The 1080 is currently the best GPU on the market, but what part of this GPU is different or more impressive than the previous ones (such as the 780).
bjay
@Drizzle this probably contributes, but is not the whole answer: A 256kb register file is huge. Since GPUs are all about throughput, the ability to hold massive amount of data on-chip pays huge dividends.
Firephinx
@Drizzle The 1080 has about twice the performance as the 780. This is due to the following:
The 780 has only 12 SM cores vs the 1080's 20. However, each SM core in the 780 has 192 CUDA cores vs. 128 CUDA cores in the 1080. This results in the 780 having 12 x 192 = 2304 CUDA cores vs 128 x 20 = 2560 CUDA cores in the 1080.
The 1080 has a higher memory clock speed of 10 Gbps vs 6Gbps.
The 1080 also has a higher memory bandwidth (320 GB/s) vs. (288.4 GB/s) in the 780.
The 1080 has a higher base clock speed (1607 MHz vs 863 MHz) as well as a higher Boost clock speed (1733 MHz vs 900 MHz).
Finally the 1080 has 8GB of GDDR5 memory vs 3GB in the 780.
All the while, the 1080 only uses 180W vs 250W for the 780. The Nvidia GTX 1080 with Pascal Architecture is quite an engineering feat and has made dramatic improvements in just a few years. I think the main reason for the performance increase is due to the near doubling of the 780's clock speed.
I got my information from Nvidia's official specification pages for the 1080 and the 780.
P.S. Also the GTX 1080 isn't the best chip on the market right now. Nvidia's Titan X beats it by about 20% (but it is also double the cost of the 1080).
metainf
What is the difference between the shared memory and the L1 cache? Is the shared memory just a chunk of memory that is even faster than the L1 cache?
lya
@metainf Shared memory is faster than L1 cache. Also, shared memory is the working memory of each core according to Slide 65.
The 1080 is currently the best GPU on the market, but what part of this GPU is different or more impressive than the previous ones (such as the 780).
@Drizzle this probably contributes, but is not the whole answer: A 256kb register file is huge. Since GPUs are all about throughput, the ability to hold massive amount of data on-chip pays huge dividends.
@Drizzle The 1080 has about twice the performance as the 780. This is due to the following:
The 780 has only 12 SM cores vs the 1080's 20. However, each SM core in the 780 has 192 CUDA cores vs. 128 CUDA cores in the 1080. This results in the 780 having 12 x 192 = 2304 CUDA cores vs 128 x 20 = 2560 CUDA cores in the 1080.
The 1080 has a higher memory clock speed of 10 Gbps vs 6Gbps.
The 1080 also has a higher memory bandwidth (320 GB/s) vs. (288.4 GB/s) in the 780.
The 1080 has a higher base clock speed (1607 MHz vs 863 MHz) as well as a higher Boost clock speed (1733 MHz vs 900 MHz).
Finally the 1080 has 8GB of GDDR5 memory vs 3GB in the 780.
All the while, the 1080 only uses 180W vs 250W for the 780. The Nvidia GTX 1080 with Pascal Architecture is quite an engineering feat and has made dramatic improvements in just a few years. I think the main reason for the performance increase is due to the near doubling of the 780's clock speed.
I got my information from Nvidia's official specification pages for the 1080 and the 780.
P.S. Also the GTX 1080 isn't the best chip on the market right now. Nvidia's Titan X beats it by about 20% (but it is also double the cost of the 1080).
What is the difference between the shared memory and the L1 cache? Is the shared memory just a chunk of memory that is even faster than the L1 cache?
@metainf Shared memory is faster than L1 cache. Also, shared memory is the working memory of each core according to Slide 65.