![]() ![]() So while the RTX 3060 is mightier by CUDA count, it actually has two fewer SMs than the RTX 2060. Since gaming workloads can vary wildly the flexibility must be there to offer both data paths in good amount, but there's often a leaning towards greater FP32 operations and therefore a boon to performance by offering more capability to crunch those numbers. You can see it the Ampere SM diagram, a fixed FP32 unit besides a FP32/INT32 path within each of the four SM processing blocks. It's the FP32 units that are effectively being doubled here, what Nvidia classes as CUDA Cores, while the remaining silicon remains at a close count to its Turing counterpart.Įven the doubling of FP32 units isn't all that simple-the use of a mixed function path that can jump between either integer or floating point tasks is the key to the new core counts all-round. That's significantly higher than the 1,920 CUDA Cores of the RTX 2060 (opens in new tab), which is our main point of comparison for this review, although it must be said that the way the Ampere generation works doesn't make this a straightforward doubling of performance. Making for an overall CUDA Core count of 3,584. Thanks to the advancements in the Ampere architecture, those SMs each carry 128 CUDA Cores, or more precisely 128 potential FP32 units, a piece. There's a little redundancy in the actual RTX 3060 chip, though, which comes with two fewer for a total of 28 SMs. It touts three GPCs, each one with the potential for up to 10 SMs in total. It's a brand new chip, silicon not yet seen in our gaming PCs, and that means that the RTX 3060 is in some ways unlike any other Ampere GPU to date. The GA106 GPU is at the heart of the RTX 3060. And there's plenty to like about the card when you get down to the metal, too. Yet 12GB of memory is welcome at a price of $329, especially considering the RTX 2060 launched for $349 and with just 6GB of GDDR6. That results in slimmer memory bandwidth compared to the other Ampere generation graphics cards at 360GB/s. Let's stick to what we know, and that's 12GB of GDDR6 rated to 15Gbps over a 192-bit bus. A question I need to follow up with an engineer, no doubt. I'm not too certain on the specifics of mixing 8Gb and 16Gb modules on the same package either, for an 8GB GA106 card, but that sounds pretty dicey for more than a handful of reasons. Nvidia has done something similar in the past, with the RTX 2060 actually incorporating some pared back TU104 GPUs, and perhaps could be a way to use up unsuitable silicon, but we're far from more than tacit mention of any such configuration today (opens in new tab). That very GPU could be pared back to replace GA106, in theory. Those can be occupied in either 6GB or 12GB configurations with 8Gb or 16Gb GDDR6 modules from the likes of Micron, Samsung, and SK Hynix.Īn 8GB configuration would only be possible with a larger bus, such as the one on the GA104 GPU. That's still limited to a 192-bit bus, however, and that means six 32-bit memory controllers on the die. ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |