YouTube channel Moore’s Law is Dead published a rather juicy reveal of NVIDIA’s next-generation “Ampere” flagship graphics card, which will be the GeForce RTX 3080 Ti. Apparently they’ve had correspondence with sources at NVIDIA. This report goes into detail on the Founder Edition board design as well as what is on the actual silicon.
So to start things off the RTX 3080 Ti reference card (Founder Edition) will come with a triple-fan cooling solution, unlike the RTX 20-series, which only has 2 fans. The cooler is said to be quieter than the RTX 2080 Ti FE cooling solution. The card will pull power from a pair of 8-pin PCIe power connectors. For display outputs you have three DisplayPort, one HDMI, and one VirtualLink USB-C. The source also confirms that “Ampere” will implement PCI-Express gen 4.0 x16 host interface.
When it comes to the silicon NVIDIA will have three tiers of high-end GPUs, with the “GA102” leading the pack. This will of course succeed the “TU102”. Then there will be a “GA104” which will hold the upper-performance segment which succeeds the current “TU104”. There is however a new silicon between the two, “GA103” which does not have a predecessor from the current generation. “GA102” will have 5376 “Ampere” CUDA cores (up to 10% higher IPC than “Turning”). The silicon will be based on the 7nm-class fabrication node (most likely from TSMC). Clock speeds will be dialed up well above 2.2 GHz, even for the “GA102”. Smaller chips in the range will able to boost beyond 2.50 GHz. The “GA102” could end up with a FP32 compute performance in excess of 21 TFLOPs. The card will use faster 18 Gbps GDDR6 memory, giving it 863 GB/s of memory bandwidth. That is 40% higher than the RTX 2080 Ti (if the memory bus ends up at 384-bit). With this configuration the prototype card is getting 40% higher performance than t he RTX 2080 Ti at 4K UHD resolution in poorly optimized games and 50% higher performance in optimized games.
NVIDIA will be increasing the number of streaming multiprocessors, which doubles the CUDA core / RT core ratio compared to Turing. This results in more RT cores per tier so ray-tracing performance will be improved as well. Tensor core count is also going to see an increase.
NVIDIA could introduce DLSS 3.0, and updated image and performance quality enhancement. NVIDIA could also resurrect a hybrid memory technology which would be similar to AMD’s HBCC, which is called NVCache. This would spread video memory across not only the video memory, but the system memory and flash-based storage.
NVIDIA will be doing an overhaul of their software, merging the GeForce Experience with GeForce Now. There won’t be a login required anymore and RTX Voice will get an upgrade.
It appears NVIDIA’s entire line will be “RTX”, so even low-end cards will get Tensor and RT cores. HPC “Ampere” parts will be available in May, while consumer gaming cards will be released in September, likely close to the release of Cyberpunk 2077.