During this time, the NVIDIA, AMD each of the next generation of flagship GPU core exposure, but are a variety of rumors, unpredictable. Now, we've dug up two new cores of NVIDIA from the GeekBench5 test database, both fairly large. One display has 118 computing units (NVIDIA is called SM streaming multiprocessor). If each group continues the configuration of 64 streaming processors (CUDA cores), the total number is 7552, and the core frequency is 1.11ghz, with 24GB of video memory (detection value 23.8gb).
The other is 108 computing units, corresponding to 6912 stream processors, with a core frequency of 1.01ghz and 48gb (detection value of 46.8gb) of video memory.
Obviously, they are the core of NVIDIA's next generation. Maybe they are legendary"Ampere", 7Nm technology, maybe even the next generation"Hopper"。
NVIDIA has also done dual architecture core parallel work, for example, Tesla V100 series is the Volta (volt) core, only for professional market, geforce RTX 20 and GTX 16 series are Turing (Turing) core, only for consumer market.
Due to the lack of other evidences, it is still impossible to confirm which sequence of these two cores belongs to Tesla, Quadro, Titan and geforce,According to NVIDIA's product development and release rules, Tesla is very likely, such as Tesla A100, but geforce, such as RTX 3080 Ti, is not excluded.
If RTX 3080 ti is the core of 118 units and 7552 stream processors, the scale of RTX 2080 ti is 74% larger than that of 68 units and 4352 stream processors.
And it's probably not the full core,128 units is the most reasonable strength, which is 8192 stream processorsCompared with RTX 2080 Ti, it is 88% larger, even compared with Titan RTX with 72 units, 4608 stream processors and complete tu102 core, it is 78% larger.
If the 24GB video memory is the gddr6 of the game card, it has more than doubled compared with the RTX 2080 Ti, and if it is the hbm2 of the professional card, it is only a low configuration version compared with the current Tesla V100 16 / 32GB.
The core of 108 units and 6912 stream processors is certainly not RTX 3080. This will increase the scale of RTX 2080 by 135%, which is very outrageous. Secondly, the display memory capacity is as high as 48gb, which is impossible to be used in game cards. It can only be Tesla or even Quadro professional cards.
As for the low frequency of 1.01ghz and 1.11ghz, it is easy to understand that they are all early engineering samples, which appeared at the end of October and November last year respectively.At the same time of the rapid growth of core scale, even at such low frequency, the performance is quite fierce. The OpenCL calculation run points of 118 units and 108 units are 184096 and 141654, respectively. As a reference, Tesla V100 154606, Quadro gv100 142837 and Titan RTX 132804.
The next generation of computer card, the radon insight mi100, codenamed Arcturus, is said to double the Vega core, with 128 computing units, 8196 stream processors and a power consumption of 200WWe have the strength to fight the first World War.