According to a tweet by @momomo_us, four Russian listings have confirmed the specifications of Nvidia's upcoming GeForce RTX 4060 Ti graphics card, including its underwhelming 8 GB memory capacity. The listing shows four video card models from Palit, detailing the memory configuration and display connection outputs.
The listed specifications align perfectly with previously reported leaks outlining the RTX 4060 Ti's full specifications. Essentially, we see a very underwhelming memory configuration from these reports, surrounding a 128-bit bus and 8 GB memory configuration.
If previous leaks hold true that this memory configuration will operate with 18 Gbps GDDR6 modules, the RTX 4060 Ti will have roughly half the memory bandwidth of the RTX 3060 Ti and even the RTX 2060 Super, with just 288 GBps. Adding the fact that the RTX 4060 Ti is also reported to have just 8 GB of memory, this GPU is starting to look very underwhelming compared to its predecessors.
pic.twitter.com/BhFvcGkAROMay 5, 2023
The good news is that the RTX 4060 Ti's Ada Lovelace GPU architecture does have substantially more L2 cache than its predecessor (32MB), which should counter many potential downsides with its memory configuration. It reportedly has 36% more TFLOP compute performance, so it could be significantly faster than its predecessor in non-gaming situations.
But going back to the gaming aspect of the card, with its slower memory bandwidth combined with just 8 GB of memory, it'll be interesting to see how much of a bottleneck 288 Gbps will be.
The problem is that modern video games stream assets and textures on the fly to the GPU memory constantly as the player roams around the game world. For example, Unreal Engine 5 titles utilize Nanite because open-game worlds have become far too big to house all the necessary textures and assets in video memory. As a result, having lower memory bandwidth could reduce frame rates when the GPU has to wait for new textures to be streamed in before rendering out a finished frame.
Current rumors indicate that the RTX 4060 Ti will be unveiled during Computex 2023 later this month for $400 to $450. So at the very least, we're getting a cheap Ada Lovelace GPU, even if it won't be that much faster than its predecessor in memory-bound scenarios.
Based on the rumored 2685 MHz boost clock on some custom models, the card could deliver up to 23.3/24 TFLOPs of compute horsepower which seems roughly 44% higher than the NVIDIA GeForce RTX 3060 Ti.
But the actual performance increase should be 30-40% faster than the RTX 3060 Ti.
How much longer until Nvidia starts producing Light Artificial Intelligence (LAI) cards? They had Light Hash Rate (LHR) cards. Bring on the LAI variants!
Yeah, I read about it on Nvidia's blog, the neural texture compression method. The method sounds kind of interesting.
Random-Access Neural Compression of Material Textures/NTC:
They also shared a lot of technical details on this texture compression method in this PDF file.
so imho it wont be 30%+ better.
not the 1st time a card would be heldback by its memory bus being shafted.
While more cache helps, it doesn't eliminate the need for fast VRAM.