NVIDIA GTC 2010 直播 发布未来两代 GPU代号开普勒与Maxwell 及性能目标

Edison

速度的确很快（当然，是相对于 CPU 来说），想象一下你（用传统 CPU 方式）花了 30 分钟渲染好一个画面后，拿给客户看，客户看了后说，我想换个角度看看...结果：

10:15
michael's taking a rendered scene and adding objects to it, like a glass vase that's actually reflecting back scenes in the room. now he's changing the time of day, all in real time.

Edison

10:16
jhh thanks ken p and michael and they head off, feeling pretty good about the audience's reception, which included some spontaneous applause.


10:17
JHH is shifting gears. He announces three new OEMs with Tesla systems in production:

10:18
the first is IBM, with its IBM BladCenter, which now incluldes Tesla. Another is Cray, with its T-Platforms TB2; a third is Cray, with its XE6.

10:19
Tesla's rapidly reaching a broad footprint, he says.

10:19

So, what's next? Another gear shift....Over to Adobe, which is democratizing digital photography and editing. The stuff they're working on will blow your mind, in jhh's phrase.

10:20
Two Adobe heavy hitters come on stagej, david salesin, its' senior principal scientist, and todor georgiev, a senior research scientist.

Adobe 的人来了。




10:21
David is recalling trips he took in the past and apparently wasn't such a good photographer. he was frustrated by the poor images of his pics, which included lots that were out of focus.

10:22
there's a lens they're showing, a plenoptic lens, which captures multiple views of a scene from different view points. each is high res. and with software you can stitch these together into one high res image.

10:24
so, instead of taking one image of a subject, you have little lenses that take lots of smaller pics -- it creates a 4d image that allows computation to take the place of optics. after the fact, you can change the angle you're viewing and even the focus.

10:24
gpus can stitch these together 500x faster than cpus.

Edison

Adobe 的人过来不是吃素的，这次他们搞的 demo 类似于 microlens 的镜头。现在的相机像素数量可以说是已经超出一般人的应用需求（动辄几千万像素了），但是如果你使用的是 microlens 矩阵，虽然分辨率下降，但是却相当与能把更多的场景画面拍摄进去，经过一些简单的处理，甚至能让你的画面做些微的转向，这是非常疯狂的 idea。













10:26
it used to take an hour to distill a photo from this and now it's exceptionally fast.

10:27
he's showing a photo of a girl taking a picture. the focus on the picture was wronog, it was on some trash cans, not on the cute little girl. they're changing the focus to the girl. now to the branches near her. this is the triumph of algorithems!

10:28
whoah, the adobe guys are talking about this getting done on video, which would give huge flexibility

10:29
jhh: so far, digital photography has been about digitizing analogue facts. now, with plenoptic lenses, you can gather far more information and do enormously more. this moves optics into the computer!


10:30
we're now commaneded to put our 3D glasses on. and it turns on the picture that we were looking at is 3D. they're even changing the focus of 3d pics.

10:30
"you guys are going to change the world," jensen tells them, as they scamper off.

Edison

医学界梦寐以求的实时立体成像远程心脏手术看来不远了。





10:31
Nice segue here. jensen talks about a doctor changing the world.


10:32
he introduces dr michael black of the california pacific medical center, where he's chief of the congenital heart program. he was bored, he said, and began to get into using robotics for surgery.

10:33
the good doctor is recounting how he operated on a 22-week baby that weighed under a pound.


10:34
back to 3D glasses, which everyone has donned. the video didn't work. ughh.


10:34
now it shows up!

10:36
they're talking about surgeons using small incisions and gpu imagery that effectively makes it appear that a beating heart has stopped. except it hasn't. thus, heart surgery can be done robotically on beating hearts. this would open up the door to procedures that aren't yet even possible.


10:37
what's the implication of this, jhh asks? the good doctor says you can operate on a beating heart and sending a patient home the next day. or, you can perform surgery remotely on an astronaut. eventualy, little robots will work in the body.

10:39
Very cool stuff. It can be used not just on beating hearts. You could operate within the brain, on very small fetuses, it's all dependent on visual information, and with gpus we can replicate this, the doc sez.

Edison

发布下一代 GPU 架构代号开普勒以及下下代 GPU 架构代号 Maxwell

10:43
we talked about how parallel computing is going to change the world. then, how momentum is accelerating, third how we're reaching an ever-larger footprint, and finally we gave a peek into the future.

but the question is, where are we in the advancment of computing


10:43
for the first time-ever in NVIDIA's history, we're going to share with you our direction:


10:44
today, jhh says, we're shipping fermi. kepler is coming in s2011 and maxwell in 2013. these are codenames

10:45
fermi is just the beginning, he says. he shows a chart plott time and performance per watt. "perf per watt" equals perf. Fermi is 1.5 double precision gigaflops per watt. kepler, based on 28 nanometers, will be 3-4 times per watt of Fermi; we will invest a couple of billion in r&d on it.

Fermi 的每瓦双精度性能是 1.5 GFLOPS，基于 28nm 的开普勒每瓦双精度性能将是 Fermi 的 3~4 倍。


10:46
we're constaintly learning about barrier and walls to achieve the speed of light of parallel computing.


10:47
the fanatical drive of perf per watt will take us way beyond kepler wtih maxwell. we believe we will continue to be on this ramp. with parallel computing transistors are free, power's not.

10:47
maxwell will bring 16 gigaflops per watt. this will be 16x improvement over fermi for parallel computing.

maxwell 将实现每瓦 16 GFLOPS 性能， 在并行计算性能将是 Fermi 的 16 倍以上。




10:49
between now and maxwell, we'll introduce virtual memory, pre-emption, enhance the ability of gpu to autonomously block. these will take gpu computing to the next level, along with a very large speed up in performance.


10:49
JHH winding down now. GTC is about you, he says, it's about the democratiziation of parallel computing. Thanks for support and have a great week.


10:50
music goes back up, jensen moves off and now dan vivoli comes up on stage to talk about logistics. They both show off their tattoos.


10:51
that's it for the live blogging. i had great fun and may stick with this gig!

gzpony

开普勒和麦克斯韦都点名了。

看来不会太久会见到爱因斯坦。

而法拉第，薛定谔，波尔，狄拉克这些也有可能登场

而牛顿好像是苹果的注册商标，80年代注册的，现在不知过期了没有。

Edison

直播完毕

qc8241

回复 29# Edison


    继续顶

gzpony

直播完毕
Edison 发表于 2010-9-22 01:55

    版主请点评一下

Edison

E大对技术的熟悉，实在让人敬佩。能爆这么多前瞻性的料。
对于远程心脏手术，不是很理解和N有啥关系。N的3 ...
nemol_xc 发表于 2010-9-22 01:42

这么说吧，MRI 等医学成像技术其实现在也比较成熟，但是要做到立体实时显示却还是非常难，GPU 在这个上面就是展示了一些实力，医生不需要再只是看 2d 画面操作，而是好像 ironman 里的立体成像技术那样，透过三维界面操作机械臂。

Edison

版主请点评一下
gzpony 发表于 2010-9-22 02:02

要冷静    。

WG_Baby

"GTX490" and "GF110"?????哪裡去了。。。

gzpony

这么说吧，MRT 等医学成像技术其实现在也比较成熟，但是要做到立体实时显示却还是非常难，GPU 在这个上 ...
Edison 发表于 2010-9-22 02:03

      可以改变世界的GPU通用计算技术。

AMD和intel虽说有CPU，动作慢一些，但还是会跟进上来的

gzpony

"GTX490" and "GF110"?????哪裡去了。。。
WG_Baby 发表于 2010-9-22 02:09

    看E大的宣讲内容，这次主要是讲解概念技术，也就是远景啥的。
不是产品发布，这么具体的怕不会出现。

clawhammer

直接上GF200吧

pcpvp

用CPU跑CUDA C有什么意义吗，能比C好多少吗

ifu

谁来简单总结一下开普勒到底游戏性能 到底会提升多少？ 双精度浮点性能心脏手术等和普通人又搭不上边。。。

这算是换个方式重演费米纸面发布阻击5系列的情形么？赶在6系列发布前搞幻灯片发布，这次猛把下代和下下代开普勒Maxwell都搬出来了

yangdaye03

我擦，nv太nb了。