NVIDIA-Partner beginnen mit der Ausgliederung von RTX 3070 Ti von größeren GA102-Dies



NVIDIA manufactured a heap of large “GA102” Ampere silicon to cater to demand from the crypto-mining boom; only to see that demand vanish. With next-gen RTX 40-series awaiting ramp; the company has to digest these GA102 chips somehow, and is apparently letting its partners use them on performance-segment SKUs such as the GeForce RTX 3070 Du. Der RTX 3070 Ti is normally based on the GA104 silicon, which it maxes out, alles aktivieren 6,144 CUDA-Farben, 48 RT-Kerne, 192 Farbtensor, 192 TMUs, und 96 ROPs, besides the chip’s full 256-bit wide GDDR6X memory interface. This SKU is now being carved out on the larger GA102, by enabling 48 von 84 Streaming-Multiprozessoren (gerade 57% of the CUDA cores enabled); and narrowing the memory bus from its normal 384-bit, down to 256-bit.

The memory size remains at 8 GB, memory type at GDDR6X, and memory speed at 19 Gbit/s, trainieren zu 608 GB/s Bandbreite. The most interesting aspect of carving the RTX 3070 Ti out of the GA102 has to be board power; with a ZOTAC-branded card listing it at 320 W, higher than the 290 W of GA104-based cards from the company. Leider, this is a China-only SKU. Every custom-design graphics card, especially from a reputed AIC such as ZOTAC, has to go through qualification with NVIDIA; which means NVIDIA is not only aware of GA102-based RTX 3070 Ti-Karten, but is behind fusing the SMs to carve out the SKU, and developing the video BIOS and driver support. ZOTAC is kind enough to list the ASIC code on its website, and for this SKU it isGA102-150-xx.