The most conscientious n-card in ten years! NVIDIA RTX 3080 launch evaluation
1、 Foreword: the arrival of RTX 3080 is totally different from what was said
The first revolution in the field of 3D graphics cards was the T's T of integration in 1999
In the next 10 years, there was no breakthrough technology in the field of GPU. Dx12 was a thunderclap with small raindrops, and its performance improvement was not as good as expected, while PhysX was not as high as expected. Until August 2018, the emergence of real-time ray tracing technology pushed the 3D game picture to a new height. For the first time, players saw the light and shadow world like science fiction movie in the game.
After the release of Turing GPU, NVIDIA's market value soared and even surpassed Intel. Rich old Huang didn't let us wait too long. In only two years, NVIDIA innovated the GPU architecture again. The geforce RTX 30 series graphics card based on the new generation ampere architecture released on September 4, 2020 unifies int32 and fp32 units. The int32 unit, which could only do integer operation, can now do both integer operation and single precision floating-point operation.
This technological innovation directly doubles the number of traditional stream processors and doubles the floating-point computing capability of GPU. So we can see that the number of RTX 3080 stream processors has increased from 2954 to 8704 in RTX 2080, which suddenly seems to be back to the 8800 GTX era.
There are 7 groups of GPCS in ga102. There are 12 SM units and 16 ROP units in each GPC. There are 84 SM units and 112 ROP units in total.
According to different specifications, the total number of SM units enabled by RTX 30 series graphics cards is different. RTX 3090 is 82 sets of SM units and 116 ROPS, RTX 3080 is 68 sets of SM units and 96 ROPS, and RTX 3070 is 46 groups and 96 ROPS.
This is the SM unit architecture of ga102 core. Each SM unit has 64 int32 integer cores, 64 fp32 single precision floating-point cores, four third-generation sensor cores and one second-generation RT core.
The int32 unit here can perform single precision integer or floating-point operations according to requirements, that is to say, a sm unit can have up to 128 fp32 units.
The total cache capacity in each SM unit is 128 KB, which can be flexibly assigned to the first level cache and shared cache (Shared Memory), which can be a combination of 64 KB 32KB or 32 KB 64KB.
Nvidd released a total of three ga10x core graphics card models.
RTX 3090：It has 7 sets of GPCS, 82 sets of SM units, a total of 10496 stream processors, 112 ROPS, 328 texture units, 328 third-generation tensor cores and 82 second-generation RT cores. With 24GB gddr6x memory, 19.5GHz
The bit width is 384bit and the bandwidth is 936gb / s. Compared with Titan RTX, it is 50% faster than Titan RTX. It will be launched on September 24 with the price of 11999 yuan.
RTX 3080：It has 6 sets of GPCS, 68 sets of SM units, a total of 8704 stream processors, 96 ROPS, 272 texture units, 272 third-generation tensor cores and 68 second-generation RT cores. With 10Gb gddr6x memory, the frequency of video memory is 19ghz
The bit width is 320bit and the bandwidth is 760gb / s. Its comprehensive performance is twice that of RTX 2080. It was put on the market on September 17 and the price was 5499 yuan.
RTX 3070：Ga104 core, with 6 groups of GPCS, 46 sets of SM units, a total of 5888 stream processors, 96 ROPS, 184 texture units, 184 third-generation tensor cores, 46 second-generation RT cores. With 8GB gddr6 video memory, the video memory frequency is 14ghz, the bit width is 256bit, and the bandwidth is 448gb / s. It has comprehensive performance and RTX 2080 Ti, but its price is only half. It will be launched in October with a price of 3899 yuan.
The detailed specifications of RTX 3080 are as follows:
Related reading:RTX 3090 graphics card release: violent 5 fan 20 phase power supply