Nvidia launched its latest line of Tesla GPU compute engines at thecompany s Graphics Technology Conference in San Jose today. Onemodel shipping immediately is based on the existing GK104 chip usedin the recently released GTX 680 . Dubbed the Tesla K10, the board delivers as much as 4.6 teraflopsof single precision floating point performance, roughly three timesthe single precision FP of the older, Fermi-based Tesla. The cardcan also handle an aggregate memory bandwidth of 320GB per second.This board is targeted towards oil exploration, signal processingand seismic processing applications. The more intriguing announcement is the Tesla K20. Built on amonster chip with 7.1 billion transistors, the K20 isn t slatedfor release until Q4. Nvidia s CEO, Jen-Hsun Huang noted that theK20 was the largest, most complex semiconductor chip ever built. Itwill likely use the same 28nm manufacturing process used for theGTX 680. The K20 is designed for computationally intensive HPCenvironments, particularly Finite Element Analysis (FEA), financeand physics applications. It offers three times thedouble-precision floating point performance of previous generationTesla products. In addition to the huge transistor count, the K20will sport a 384-bit memory interface. New Features In addition to improved compute performance, the K20 will supportseveral key features to keep the chip busy when being fed computechores. Hyper-Q increases the number of work queues from a singlequeue in the previous generation Fermi chip to 32 work queues. Thisimproves GPU utilization, keeping more of the compute cores hummingwhen running parallel compute applications. Dynamic Parallelism behaves like a kind of parallel branchpredictor. When fed tasks, the K20 can keep track of dependenttasks and spawn new compute kernels to complete those tasks, ratherthan having to request more work from the CPU. Huang demonstrated a simulation of particles colliding, firststarting with the last generation Fermi chip. That GPU could handle20,000 bodies colliding in real time at high frame rates. Then hewent on to demonstrate real-time modeling of the Andromeda andMilky Way galaxies colliding not something we need to worryabout for the time being, since it won t happen for 3.8 billionyears. That simulation ran on a Kepler-based Tesla, showing over208,000 bodies colliding. The GPU in the K20, code-named GK110, is expected to be used in thenet Titan supercomputer being built at the Oak Ridge NationalLaboratory and the Blue Waters system at the National Center forSupercomputing Applications at the University of Illinois atUrbana-Champaign. We are high quality suppliers, our products such as China Kraft Clasp Envelopes , China Poly Bubble Envelope for oversee buyer. To know more, please visits Kraft Bubble Mailer.
Related Articles -
China Kraft Clasp Envelopes, China Poly Bubble Envelope,
|