GPU 101: The Beast In Data Centers

Editor's Notes

• #3: Looking into the future, we can see what direction high performance computing is going. HPC : Computational science Numerical models to understand and predict physical and biological behavior. There are so many laws of nature acting on the smallest scale imaginable. The closer we get to modeling those laws and the world around us, the more powerful our analysis can be. AI : Data science The first step to solving data science problems used to be getting the complete source of information. Deep learning extracts multi-dimensional features from data through layers and layers of connected computational networks. Data science can now solve problems with incomplete understanding.
• #4: AI is revolutionizing our world Intelligent Internet services (search, translation, recommendations) Healthcare DL Farming Arterys, which is bringing deep learning to medical image analysis, earned FDA clearance for its software that provides automated ventricle segmentations from conventional heart MRI scans that are as accurate as manual segmentations performed by physicians, the company announced Monday. The Arterys Cardio DL software automates time-consuming analyses of cardiac MRI images, generating editable contours showing the inside and outside of the ventricles of the heart. The software processes a scan in 10 seconds, far more quickly than a clinician would. The application uses a deep learning algorithm that was trained using data from several thousand cardiac cases, the company said in a statement. The algorithm produces results comparable to those of an experienced clinical annotator, Arterys said.
• #5: With 100 ExFLOPS it would take more than 2 years on a dual socket CPU server to train
• #7: Case Western Reserve University Products NVIDIA® GPUs Summary No cancer patient wants chemotherapy if it isn’t needed. But when doctors aren’t certain, they’re most likely to treat the cancer aggressively. Breast cancer is the second leading cause of cancer death for women worldwide. Gene expression based tests help doctor's determine a cancer’s aggressiveness so they can prescribe an appropriate treatment, for instance whether to not to administer adjuvant chemotherapy. However these tests are tissue destructive, expensive and have turn-around times of 10-14 days. Case Western Reserve is using GPU deep learning to develop an automated risk assessment of routine tissue pathology slides for potentially more accurate and rapid assessment of cancer risk, at 1/20 the cost of current genomic tests. Problem Researchers at Case Western Reserve University are working to eliminate unnecessarily harsh treatment (i.e., chemotherapy) for the most common type of breast cancer. Their work is one of the many ways that GPUs and AI are advancing diagnosis and treatment for breast cancer, the second leading cause of cancer death in women worldwide. Early stage estrogen receptor positive (ER+) breast cancer (BCa) treatment is based on the presumed aggressiveness and likelihood of cancer recurrence. Oncotype DX (ODX) and other gene expression tests have allowed for distinguishing the more aggressive ER+ BCa requiring adjuvant chemotherapy from the less aggressive cancers benefiting from hormonal therapy alone. However these tests are expensive, tissue destructive and require specialized facilities. To determine the cancer’s aggressiveness, doctors use the Oncotype DX test. It determines whether a patient needs chemotherapy – with the hair loss, nausea, fatigue and other side effects that come with it – or hormone therapy, which has milder side effects. The test can easily distinguish high-risk and low-risk patients, but others fall in a sort of intermediate-risk limbo, said Madabhushi. And at $4,000, the test is out of reach for many people, including most in the developing world. Solution To replace the costly test, Anant and his team are using GPU-accelerated deep learning to develop an automated MRI Analyzer test that could speed diagnosis and improve accuracy without destroying the tissue specimen. The estimated cost: $200. Result(s) .The Case Western Reserve approach can make a risk assessment from existing pathology slides. .Automated tubule quantification could be potentially useful in streamlining clinical pathology workflows. The automated quantification aims to standardize the breast cancer grading and risk assessment process and reduce inter-reader variability – thus avoiding having to do genetic testing and avoiding wasted time and cost. Impact -Reduces test cost from $4K to $200 -Speeds Diagnosis and increases accuracy Quote(s) “Anyone who can get biopsy can get our test,” said Madabhushi. “This could have huge global impact.” About Case Western Reserve University (also known as Case Western Reserve, Case Western, Case, and CWRU) is a private doctorate-granting university in Cleveland, Ohio. The university was created in 1967 by the federation of Case Institute of Technology (founded in 1881 by Leonard Case Jr.) and Western Reserve University (founded in 1826 in the area that was once the Connecticut Western Reserve). Time magazine described the merger as the creation of "Cleveland's Big-Leaguer" university. More information http://www.nature.com/articles/srep32706 https://blogs.nvidia.com/blog/2016/10/01/ai-advancing-fight-against-breast-cancer/ http://blog.case.edu/think/2016/02/18/new_image_analytics_may_offer_quick_guidance_for_breast_cancer_treatment https://blogs.nvidia.com/blog/2016/10/01/ai-advancing-fight-against-breast-cancer/ http://www.cdc.gov/cancer/dcpc/data/women.htm 
• #8: Today, for traditional CPUs, the ability to double performance by doubling transistors every 18-24 months (“Moore’s Law”) is coming to an end. He noticed that the number of transistors per square inch on integrated circuits had doubled every year since their invention.
• #9: OpenACC is the only programming model in HPC that can deliver perf portability. Primary goals are three fold: simple, portable, and powerful. Parallel programming is hard. It’s harder when there’s different archs, with different optimization methods. OpenACC’s goal in life is to address this problem by hiding it from developer. Let the compiler do all the heavy lifting that so programmer can worry about their science. Portable. OpenACC runs on all these architectures TODAY! (xeon phi is in “alpha stage” with PGI compiler). And it’s powerful. When you give this tool to developers, they can iterate on optimizations, see improvements immediately, and not worry about it being portable.
• #15: This might look good as waves!