NVIDIA is expected to launch its Ampere refresh (RTX 30 Super series) sometime in early 2022. Retaining the same microarchitecture as the existing 30 series parts, this is going to be a mid-generation upgrade of sorts. Samsung’s 8nm process node will be retained for the Super refresh with better yields. The last time, only the RTX 2060, 2070, and 2080 got Super refreshes, and the 2080 Ti and 2090 were left alone. This time we’re looking at a refresh of the might RTX 3090 as well.
|GPU||Die||CUDA Cores (FP32)||Boost Clock||Memory||Memory Bus||B/w (Gbps)||SMs: RT Cores||Tensor Cores||TDP (W)||Price ($)|
|GeForce RTX 3060||GA106||3,584||1.77GHz||12GB GDDR6||192-bit||360||28||112||170||329|
|GeForce RTX 3060 Super||GA106||3,840||~1.75GHz||12GB GDDR6||192-bit||360||30||120||200||349?|
|GeForce RTX 3070||GA104||5,888||1.72GHz||8GB GDDR6||256-bit||448||46||184||220||499|
|GeForce RTX 3070 Super||GA104||6,144||1.7GHz||16GB GDDR6X||256-bit||608||48||192||390||599?|
|GeForce RTX 3070 Ti||GA104||6,144||1.7GHz||8GB GDDR6X||256-bit||608||48||192||390||599|
|GeForce RTX 3080||GA102||8,704||1.71GHz||10GB GDDR6X||320-bit||760||68||271||320||699|
|GeForce RTX 3080 Super||GA102||8,960||~1.7GHz||12GB GDDR6X||384-bit||912||70||279||350||699?|
|GeForce RTX 3080 Ti||GA102||10,240||1.65GHz||12GB GDDR6X||384-bit||912||80||320||350||1,199|
|GeForce RTX 3090||GA102||10,496||1.7GHz||24GB GDDR6X||384-bit||936||82||328||350||1,499|
|GeForce RTX 3090 Super||GA102||10,752||~1.7GHz||24GB GDDR6X||384-bit||936||84||336||400+||~2,000?|
Starting from the bottom of the stack, we have the RTX 3060 Super. There aren’t many options here. The only one that makes sense is a full-fledged GA106 die with 3,840 FP32 cores, 30 RT, and 120 Tensors. NVIDIA could also opt to use the GA104 die, with a 256-bit bus and a higher core count, but that’s unlikely.
A step above we have the RTX 3070 Super. This is mere speculation on my part, but I believe this SKU will replace both the RTX 3070 as well as the 3070 Ti. The latter is hard to buy (harder than the rest, anyway) and will be phased out by early 2022. We’re looking at the full-fledged GA104 die paired with 16GB of GDDR6X memory (this is hard to predict) through a 256-bit bus.
The RTX 3080 Super can’t go many ways either. A GA102 die with 70 SMs enabled, and a 384-bit bus connected to 12GB of GDDR6X memory is the most likely configuration. This should offer a decent upgrade over the vanilla 3080, all the while keeping the 3080 Ti relevant (sort of).
At the top, we have the RTX 3090 Super. This will be the first SKU to feature the full-fledged GA102 die with a total of 10,752 cores and 84 SMs, resulting in a TDP of 400W+. Accordingly, the price should be on the high side as well.
Finally, by the end of 2022 or in early 2023, we should see the RTX 40 series graphics cards based on the Ada Lovelace architecture, and fabbed on TSMC’s 5nm EUV process. These parts (due to their large dies and expensive process node) should be as expensive as Turing, if not more. It would seem those core counts of up to 18,000 and a TDP of nearly 500W will come at a steep cost.