New leak pegs Nvidia Lovelace RTX 4090 at 16,384 CUDA cores and RTX 4070 at 7,168 CUDA cores, but RTX 4080 gets nerfed GDDR6 memory compared to Ampere RTX 3080
With Nvidia expected to launch the RTX 40 series Lovelace GPUs sometime in the September timeframe, more details about the cards are spilling out in the open by the day. Recently, we have seen possible power limits for Lovelace cards — both desktop and mobile. Now, we get to learn about a few additional specs including memory, CUDA core counts, and other tidbits.
As usual, this information comes via @kopite7kimi on Twitter. According to Kopite, the RTX 4090 will have 16,384 CUDA cores or 128 streaming multiprocessors (SMs), which is an increase over the 16,128 CUDA cores or 124 SMs that has been thought of so far for this card. The RTX 4090 uses a 384-bit 21 Gbps 24 GB GDDR6X VRAM and will run at a 450 W TGP.
The RTX 4080, which is likely to be pegged as the mainstream flagship gaming GPU, will feature 10,240 CUDA cores, 256-bit 18 Gbps 16 GB GDDR6 VRAM, and will likely be rated at a 420 W TGP. It is interesting that the RTX 4080's rumored memory specifications are lower than what we had seen with the RTX 3080.
Wearing the tin foil hat, this is likely because Nvidia might have had to choose between increasing VRAM capacity by using a narrower memory bus or continue increasing the VRAM comparable to that of the RTX 4090's by reusing Ampere's memory bus.
There is also likely cost dynamics at play here given that 2 GB Micron chips are still expensive and that may limit the reach of the RTX 4090.
That being said, the limitations of the slower 18 Gbps memory (RTX 3080 does 19 Gbps) are likely offset with the larger cache of the Lovelace architecture. Faster clocks may help too, which is still the missing piece of the puzzle so far.
We also get to know the RTX 4070's specs for perhaps the first time. The RTX 4070 is touted to sport 7,168 CUDA cores, 160-bit 18 Gbps 10 GB GDDR6 VRAM, and a 300 W TGP.
While these specs do seem probable, it is still advisable to keep a pinch of that proverbial salt within reach. Especially, since the current leak actually does not corroborate with an earlier such leak in many aspects.
|Spec||RTX 4090||RTX 4080||RTX 4070||RTX 4060|
|Streaming Multiprocessors||128||80||56||36 (?)|
|CUDA Cores||16,384||10,240||7,168||4,608 (?)|
|Memory Configuration||384-bit 21 Gbps 24 GB GDDR6X||256-bit 18 Gbps 16 GB GDDR6||160-bit 18 Gbps 10 GB GDDR6||192-bit 8 GB GDDR6X (?)|
|TGP||450 W||420 W||300 W||200 W (?)|
Kopite also hints that the upcoming Nvidia Lovelace GPUs are unlikely to launch at lower MSRPs than their Ampere counterparts. Nvidia is expected to launch Lovelace GPUs sometime in Q3 2022 with expected availabilities from Q4 onwards.
And TBP, 450/420?/300W.— kopite7kimi (@kopite7kimi) June 23, 2022