Details of the new Tesla K10 and K20 continue to come out, and the K10 is proving to be a bit different than anticipated. The new Tesla K10 is using the GK104 chip, available in the GeForce cards. This chipset lacks most of the usual Tesla features, and NVidia is getting around this by marketing it specifically to a narrow slice of the market:
NVIDIA’s market strategy here is actually summed up rather well in their K10 press release: “NVIDIA Tesla K10 GPU Accelerates Search for Oil and Gas Reserves, Signal and Image Processing for Defense Industry.” GK104 lacks the ECC and compute flexibility of the Fermi Tesla cards, but what it doesn’t lack is single-precision compute performance and memory bandwidth; and with a dual-GPU card in particular it has both of those in spades. Accordingly, NVIDIA’s goal for K10 is to go after the specific market segments that don’t need ECC and don’t need flexibility, but do need all the raw compute performance they can get. This as it turns out is something gamers are already familiar with: image processing.
I’m busy watching the NVidia GTC Keynote, but just got a press release too exciting not to share. I don’t think Jen-Hsun has said it yet, but this press release indicates that he’s about to announce new Kepler powered Tesla cards, the K10 and K20.
The NVIDIA Tesla K10 GPU delivers the world’s highest throughput for signal, image and seismic processing applications. Optimized for customers in oil and gas exploration and the defense industry, a single Tesla K10 accelerator board features two GK104 Kepler GPUs that deliver an aggregate performance of 4.58 teraflops of peak single-precision floating point and 320 GB per second memory bandwidth.
A group of 100 scientists, engineers, and developers are working together for a bid at Google’s Lunar X Prize, a $30 million award to the first private funded team to land a rover on the moon. Any bid will require tons of work in computations, hardware, and physics, but the German team is beefing up their systems with the power of NVidia Tesla GPU’s.
The PTS team will benefit from the Tesla GPUs at all stages of the mission. During preparation and planning, GPUs will be used to simulate millions of different mission scenarios. This will enable the team to improve launch and landing techniques by, for example, adjusting the timing and duration of thruster burns for course corrections, while minimizing the margin of error.
Once Asimov has reached its destination, the PTS team will use the computational power of Tesla GPUs to navigate and monitor the rover’s activities and generate highly detailed lunar maps from the transmitted stereoscopic 3D images.
A few weeks ago Cray CTO Steve Scott made waves in the HPC News media with the announcement of his resignation and move to a “computing partner” that wouldn’t be named. Most people figured they meant AMD. Well, it’s been named now and it’s a shocker: He’s the new CTO of NVidia’s Tesla business unit.
“There are few people on the planet that have Steve’s deep system level understanding of high performance computing,” said Bill Dally, Nvidia’s chief scientist. ”Steve’s decision to join Nvidia is a resounding endorsement that GPU accelerated computing is the future of HPC. He will play a central role in architecting the world’s most powerful supercomputers.”
Two new offerings from NVidia today, first the new Geforce GTX560 card. In a new video on YouTube you can see it running the new Duke Nukem Forever in 1080p Stereo. Looks like a nice little bump from their existing stuff, nothing revolutionary. See the video below.
Perhaps bigger news is the new Tesla M2090 card, boasting some impressive new computing figures for GPU compute in scientific spaces.
It’s said that Tesla M2090 GPUs coupled with four CPUs delivered record performance of 69 nanoseconds of simulation per day.The fastest AMBER performance recorded on a CPU-only supercomputer is 46 ns/day.
To put it a bit simpler, a simulation of 1 microsecond of time that took 22 days previously, now takes 14, a nice saving of a week.
Over at the NVidia Forums, they’ve published some impressive benchmarking results of the new-ish Telsa C2050 Fermi-driven board running various CUDA, CULA, and HPC simulation codes. They compare the results to the previous generation Tesla cards, as well as the newest Intel CPU’s. The results are impressive. Just check out these graphs:
Be sure to check out the full PDF on their website (forum registration required, unless of course someone hosts it else where and posts the link in the comments…)
HPC people have been investigating Tesla & other CUDA technologies for several years, but mainly through test & development servers and homebrew configurations. No longer will that be required, as now IBM is jumping in the game by packing the new Tesla M2050 card into their new systems.
“NVIDIA provides an innovative solution for customers who push the envelope in high-performance computing,” said Dave Turek, vice president, Deep Computing, IBM. “GPU acceleration provides performance boosts for many applications in energy exploration, science and financial services. It is among the significant emerging supercomputer technologies to watch in the years ahead.”
The industry standard benchmarking tool LINPACK boasts an 8x improvement when run on a server with 2 M2050′s installed, so the Top500 might see some amazing upsets from smaller IBM clusters this year.
I never have seen much use for the Quadro or Tesla series of graphics processing units from Nvidia. Most of the time you can use the cheaper GeForce series to accomplish the same task at a fraction of the cost. With the new Tesla series, you get full double precision support and ECC. I can now see a reason to buy them over the GeForce series.
Here’s another behind the scenes look at NVIDIA. General manager Andy Keane explains how Tesla-based servers will help engineers and scientists solve complex problems, with teraflop-class computing at their fingertips. With the first wave of Tesla-based servers launching earlier this week – and more news from OEMs expected in the weeks ahead – this is an exciting time for the Tesla team.
The Tesla is a high-end product from Nvidia that is meant to serve the General Purpose computing on Graphics Processing Units (GPGPU). Nvidia’s partners have started announcing products based on the latest version of the Tesla. For example, Supermicro has found a way to fit the Tesla M2050 – which is really just a professional version of the GeForce GTX 470 – into a 1U 19-inch rackmount chassis.
Today we saw the first of a series of launches from our partners of new products based on our new Tesla M2050 GPU Computing modules. What’s particularly exciting about these launches is that they are all server products, an important area of the Tesla business and one that’s about to take off exponentially. For the last 3 years we have seen a lot of pilot projects — in banks, in the military, in science — projects that enabled scientists and researchers to experiment with GPUs and see how they can be used to increase the pace and scope of their work.
At the SuperComputing conference back in November of 2009, one could find the still-yet-to-be-released Fermi graphics card running in the NVidia booth. At the recent CEBIT conference, SemiAccurate has spotted a system with four Fermi-based Tesla cards, and even has the picture to prove it. To see a picture of it, you will have to click on through the link below. There is just one small catch however:
Should you want to move to greener pastures, SuperMicro has the servers for you. There were 2 and 4 Tesla/Quadro servers featuring that loveable scamp, Fermi. These were genuine 'puppies', meaning they were not functional, just mockups. There are not enough samples to go around to bring working cards, much less six of them, to a trade show. (As a note to Nvidia PR, SuperMicro was honest when asked about the status of the cards on display. It can be done, no one died!)
Details of the new Tesla K10 and K20 continue to come out, and the K10 is proving to be a bit different than anticipated. The new Tesla K10 is using the GK104 chip, available in the GeForce cards. This chipset lacks most of the usual Tesla features, and NVidia is getting around this by marketing it specifically to a narrow slice of the market:
A group of 100 scientists, engineers, and developers are working together for a bid at Google’s Lunar X Prize, a $30 million award to the first private funded team to land a rover on the moon. Any bid will require tons of work in computations, hardware, and physics, but the German team is beefing up their systems with the power of NVidia Tesla GPU’s.
A few weeks ago Cray CTO Steve Scott made waves in the HPC News media with the announcement of his resignation and move to a “computing partner” that wouldn’t be named. Most people figured they meant AMD. Well, it’s been named now and it’s a shocker: He’s the new CTO of NVidia’s Tesla business unit.
Two new offerings from NVidia today, first the new 
HPC people have been investigating Tesla & other CUDA technologies for several years, but mainly through test & development servers and homebrew configurations. No longer will that be required, as now IBM is jumping in the game by packing the new Tesla M2050 card into their new systems.
VizWorld.com 
Comments