NVIDIA 下一代架构"Fermi" 猜测、讨论专题

Edison

有 GDDR5 后没必要硬拉到 512-bit，比 RV870 大 50% 其实也不错了。

Eji

有 GDDR5 后没必要硬拉到 512-bit，比 RV870 大 50% 其实也不错了。
Edison 发表于 2009-9-21 00:49

不要用頂級的也成啊....XD
反正腳位做了空焊也成，有需要的時候衝上去總是不錯。

但是如果把ECC的需求考慮進去的話，那就沒那麼簡單了：
512bit + ECC就到640bit了，成本太高.... 改成384bit + ECC = 480bit可能比較合算。

Edison

ECC 一般都不会考虑作总线位宽，虽然实际上的确占用了线路。

PS5

512bit的GDDR5显存容量只有2GB稍微小了点，如果有4GB就好了。

Edison

Some info about the Fermi is coming soon:)

http://www.nvidia.com/object/gpu_technology_conference.html

Day/Time	Detail	Room
WEDNESDAY
1:00 PM	Opening Keynote with Jen-Hsun Huang	Regency 1+2
3:00 PM	General Session – Important Trends in Visual Computing	Regency 1+2
4:30 PM	General Session – Breakthroughs in High Performance Computing	Regency 1+2
6:00 PM	Exhibits, Research Poster Preview and Networking Happy Hour	Imperial Ballroom
8:00 PM	Special Opening Night Charity Event	Imperial Ballroom plus other venues
THURSDAY
9:00 AM	Day 2 Keynote with Hanspeter Pfister, Professor, Harvard University	Regency 1+2
11:00 AM	Emerging Companies Opening Address	Regency 2
11:00 AM	Research Summit Fast Forward with Bill Dally	Regency 1
11:00 AM – 6:00 PM	Nexus Lab (Hands-On Tutorial)	Piedmont
12:00 PM	Future of GPU Computing Panel (Emerging Companies Summit)	Regency 2
12:00 PM	Exhibits and Networking Lunch	Imperial Ballroom
3:00 PM	Supercomputing Super Session	Regency 2
6:00 PM	Exhibits and Networking Happy Hour	Imperial Ballroom
FRIDAY
8:30 AM	Day 3 Keynote with Richard Kerris, CTO, Lucasfilm	Regency 1+2
10:00 AM	Fireside Chat with Jen-Hsun Huang, moderated by Jon Peddie	Regency 1+2
10:30 AM – 6:00 PM	Nexus Lab (Hands-On Tutorial)	Piedmont
12:00 PM	Exhibits and Networking Lunch	Imperial Ballroom
3:30 PM	Advanced C for CUDA	Gold
3:30 PM	GPU Revolution in Film Production: Report from the Battlefield	Sacramento

Edison

512bit的GDDR5显存容量只有2GB稍微小了点，如果有4GB就好了。
PS5 发表于 2009-9-23 21:34

还是那句话，比 RV870 快 50% 就不错了。

airforce18

意思就是  384  1.5GB或者是3GB

Edison

可能更多吧，取决于产品型号。

airforce18

民用的我相信  1.5和3.0就差不多了

Edison

嘻嘻，泄密咯 ：

GPU specifications
This is the meat part you always want to read fist. So, here it how it goes:
* 3.0 billion transistors
* 40nm TSMC
* 384-bit memory interface
* 512 shader cores [renamed into CUDA Cores]
* 32 CUDA cores per Shader Cluster
* 1MB L1 cache memory [divided into 16KB Cache - Shared Memory]
* 768KB L2 unified cache memory
* Up to 6GB GDDR5 memory
* Half Speed IEEE 754 Double Precision

Edison

不过之前我所了解的 DP FP 性能应该 1:4，不知道为啥这次公布的资料成了 1:2。

Edison

这下 NVIDIA 也可以列席 10PFLOPS 俱乐部了：

http://www.nvidia.com/object/pr_oakridge_093009.html

SANTA CLARA, Calif. —Sep. 30, 2009—Oak Ridge National Laboratory (ORNL) announced plans today for a new supercomputer that will use NVIDIA®’s next generation CUDA™ GPU architecture, codenamed “Fermi”. Used to pursue research in areas such as energy and climate change, ORNL’s supercomputer is expected to be 10-times more powerful than today’s fastest supercomputer.

目前最快的 Roadrunner 是 1PFLOPS 的级别，这意味着 ORNL 的 CPU+Fermi project将会达到 10PFLOPS。

除了 ORNL外，当前还有其他若干个 10P FLOPS supercomputer 在进行中，例如日本方面富士通和 RIKEN 采用 Sparc64 VIIIfx 的 10PFLOPS 超级电脑以及 IBM 的 blue water 10PFLOPS 超级电脑。

iphone

NV走这一步很冒险。