-
00 : 01
© Distribution of this video is restricted by its owner
Transcript
×
Auto highlight
Font-size
| 00:00 | Hi my name is Peru and I'm to introduce you to my research in |
|
|
|
| 00:04 | mini talk. Let's get started My overall research interest is in the |
|
|
|
| 00:11 | , high performance computing HPC area which a victim that says something like computing |
|
|
|
| 00:18 | performance matters a lot mostly. This heavy numerical computations such as simulations of |
|
|
|
| 00:24 | chemical biological social phenomenon at a very scale and with very high fatality. |
|
|
|
| 00:30 | of my main research topic is in extension computations. And the goal is |
|
|
|
| 00:36 | make it fast scalable and energy efficient make use of the latest advancements in |
|
|
|
| 00:43 | hardware such as Gpus clusters, clouds special units such as tension processing |
|
|
|
| 00:48 | And potential core also have some other in algorithms, especially in numerical algorithms |
|
|
|
| 00:55 | computer systems in complete scientific software and a visualization as you can see, |
|
|
|
| 01:03 | research is at the intersection of numerical , HPC software system and libraries and |
|
|
|
| 01:10 | of computer hardware. We investigate the of the of the advancements in these |
|
|
|
| 01:16 | areas. Let me introduce a few snapshots. So the first snapshot is |
|
|
|
| 01:23 | project that I call later short for algebra on terms of course. So |
|
|
|
| 01:28 | is tensile court basically in terms of is a device on your on NVIDIA |
|
|
|
| 01:33 | that are designed to accelerate deep neural . Um it comes it comes with |
|
|
|
| 01:39 | video GPU that appears since 2018. big deal about tensile core is that |
|
|
|
| 01:45 | specializes in computing Matrix matrix multiplication at very high speed hundreds of teraflops. |
|
|
|
| 01:53 | one teraflops means 10 to the power 12 numerical operations per second. So |
|
|
|
| 02:00 | is equivalent to about 1000 CPU cores the same thing at the same |
|
|
|
| 02:06 | They are designed for neural networks. , we'd like to make use of |
|
|
|
| 02:10 | to do more than that to compute sentence for factories. Asians, for |
|
|
|
| 02:15 | , this open doors too much wider in scientific computing engineering and data science |
|
|
|
| 02:21 | , it turns out that to do we need some innovation in numerical |
|
|
|
| 02:26 | software and performance engineering and basically a level concerted algorithmic and performance oriented |
|
|
|
| 02:35 | We are able to achieve 3 to times speed up for major Matrix Factory |
|
|
|
| 02:41 | called Q. R. Factory This is a significant improvement with the |
|
|
|
| 02:45 | of tensile court. This kind of operations are very useful and they are |
|
|
|
| 02:50 | the standard two blocks of every major processor vendors and they are generally highly |
|
|
|
| 02:58 | at better companies because they are the blocks for many, many applications. |
|
|
|
| 03:04 | second research snapshot is on matrix or decomposition and a competition. These are |
|
|
|
| 03:10 | diverse and useful for many problems in learning, signal processing, bioinformatics and |
|
|
|
| 03:17 | areas. For example, in data , a powerful unsupervised model for topics |
|
|
|
| 03:22 | of documents. Gaussian mixture model, recognition recommendation systems and hyper spectral and |
|
|
|
| 03:30 | just maybe a few examples. The is that it's very time consuming and |
|
|
|
| 03:36 | requires a lot of memory to compute scale major extensive decomposition or competition. |
|
|
|
| 03:44 | , so how do we scale and this particular operation? So we do |
|
|
|
| 03:49 | in parallel in distributed memory system with help of GPU and we also borrow |
|
|
|
| 03:55 | innovate uh, the algorithms such as and communication, avoiding in algebra and |
|
|
|
| 04:05 | order optimization algorithms. And we work high performance platforms with M P |
|
|
|
| 04:11 | code A and C plus plus. third snapshot is on the class of |
|
|
|
| 04:18 | models called kernel machines for classification and . This include support vector machine and |
|
|
|
| 04:26 | model and Gaussian process. When you more points, the model fits the |
|
|
|
| 04:32 | better and also reduces uncertainty at the that he has never seen before with |
|
|
|
| 04:38 | two models. The challenges again that becomes quite challenging at large scale. |
|
|
|
| 04:44 | we also try to solve this problem HPC. These are the most relevant |
|
|
|
| 04:50 | for my research and thank you very for listening. |
|
|
|
A simple tooltip for the link