NVIDIA NVIDIA Tesla K20 Tesla C2075 and M2090 are operational card stock high resistance ..

GPU number and type: 1 Kepler GK110 CUDA cores: 2496 pairs of precision floating point performance: 1.17 Tflops single-precision floating point performance: 3.52 Tflops dedicated memory total capacity: 5GB power: 225W thermal design power Tesla K20 and NVIDIA GK110 architecture for K20 described as "double precision floating point performance three times" and have Hyper-Q, Dynamic Parallelism and other parallel computing blessing, these are existing GK104 architecture does not have. NVIDIA's GK110 PDF document introduces the SMX architecture is 192 CUDA cores must admit, before news leaked about the GK110 architecture is wrong, GK110 GK104's SMX architecture in fact, with still the same, all 192 CUDA cores 32 group SFU units and 32 LD / ST unit. Remove among other functional units, GK110 core group of a total of 15 SMX units, 2880 CUDA cores, but Heise claimed that not all units are enabled, may actually be only 13-14 group SMX unit, the actual CUDA core is 2496 or 2688. Memory interface is 384bit, has been identified as Huang and NVIDIA CTO. Since CUDA cores have been lower than previously reported, down 384bit memory interface is very natural thing, if you keep the memory speed 6Gbps GK104 and GK110 bandwidth that will reach 288GB / s, and finally over AMD GCN architecture of 260GB / s a. NVIDIA is given three times a double-precision floating point performance is relatively unknown with the card or with the GF110 GF110 core of Tesla accelerators do, GF110 single-precision floating-point capability 1.58TFLOPS, graphics double-precision to single precision of 1 / 4, which is 0.4TFLOPS, but the GF110 core of Tesla card double capacity of up to 1/2 of single-precision, about 0.8TFLOPS. Thus, if the card is based on double-precision floating point performance GK110 about 1.2TFLOPS above, if it is three times the Tesla card that 2.4TFLOPS above, in view of the latter is beyond the capacity of previous rumors 2TFLOPS of, GK110 The double-precision floating-point capability should be 1.2TFLOPS or higher. Tesla K20 configured 6pin + 8pin power supply interface between the core area and the TDP is unknown, but the K20 is equipped with 6pin and 8pin power supply interface, the maximum TDP will not exceed 300W. The number of transistors is a 7 billion, to be exact is 7.1 billion. The basic specifications of the graphics card is so much information, and then look at the NVIDIA GK110 increase as new technologies it. Dynamic Parallelism (dynamic parallel) GK110 architecture One of the primary goals is to make it easier for programmers to call the powerful GPU parallel computing capability. Under the traditional model, GPU CPU every operation need to be involved, and Dynamic Paralleliom presence makes GPU receives data dynamically refreshed thread without CPU involvement. Because the kernel with an independent load capacity workloads, dynamic parallel technology allows programs to run directly on the GPU. The benefit of this technology is that you can reduce programming complexity, 200-300 lines of code originally required to complete the work on the GK110 graphics needs only 30 lines on it. Hyper-Q on a technical emphasized that simplify operations, is to offload CPU, and Hyper-Q is to increase the number of CPU cores while loading work is in ascension = high CPU utilization, and avoid excessive CPU idle. Fermi architecture CPU can only run one MPI (Message Passing Interface Message Passing Interface) task, but the number of tasks in the GK110 architecture MPI CPU running up to 32. MPI task is mainly based on traditional multi-core CPU applications, compared to the powerful GPU parallel computing power, MPI amount of CPU processing task is too small, often bring false GPU-dependent, resulting in GPU performance can not be effectively utilized, Hyper-Q a substantial increase in the amount of CPU MPI tasks can be assigned to the GPU, if passing the 32 tasks to GPU, it will be up to 32 times the theoretical performance of the Fermi architecture, practical applications, although not so exaggerated, but after Optimal GPU parallel computing power will still be improved. GPU Direct GPU Direct NVIDIA official PDF Direct Connect is not mentioned, but still worth explaining. NVIDIA has launched the Kepler architecture-based GeForce GRID cloud gaming technology, use Kepler graphics server will inevitably have to exchange data with each other. GPU Direct technology allows servers in different memory card directly read data, graphics and even between different servers can also read another piece of data in the memory card through the network card, is simply to improve the graphics data exchange capabilities, fewer steps needed, lower latency. CUDA 5 To use the technology described above it is necessary to use the new CUDA 5, GTC conference NVIDIA has released a preview version of the CUDA 5 SDK, the official version will be released in the third quarter of this year. After Kepler graphics published, Tesla family has finally ushered in the schema update, and it will soon have to update the schema of Tesla accelerator, thanks GK104 good performance ratio, NVIDIA's Tesla accelerator also have this kind of capability, performance and more strong, while consuming less power. GK110 architecture focusing on a new generation GPU computing performance did strengthen, double precision floating point capabilities up to three times before the architecture and dynamic parallelism, Hyper-Q, GPU Direct and other technical assistance, both ease of use and performance have improved significantly, to assume the glorious mission of GPU computing.