8 路小蜜蜂 GPU 的渲染服务器

Edison

http://www.xbitlabs.com/news/graphics/display/20110914145504_Intel_Demonstrates_8_Way_Multi_GPU_MIC_Knights_Ferry_System.html

At the Intel Developer Forum 2011, Intel Corp. showcased a system running a new version of its real-time “Wolfenstein” ray-tracing demo. In order to render the game scenes in full HD (1080p) resolution and with additional visual effects, Intel had to use a 2-way Xeon system with eight code-named Knights Ferry compute accelerators.
The up-to-date Wolfenstein game is rendered through a real-time ray tracer with several special effects that haven’t been possible before in games with such an accuracy. The latest Wolfenstein game demo also includes such effects as depth of field, HDR bloom, inter-lens reflection, smart anti-aliasing.

Las year Intel rendered Wolfenstein demo using a “cloud” of four Intel's Knights Ferry/Aubrey Isle compute accelerate cards in 720p (1280*720 resolution). This year Intel decided to install eight KNF compute cards and render the demo in 1080p (resolution). As ray tracing is a highly parallel application it can therefore take very good benefit of the many cores that are in a single chip on the Knights Ferry board.

The demonstration is a proof-of-concept of real-time ray-tracing as well as many Intel core (MIC) architecture. However, the show off does not have a lot of practical meaning since real-time ray-tracing will hardly be used in the consumer applications for many years.
The Knights Ferry has 32 x86 cores clocked at 1.2GHz and featuring quad-HyperThreading. The unit, aimed at PCI Express 2.0 slots, has up to 2GB of GDDR5 memory. The chip itself has 8MB of shared L2 cache, which is quite intriguing by itself since highly-parallel applications do not require a large on-chip cache.
According to Intel, eight KNF compute boards (on 2-way Intel Xeon X5690 [3.46GHz, 6 cores, 12MB L3 cache] with 24GB DDR3 RAM at 1333MHz) achieved 7TFLOPS performance in SGEMM [single precision general matrix multiply] operations, which is not a truly high result.

asdfjkl

achieved 7TFLOPS performance in SGEMM [single precision general matrix multiply] operations, which is not a truly high result.
这是关键！

mooncocoon

8路……
NV和TG顿觉INTEL良心大大滴坏了{titter:]

disruptor

8路的单精度，换做580，双精度都比这强啊

Edison

这里是 SGEMM = 7TFLOPS 而不是理论峰值 7TFLOPS。

448SP GF100 C2050 的 SGEMM 也就是 580GFLOPS。

dracocephalum

怎麽這機器這麽像是 Tyan 的 FT77B7015 ?

xmap

好吧，我只是想吐槽一下这糟糕的画面感。。。

fengpc

光线追踪啊，像是geforce2年代的宣传亮点~~要比单精度理论性能啊，拿个A卡出来立刻让他满地找牙了~~等开普勒正式发布，intel这卡立马浮云了