NV Maxwell Architecture

According to NVIDIA's roadmap, the Kepler architecture adopted by GTX 600 and GTX 700 series is already quite old. as early as the first quarter of next year, the successor's Maxwell architecture may officially meet us. At present, foreign media has released the maxcompute Architecture Specification, but it is still unable to determine whether it is true.

According to wccftech, the Maxwell architecture still adopts the 28nm manufacturing process from TSMC, while NVIDIA plans to launch four Maxwell chips: gm104, gm106, gm108, and gm100, compared with the current Kepler, the most obvious feature of Maxwell is that the Cuda core will be explosively improved, while the deeper change is to enhance the design of smx units and optimize dual-precision logical units.

Specifically, each smx unit of Maxwell will have 256 Cuda cores, and this number is currently 192 in the Kepler architecture. Of course, because the number of DP and texture units remains unchanged, therefore, the ratio of DP Double Precision Floating Point units will be reduced from of Kepler. Of course, Maxwell will continue the feature that the DP Double Precision Floating Point Unit in Kepler can be disabled, thus reducing power consumption.

Apart from the changes in the number, the number of registers in the Maxwell meizu smx unit will increase by two, and the L1 cache will increase from 64kb to 128kb, in addition, 32/96, 64/64, and 96/32 can be allocated as cache and shared memory. In addition, each GPC Unit also has a kb L2 cache as the instruction set cache.

GPU architecture Roadmap

Let's take a look at the specific core:

Gm108:

Gm108 has 576 Cuda cores, 48 texture units, 8 grating units, and 64bit memory width. Its core area is similar to that of gf117 in the current year. It is obviously a low-end entry-level product.

It is said that there is still a gm107 before gm108, but the specific specification cannot be determined yet.

Gm106:

Gm106 is naturally to replace the current gk106, positioning the middle end, the specific replacement of the product is GTX Ti and gtx 660 family.

The number of Cuda cores in gm106 has soared from 960 in gk106 to 2304. It has 144 texture units and 24 grating units. Its memory width is bits and its capacity is 3 GB, the equivalent memory frequency can reach about 7 GHz.

Gm104:

From the perspective of naming, it is obvious that he is trying to replace the current gk104 core, that is, positioning in the middle and high-end.

Cm104 has five GPC units (four gk104 groups), with a total of 3840 Cuda cores, 240 texture units, and 40 grating units. The bit width is 320bit and the capacity is 3 GB, or 2.5 GB. The core frequency may be around 1 GHz, and the equivalent memory frequency can reach about 7 GHz.

Gm100:

Gm100 is a complete Maxwell architecture with a total of 8 GPC units. Each GPC unit has three smx units, and then 6144 Cuda cores, with 384 texture units and 64 grating units, the memory width is bits, and the capacity has reached an exaggerated 8 GB.

In terms of frequency, the gm100 core frequency of the geforce series is about 2.61 MHz, the acceleration frequency is about 1 GHz, while the gm100 core frequency of Tesla is about MHz, and the double precision floating point performance is tflops, this is twice the current GTX Titan.

If the Maxwell architecture is so powerful as described in the article, the performance of the Next Generation NVIDIA graphics card is indeed worth our expectation. What will amd do to resist?