Il whitepaper di NVIDIA Hopper rivela le specifiche principali di Monstrous Compute Processor
The GH100 compute die is built on the 4 nm EUV process, and has a monstrous transistor-count of 80 miliardi, a nearly 50% increase over the GA100. Interestingly though, a 814 mm², the die-area of the GH100 is less than that of the GA100, con i suoi 826 mm² die built on the 7 nm DUV (TSMC N7) nodo, all thanks to the transistor-density gains of the 4 nm node over the 7 nm one.
The component hierarchy of the GH100 is similar to that of previous generation NVIDIA architectures. The main number-crunching machinery is spread across 144 multiprocessori in streaming (SM). The chip features 18,432 Core CUDA FP32, e 9,216 double-precision (FP64) Colori CUDA. Le grandi cose vengono in due 576 fourth-generation Tensor cores. One of the GPCs on the silicon has raster graphics hardware, so the silicon has certain GPU capabilities.
The GH100 features a 6144-bit HBM3 memory interface, e 80 GB is the standard memory amount for the H100 compute processor. The memory bandwidth on offer is expected to exceed 3 TB/sec, including ECC support. The host-interfaces get a major upgrade, troppo. The SXM form-factor board comes with the latest-generation NVLink interconnect that has 900 Larghezza di banda GB/s.
The AIC (add-in card) form-factor model has PCI-Express 5.0 x16 (128 GB/s). Both interfaces introduce resource-pooling features. Infine, NVIDIA is pushing up the power envelope in pursuit of performance, with the H100 having a typical power value of 700 W, rispetto a 400 W of the A100. Interestingly, the H100 does not max out the GH100 silicon, with the high-density SXM form-factor board featuring 132 su 144 SM, while the PCIe AIC board only has 114 su 144. Both models could come with clock speeds as high as 1.80 GHz.