Quote:Nvidia's Ampere architecture has just fairly recently arrived to consumers (at least for those who've managed to find it in stock), but that hasn't stopped leakers from figuring out what's next on Nvidia's agenda, at least for its future gaming architectures. 3dCenter.org Tweeted on new rumored specifications for Nvidia's future "Lovelace" AD102 chip, which might include 144 SMs and 18432 CUDA cores with around 66 TFlops of computing performance.

Beware, this is entirely based on rumors, and with how little we know about Nvidia's next GPU architectures, there's a huge chance most of this info is wrong. But, at least this post can give us some clues as to where Nvidia is headed.

One part of this rumor that does make sense is the codename "Lovelace"; Jensen a few years back at CES of 2018 was wearing a T-shirt with several names of popular mathematicians all through history. And with Nvidia's love for codenaming architectures with names of mathematicians, many believe Jensen was leaking the names of future architectures on his T-Shirt. One of those names was Lovelace.
...
If you were wondering about Nvidia's other rumored architecture, Hopper, it seems right now that both Hopper and Lovelace are in the works. What makes Hopper interesting is its multi-die design, which would be a first for Nvidia. What we could see is Lovelace being Nvidia's gaming architecture, as one die decreases latency, then we could see Hopper as being a data center exclusive.

Quote:The rumor mill has begun grinding with details about NVIDIA's next-gen graphics processors based on the "Lovelace" architecture, with Kopite7kimi (a reliable source with NVIDIA leaks) predicting a 71% increase in shader units for the "AD102" GPU that succeeds the "GA102," with 12 GPCs holding 6 TPCs (12 SMs), each. 3DCenter.org extrapolates on this to predict a CUDA core count of 18.432 spread across 144 streaming multiprocessors, which at a theoretical 1.80 GHz core clock could put out an FP32 compute throughput of around 66 TFLOP/s.

The timing of this leak is interesting, as it's only 3 months into the market cycle of "Ampere." NVIDIA appears unsettled with AMD RDNA2 being competitive with "Ampere" at the enthusiast segment, and is probably bringing in its successor, "Lovelace" (after Ada Lovelace), out sooner than expected. Its previous generation "Turing" architecture saw market presence for close to two years. "Lovelace" could leverage the 5 nm silicon fabrication process and its significantly higher transistor density, to step up performance.

Quote:The interesting thing about that prediction is that it doesn’t square with what’s been conventionally predicted for 5nm GPUs. According to TSMC, 5nm is only expected to introduce modest performance and power consumption improvements of ~15 percent and ~20 percent, respectively. The big winner on 5nm is supposed to be density, with up to a 45 percent gain over 7nm, though these improvements tend to depend on exactly what kind of chip you are trying to build in the first place. Larger, more power-hungry structures intended for high-speed operation tend to draw more power than a more modest implementation.

Nvidia’s huge core count expansion would make sense given predicted density improvements, but power consumption is a major unknown. The RTX 3090 significantly outperforms Turing, but Nvidia had to expand the GPU’s power consumption to do it, up to 350W from 280W. It’s not clear how much additional headroom exists to keep pushing GPU power consumption. I won’t claim to know exactly where the cutoff would be, but it’s difficult to imagine Nvidia shipping 450W-500W cards for consumer systems. At some point, Nvidia is going to have to limit its own growth. Intel and AMD will allow their respective CPUs to draw over 200W of power in short boosts, but they don’t sit at those TDPs long-term by default.
...
Lovelace is currently expected in 2022. It’s not known if Nvidia will launch a true Ampere refresh cycle in 2021, or if the company will instead opt to launch high-VRAM variants of cards. There are rumors of an RTX 3080 Ti (20GB) and an RTX 3060 Ti with 12GB of RAM — an RTX 3070 Ti with 16GB of RAM would fit neatly in the stack. Nvidia could potentially pair these VRAM jumps with higher clocks or slightly more GPU cores across the new hardware for any 2021 refresh cycle.