GPUsNews

NVIDIA Hopper GPUs (RTX 50?) May Feature Dual Chiplet Design w/ 18,432 Cores, 288 SMs

According to some new info shared by @kopite7kimi on Twitter, NVIDIA’s Hopper GPUs will feature a chiplet design with two compute dies, each featuring a core count of 9,216 packed into 144 SMs (Streaming Multiprocessors), resulting in an overall core count of 18,432 or 288 SMs. Keep in mind that this is a conservative claim as we’re expecting each SM to pack 64 cores instead of the 128 marketed with Ampere. NVIDIA actually markets the FP32 core count with Ampere with the INT32 core count being half as much. Furthermore, the GA100 accelerator SM features an equal number of FP32 and INT32 cores, bringing down the single-precision core count to 64. However, they also feature 32 FP64 cores per SM to boost double-precision performance.

Now, Hopper is expected to come after Ada Lovelace which is going to succeed Ampere sometime in late 2022. As per earlier rumors, the former will also feature up to 18,432 cores but that’s if you assume that each SM packs 128 FP32 cores and half as many INT32 cores. If we go by that estimate, then Hopper might just pack a whopping 36,864 cores which is honestly a bit hard to digest but considering that it’s still nearly four years away, it’s not that hard to believe.

Furthermore, if NVIDIA goes by its present SM design (which is unlikely given how often it’s changed), then the data center (Tensor) GPU will feature 18,432 FP32/INT32 cores and half as many FP64 cores while the top-end GeForce part will pack up to 36,864 cores. Again, this is if the SM continues to consist of 128 shaders instead of the earlier 64. Last but not the least, considering just how far Hopper is, make sure to take these figures with a grain of salt.

Areej

Computer hardware enthusiast, PC gamer, and almost an engineer. Former co-founder of Techquila (2017-2019), a fairly successful tech outlet. Been working on Hardware Times since 2019, an outlet dedicated to computer hardware and its applications.
Back to top button