FSB 1600的Xeon要来了

Prescott

还有专为玩家准备的新V8，双socket，4个PCI-Ex16。

http://www.theinquirer.net/default.aspx?article=41868

Intel Skulltrail to break dual-socket performance barrier

100+ real GFLOPs on a desktop, anyone?


By Nebojsa Novakovic: Thursday 23 August 2007, 08:38

INTEL IS pretty much ready with the Harpertown 'cream of the crop' of initial Penryns, and the boards with the new Seaburg chipset, aka Stoakley platform, are about there by the likes of Tyan, Asus, Supermicro and of course Intel itself.

I've myself seen the dual-socket workstations and servers running 3.2 GHz X5482 FSB1600 Harpertown DP on the new chipset ever since Beijing April IDF - even then they were running stable on branded vendor boxes. The recent run at Intel labs handled even overnight non stop benchmark runs with 100.0% per-core on all cores CPU usage fine - not bad for pre-production units. The new boards, by the way, have full copper, fanned, heatsinks on the chipset north bridge, finally - these are hot babies running DUAL FSB1600 at the same time, so more decent heat dissipation is a must.



According to the industry sources in high performance computing, the larger 12 MB caches on these parts, compared to vastly increased net memory bandwidth obtainable on the Seaburg chipset and DDR2-800 CL5 FB-DIMMs, give far higher benchmark results in some cases than the current dual Xeon batch.



For instance, STREAM-type memory bandwidth benchmarks are expected to be between 40% and 60% faster, and you could get around 80 GFLOPs double precision net (Rmax) 40Kx40K Linpack performance out of 102 GFLOPs peak, with the current (not yet fully optimised) code.

This is a ~20% improvement over X5365 Clovertown DP 3 GHz FSB1333, for just ~6% clock jump - and before further tuning that may incorporate some opcodes from SSE, like dot product.

Based on the sources who saw dual-socket Barcelona running, with the most optimised new PGI compilers for the platforms, and after a good round of tuning, it should give up to 55 GFLOPs Rmax out of 64 GFLOPs peak, at 2 GHz. Now, it shows Barcelona is a bit more efficient per clock than the Harpertown, but it also has a 60% clock disadvantage that more than destroys its competitive position at the dual-socket system level. Even if AMD manages to ship a 2.2 GHz bin in October, when Intel is supposed to have X5482-based platforms on the way to the customers already, it still doesn't nearly make up.

Yeah, Linpack is of course a useless benchmark, but unfortunately many workstation or supercomputer tenders all over the world include it as a critical measure - the Top 500 ranking is still based on it.

This is just the beginning, though. Intel's "Skulltrail" high-end dual socket media & gaming platform will use this same Xeon, just in an unlocked Extreme version, and the same Seaburg chipset, but in an optimised lower latency, higher bandwidth configuration with selected FB-DIMMs running upward of DDR2-1000 speed and lower latency like CL4.

The board has FOUR PCI-E x16 GPU slots and, of course, strong power/VRM config with overclockable Dual FSB config on top of the unlocked multipliers. All the chipset and VRM heat sinks are easy to replace for optional water block/freeze block cooling.

Knowing the 4++ GHz air cooling overclock headroom on these 45 nm parts, and even higher liquid or cryo cooling potential, and coupled with fast FSB & memory - I thought "Fast FB-DIMM" was an oxymoron - we could be in for some record results when the monster E-ATX board shows up in the market in the next few months.

After all, an extra 20 - 30% speed boost on CPU, and additional 30+ memory throughput due to optimised FSB latency coupled with both higher bandwidth and lower latency by those upcoming FB-DIMMs, should give a very decent application performance increase, not to mention ability to feed, say, four Nvidia G92 GPUs simultaneously?

One of these records will surely be the above mentioned Linpack - Dual water-cooled 4+ GHz Extreme Harpertown Xeons on a Skulltrail board will be the first ever dual-socket PC / workstation / server to break the 100 GFLOPs net - Rmax - Linpack performance barrier.

Now, just think what wonders will that machine do for gaming physics or, as our Theo reported from Leipzig, real time ray tracing. µ

itany

Intel还是很聪明的，反正工作站平台NV的SLI和AMD的CF都可以使用
到时候就可以看到一片金灿灿的散热器，多YY阿
eATX显然要比4*4硬塞在ATX好得多……

awoo

For instance, STREAM-type memory bandwidth benchmarks are expected to be between 40% and 60% faster

什么时候到桌面来？

delphi5-10

这么高的 FSB，用到桌面平台得配上高频的 DDR3 才合适吧

the_god_of_pig

在得到Nehalem的详细信息前，保留意见:shifty:

BTW:linpack毕竟是残废指标(好在编辑也知道……)，但是spec cpu rate倒是可以借机小小报仇一下:thumbsup:

[ 本帖最后由 the_god_of_pig 于 2007-8-25 13:18 编辑 ]

itany

原帖由 delphi5-10 于 2007-8-25 13:07 发表
这么高的 FSB，用到桌面平台得配上高频的 DDR3 才合适吧

FSB1600高么?不就是双通道DDR2 800么
不过这个是双插座的，所以需要两倍的带宽，用的是超频4通道的FB-DIMM
不过FB-DIMM的热量和延迟一直是心中的痛啊……

itany

原帖由 来不及思考 于 2007-8-25 18:32 发表
INTEL应该收购RAMBUS

现在的高速bus专利，基本被这个B垄断了

GDDR5都得给钱

嘿嘿
有思考大大支持，我的底气就更足了！
当然还有Rock·Will 支持~
w00t)

http://we.pcinlife.com/thread-813046-1-1.html

shike_cuke

桌面上1K6是什么时候??明年年底???

九泉苍月

原帖由 <i>itany</i> 于 2007-8-25 19:45 发表 <a href="http://we.pcinlife.com/redirect.php?goto=findpost&pid=15251766&ptid=813187" target="_blank"><img src="http://we.pcinlife.com/images/common/back.gif" border="0" onload="if(this.width>screen.width*0.7) {this.resized=true; this.width=screen.width*0.7; this.alt='Click here to open new windownCTRL+Mouse wheel to zoom in/out';}" onmouseover="if(this.width>screen.width*0.7) {this.resized=true; this.width=screen.width*0.7; this.style.cursor='hand'; this.alt='Click here to open new windownCTRL+Mouse wheel to zoom in/out';}" onclick="if(!this.resized) {return true;} else {window.open('http://we.pcinlife.com/images/common/back.gif');}" onmousewheel="return imgzoom(this);" alt="" /></a><br />

<br />

<br />
嘿嘿<br />
有思考大大支持，我的底气就更足了！<br />
当然还有Rock·Will 支持~<br />
<img src="images/smilies/w00t.gif" smilieid="46" border="0" alt="" /><br />
<br />
http://we.pcinlife.com/thread-813046-1-1.html

<br />
这帖被偶说了一句IBM后就不了了之