BVH pipeline that fits in L2 cache + new hyperfast intersection algo)

 

How this would even work though? Three 16-wide SIMDs with one 32-wide warp per clock? They'll need a separate port for INT h/w then and the SM level scheduling will become complicated.
It also seem a bit excessive from typical math mix POV - unless the mix has changed more in favor of FP32 now.

 

Have we some hints about that ?

 

It always amazes me when things are cancelled that never existed. This next generation will just be bigger caches and faster RT. Bet on it.

 

I've worked on plenty of SOCs that never existed as products but went far into development and were cancelled due to changing customer needs or demand. I don't know if it's true in this case, but this stuff happens all the time in the industry.

 

That’s fair but it’s hard to imagine Nvidia cancelling entire architectures these days given their market position and expertise in the game. Cancelling an SKU or two sure but they have a massive R&D machine that should “cancel” ideas well before they’re close to production.

 

Chips can be cancelled prior to production, happens all the time, nothing weird about it.
Entire architectures are quite unlikely, and GH202 is Hopper so it's not the case anyway.
The early appearance of Blackwell chips does look a bit suspect though so it's possible that Nv decided to bring GB chips forward instead of producing GH202.

 

Blackwell?

 

https://videocardz.com/newz/nvidia-ada-ad102-107-hopper-gh100-202-blackwell-gb100-102-gpu-confirmed

 

Yeah, next gen server after hopper (or possibly separate line to co-exist with hopper specialized for something else?)
Kopite7Kimi hinted about it last summer and it was confirmed in the NVIDIA hack leaks

 

With so many cores needing to be fed, maybe AD102 will fix memory speed and heat issues with skinny 512-bit HBM3? From NextPlatform.com:

 

Not sure about the details, but big architectural improvements for RT in Lovelace are expected from a chip design perspective. For Turing Nvidia was designing their RT in a vacuum without game developer experiences and suggestions. The same happened mostly for ampere, as you can't change much in the last 2 years in an architecture. So only intersection speed was improved and motion blur. Also the reason, why AMD could only add very slow RT late in the process of RDNA2 architecture.

Lovelace is the first architecture, which should implement architectural improvements based on the first RT games and fix some of the easier to solve RT bottlenecks, which appeared.

 

And hopefully those improvements accelerate existing RT games and don’t require devs to do things differently.