Sign in with
Sign up | Sign in
Your question

Your thoughts on Dempsey......

Last response: in CPUs
Share
November 8, 2005 11:35:13 PM

You beat me to posting the links. lol

I think that the performance while it may not silence critics of Intel, should dampen them down a bit.

The article on 2cpu is meant to give a taste of Dempsey's performance as it compares a 2-way dual core Dempsey setup against a 2-way single core Opteron setup. Before people start posting about the injustice of it all, the article is more focused on how the new 3.46GHz Dempsey compares to the older 3.8GHz Nocona architecture. Most significant is the conclusion that higher latency was not an issue with FB-DIMMs and that it was a good idea for Intel to transition to quad-channel RAM. As well, they also noted that the memory controller was much improved in terms of memory bandwidth.

The German article is much more significant as it pits Intel's fastest Xeon against AMD's fastest Opteron. Going through the benchmarks, Dempsey wins in 6 out of 8 benchmarks. It's also interesting to note that Opteron's L1 cache is faster than Dempsey's in 32-bit transfers, but there L2 cache and RAM is slower. Both processors are similar in 64-bit transfers. Significantly, the Opteron appears significantly weaker in 128-bit transfers. While Dempsey is 47,340 MB/s in L1, 24,928 MB/s in L2, and 3606 MB/s in RAM, Opteron is only 18,360 MB/s in L1, 9448 MB/s in L2, and 3316 MB/s in RAM. This will probably give Dempsey an advantage in SSE calculations.

I hope people don't mind me summarizing the reviews. It's probably helped if you trust me to be unbias as going through the german site is annoying.

In any case, it appears as though claims that Intel will remain "second class" for years is false. Granted Dempsey isn't available yet, but in 2 months Intel will definitely have something extremely competitive out. Although Dempsey already beats AMD's fastest Opteron, Dempsey's performance will be even higher during launch as this it is still in validation stage. As well the lack of proper drivers also hindered its performance.

"For the Dempsey evaluation platform Intel had still no special chip set of drivers ready. Windows server 2003 worked with standard Microsoft drivers. With optimized Blackford /Greencreek drivers the memory throughput should rise."

I'm not sure when new Opterons based on Socket F are coming out, but I know that Socket M for desktop has been delayed slightly and won't be available until late April, early May. It looks like the 65nm process has given Netburst a little bit more life, and Intel some breathing room. I can't wait for Intel's next generation architecture.
November 8, 2005 11:54:24 PM

You can't call this a fair review, in the fact they are comparing Dempsey which won't be out until later next year to what AMD has out now. By then AMD's socket F will be out, and from what i've heard it will be much faster, and the integrated memory controller will be much more improved along with 2-4MB of L2 cache. But we'll just have to wait and see. I have no doubt that AMD will still be leading the way in breakthrough performance a year+ from now. :p 
Related resources
November 9, 2005 12:18:15 AM

I think that Dempsey will have a few months of free space. Intel is trying to get their new 65nm processors out the door as soon as possible, likely in early January, while AMD's new sockets won't be out until mid-Q2. The new sockets were originally scheduled for March but were delayed. That should give Dempsey at least 3 months delivering superior performance over AMD's Opterons.

The fact that people buying Dempsey and the Bensley platform can just drop in Woodcrest when its released in H2 2006 makes Intel's platform the server purchase of choice when its released. Few people would want to buy an AMD platform in Q1 2006 when its going to be replaced the next quarter. From a purchase perspective its also important to note that Bensley is not only dual core capable, but as it supports Woodcrest it should also work with the quad core Woodcrest called Cloverton.

It would also be interesting to see how well AMD's new sockets sell initially. Personally, I feel that unless the performance of the new sockets is spectacular, people aren't going to be willing to buy the new sockets. They would be willing to wait the extra few months for Intel to release its next generation processors before making a decision on which to buy.
November 9, 2005 12:18:18 AM

Here's somthing i found........
http://www.theregister.co.uk/2005/10/29/intel_xeon_2009...

"Dempsey, Intel's dual-core Xeon for servers with two processors, should be very close in performance to Opteron," IDG News Service reported, basing the remarks on comments from Mario Cooper, a manager of HP's ProLiant servers. "AMD's advantage over Intel's designs are more pronounced in servers with four processors, and the Opteron 800 series should have a clear advantage over Intel's Tulsa processor early next year."

Now, however, this tit-for-tat squabble has changed.

As stated, Intel will struggle to match Opteron in the next 18 months but instead of rolling out Opteron killers in 2007, Intel will introduce more processors tied down by its aging architecture dependencies. The company does not have a realistic chance of besting Opteron on typical server benchmarks until the new chips arrive in 2009. By that time, AMD will have new four-core designs of its own and who knows what other innovations."


2009?? Ouch......Like i said, it looks like AMD will have a lead all they way until 2007-2009 atleast. And it's not like AMD is standing still either, AMD has many more new advancements coming out in the next year. (Socket M2 for desktop and F for Server)
November 9, 2005 12:25:34 AM

Bah, no power consumption figures.
November 9, 2005 12:28:24 AM

It's important to note that the delay only refers to Whitefield which is the 4-way server segment. Intel will offer significant competition in the mobile, desktop, and 2-way server markets. In these segments, AMD's architecture doesn't offer a significant advantage of Intel's next-generation architecture. This has been debated in some of the other threads on the forum such as:

AMD's K10 Dead - Intel's Big Opportunity
November 9, 2005 12:33:34 AM

Yeah I know.

AnandTech did a power consumption test on a 3.4GHz Presler though. The results were impressive since the 3.4GHz 950 runs at the same power levels a 2.8GHz 820. This is despite the 950 having 1MB more L2 cache, which is pretty power consuming. Overall I believe the 65nm process produces processors that use only 79% the power levels as on the 90nm. It gives me hope that the 3.46GHz Dempsey will run cooler than the 2.8GHz Paxville.
November 9, 2005 12:35:20 AM

Quote:
I think that Dempsey will have a few months of free space. Intel is trying to get their new 65nm processors out the door as soon as possible, likely in early January, while AMD's new sockets won't be out until mid-Q2. The new sockets were originally scheduled for March but were delayed. That should give Dempsey at least 3 months delivering superior performance over AMD's Opterons.

The fact that people buying Dempsey and the Bensley platform can just drop in Woodcrest when its released in H2 2006 makes Intel's platform the server purchase of choice when its released. Few people would want to buy an AMD platform in Q1 2006 when its going to be replaced the next quarter. From a purchase perspective its also important to note that Bensley is not only dual core capable, but as it supports Woodcrest it should also work with the quad core Woodcrest called Cloverton.

It would also be interesting to see how well AMD's new sockets sell initially. Personally, I feel that unless the performance of the new sockets is spectacular, people aren't going to be willing to buy the new sockets. They would be willing to wait the extra few months for Intel to release its next generation processors before making a decision on which to buy.


Adding more cores to thier current architecture won't help Intel's bandwidth problems any. People won't hop onto Intels dying architecture knowing that there are limitations of speed, and bandwidth and HIGH wattage. By the end of next year AMD will have out a quad core opteron with thier socket F, with an integrated PCI-E contoller along with the Integrated memory controller 3.0 and DDR2.

Anybody else want share thier $.02?

(I'm glad we can have friendly conversations with each another along with disagreements :p )
November 9, 2005 12:37:22 AM

Quote:
It's important to note that the delay only refers to Whitefield which is the 4-way server segment. Intel will offer significant competition in the mobile, desktop, and 2-way server markets. In these segments, AMD's architecture doesn't offer a significant advantage of Intel's next-generation architecture. This has been debated in some of the other threads on the forum such as:

AMD's K10 Dead - Intel's Big Opportunity


I know i was there! :p 
November 9, 2005 12:37:27 AM

Yeah but anandtech didn't include temps. :roll:
November 9, 2005 2:24:30 AM

Quote:
It's probably helped if you trust me to be unbias

Maybe someday, if you become a performance fan, rather than inIntel fanbois.
Quote:
Opteron is only 18,360 MB/s in L1, 9448 MB/s in L2, and 3316 MB/s in RAM.

Wow, that's a big drop from the numbers I've seen. Too bad they didn't include latencies as well.
Quote:
Most significant is the conclusion that higher latency was not an issue with FB-DIMMs

Or atleast not as much an issue as the doubled power usage of FB-DIMMS.
What it really comes down to though, is that Intel just cant compete with the ODMC. You may say something about 4 fsb, but that takes a new platform, is still bottlenecked buy design, and is point to point to point, with high latencies at every point, while the opteron uses direct connect with half the latency, and by design, does not require the mem calls that xeons do. Sorry super Intel fanbois, you are just listening to Intel propaganda.
November 9, 2005 2:31:54 AM

I like having friendly conversations too. Its a lot better than being told by wusy that he is going to rape me.

"He'll be raped by me before that happen, I assue you."

In any case, I'm probably going to get yelled at:
1st because this is such a long post and
2nd because a lot of this is congecture.

All I can say is that all may conjecture is based on true information, and I just felt like speaking my mind. Besides josh_1413 asked for my thoughts, well here they are:

Anyways, I don't still don't think that bandwidth issues are that significant for Intel. It was for the Pentium 4 but the Pentium M seems to need less bandwidth. While the Pentium 4 showed a jump in performance moving from 533MHz FSB to 800MHz, Dothan doesn't seem to benefit significantly from the move from 400MHz to 533MHz. In fact, Dothan doesn't appear to use all of the 533MHz FSB. This would mean that Intels next generation dual core processors are given more than enough bandwidth by a 1066FSB. This FSB would be supplied by 800MHz dual channel memory, which likewise offers more than enough bandwidth. Dual core Woodcrest chips are also going to have even more headroom if needed with a 1333MHz FSB.

The only complication is Cloverton, the 4 core version of Woodcrest. For whatever reason Intel decided to limit it to a 1066MHz FSB. Now if this was 4 completely separate cores it would be severly bandwidth limited. However, it isn't. Cloverton has a 16MB shared L2 cache. This means that not all four cores need to request data. Often, the data a core needs is already available in the L2 cache since it was added there by another core. This will help reduce the need to access the FSB. As well, a lot of overhead is reduced by advanced prefetching techniques. Both Prescott and Dothan already had excellent ones for data prefetching as well as loop detectors to avoid pipeline misses and extra requests for data. This likewise frees up FSB bandwidth. Overall while Cloverton will be limited by bandwidth it isn't as severe as it appears.

Besides, Cloverton appears to be a short lived product. Intel is beginning to be very aggressive with its pace of technology transition. Yonah is coming out in January on a 65nm process and is to be replaced in 6 months by Merom and the rest of the next generation. 6 months after that, the entire Merom family will be replaced by a 45nm generation. Intel's 45nm process looks very promising as it solves leakage problems and will be available for early 2007. Woodcrest itself will be replaced by Harperstown and Cloverton (4-core Woodcrest) will likewise be replaced. Cloverton's replacement will likely feature a 1333MHz as well as even larger L2 caches to hide bandwidth issues. While these features may sound expensive, they won't be since the economics of the 45nm process covers these improvements.

I really don't know what to make of an AMD quad core Opteron. AMD has always been very vague on their roadmaps.

http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=256...

Since the Now column includes their current architecture and the yet to be released sockets M2 and F, the Coming Soon column appears to be what the K10 architecture was suppose to do in 2007. Now it seems the K10 architecture is either canceled or delayed. In any case, we won't see it until 2008.

Now I know that the K10 architecture was designed from the ground up as quad core, does this mean that this is the quad core Opteron AMD is planning to release? Its possible since the earliest estimates I had previously heard for a quad core Opteron was early 2007 which would have been in the K10 timeframe before it was canceled.

However, hearing the vague talk by AMD executives about multicore processors it is quite possible that they also have one based on the K8 architecture. In any case I think Cloverton will beat it to launch. As well, with AMD's 65nm process just starting up and Fab36 still producing 90nm chips, AMD won't likely have volume 65nm chips until the end of 2006. This would mean that a K8 quad core Opteron is more likely at the beginning of 2007. By this time Cloverton's 45nm replacement would already be preparing for launch to give it competition.

On the 4-way server end, things don't really look that bad either. Granted Whitefield was canceled, and with it the integrated memory controller, but Tigerton looks to still free up bandwidth in other ways. Not only will each processor have its own 1333MHz FSB, ie 4 FSBs, but it appears to also incorporate a form of Direct Connect similar to AMD. It seems each processor will also be connected to each other directly so that communication between processors doesn't go through the FSBs. This allows the FSBs to be dedicated to memory traffic. With 800MHz DDR2 to be released with Conroe in H2 2006, its likely that when Tigerton launches in 2007 it will use quad channel 800MHz memory. By 2007 1066MHz DDR2 is also possible. With Intel hinting that Hyperthreading wasn't a total waste and isn't gone for good, its possible that a 2nd generation Merom, Conroe, Woodcrest architecture like Tigerton could have it incorporated. Needless to say I'd like to see a 4 processor quad-core Tigerton system with Hyperthreading. Certainly 32 threads running at the same time would be amazing. I believe that Windows currently allows up to 32 processors so it would work.

http://www.informationweek.com/story/showArticle.jhtml?...

I await angry anti-Intel or anti-me comments. :wink:
Hmm, this wink looks more like a squint.
November 9, 2005 6:07:18 AM

A lot depends on how they execute the direct interconnect on tigerton, and when the utilize it elseware. That could be the catchup point, but only if Amd sits on their hands. They have shown a penchant for that, so, ya, some time in 07, or 08 seems practical. I'll hope for sooner, but reality says probably not.
November 9, 2005 7:17:21 PM

We'll know in a year from now which ones are faster. My money is on Opteron. :p 

ITcommander_data what do you do for a living?
November 9, 2005 11:12:21 PM

I really probably shouldn't be saying with all those internet privacy stories on the news. Anyways, I'm a student and I won't say my age or anything more specific.

On another note, does anyone know why the K10 was canceled or delayed? The Inquirer just mentioned it and hasn't said anything since. I know people have thought that AMD was so far ahead they could just wait a year, but that doesn't really make sense. I could see them delaying a quarter or half a year, but not a full year. From a business perspective you would want a new architecture to come out as a response to Intel's next generation architecture whether you feel its needed or not. Otherwise, no matter how well AMDs current architecture performs, the average customer would still feel that its "old".

As well, what types of features were to be incorporated in the K10? AMD as always is very vague. The only fact I know for sure was that its designed as a 4 core processor. It will no doubt be built on a 65nm process. It seems logical that it will include newer versions of Hypertransport and most likely DDR3 support. However, AMD really hasn't come out and said yes it will be on it. I've always found it interesting that Intel is always held to a higher standard than AMD. While Intel's roadmaps aren't a manual, they are generally more detailed than what AMD provides. In any case, now with K10 cancelled or delayed, AMD is supposed to be replacing it with the K8L. I wonder what features from the K10 are grafted onto the K8 to make the K8L?
November 10, 2005 6:56:51 AM

Quote:
I've always found it interesting that Intel is always held to a higher standard than AMD.

Boy you really are a noob aren't you? You must have missed the xp2800 release, and the Barton release.
Mind you, Amd has never had a string of missques or failures as long as the one we are seeing now from Intel.
K 10 is a major redraw. It was supposed to include quad cores, virtuallisation tech, and a major speed step. Some of these will be incorporated into 8L.
Amd is not in a position where they could take Intel by storm. Too much demand could sink them. A hostile takeover sponsored by Intel is an ongoing threat. Getting too far ahead is a guaranty of death, and so we wait.

Oh btw dual cores were supposed to be K9, but that went to the dogs.
November 10, 2005 2:09:15 PM

Dempsey's a good defenceman

but not earning me enough points on my hockey pool
!