Tuesday, August 01, 2006

IBM: Socket F Opteron 21% faster

While details are scarce, IBM claims that the socket F opterons performs 21% better than socket 940 in integer and floating point performance. We should see some nice benchmark numbers soon. IBM is very good at getting the highest scores.

As expected, IBM has unleashed five Opteron models, LS41 4way blade, LS21 2way blade, and x3755, x3655, x3455 rackmount servers. I think the x3755 is 4P, and x3655 is 2U 2P.

Some reporters asked the dumb question of why IBM didn't give more details. The answer is so obvious: AMD hasn't officially launched Socket F opteron yet. IBM was just too anxious seeing those big SUN boxes taking market share. There was an early report that AMD was going to launch Socket F on August 1st, but delayed to August 15th. I wonder who AMD was waiting for.

SUN is still quite ahead with its x4600 (8P 16 core) that smashes 16P HP Itanium 2 Superdome.

Larry Singer, Sun's senior vice president and strategic insights officer, commented on IBM's move: "We wish them well. In a couple years, they'll get competitive, and of course by then we'll have our next generation of servers out."

Indeed, SUN's x4500, x4600, and its 8400 blade are engineering wonders designed by Andy Bechtolsheim.

47 Comments:

Anonymous Ho Ho said...

IBM claims that the socket F opterons performs 21% better than socket 940

The real wording was can deliver up to 21 percent greater performance within the same power envelope as previous generations

What that basically means is that pwerormace per watt is 21% better, not neccesarily the overall performance.

My own guess is that as DDR2 consumes about half the power of DDR1 and LGA sockets are also a bit more efficient than regular pin-based ones as in s940 then big chunck of that energy efficiency comes from these two things.

11:25 AM, August 01, 2006  
Blogger Sharikou, Ph. D said...

What that basically means is that pwerormace per watt is 21% better, not neccesarily the overall performance.

Well, you may interpret that way, but I think no rational person will read it like that. The sentence states
1) 21% more performance
2) same power envelope

If it was like what you said, IBM would say: same great performance with 21% less power.

11:30 AM, August 01, 2006  
Anonymous Ho Ho said...

If it was like what you said, IBM would say: same great performance with 21% less power.

Exactly. I was only pointing it out so that people wouldn't get false ideas. I've seen too many people who assume that whenever someone sayst X berforms better than Y they automatically assume that means X is faster.

There was an early report that AMD was going to launch Socket F on August 1st, but delayed to August 15th. I wonder who AMD was waiting for.

I think I read from somewhere that although the CPU's were basically ready it wouldn't make much sense to release them without proper systems built around them.

11:38 AM, August 01, 2006  
Anonymous Ho Ho said...

I forgot to clarify that this better performance for same power most likely comes from higher clock rates. As system power usage as a whole dropped they now could increase CPU voltage to get higher clocks.

Oh how I miss the "edit" button. If it were here I wouldn't spam that much :)

11:41 AM, August 01, 2006  
Anonymous Anonymous said...

Ohh i so want a socket F opteron sytem:)
Too bad i cant Afford it :P


It be interesting to see the memory bandwith effeciancy of ddr2 on socket f versus socket 940.

On socket Am2 its still too low in my humble oponion. I think is about 60% where as socket 940 and 939 latest amd cpu have almost 100%

11:50 AM, August 01, 2006  
Anonymous Ho Ho said...

and 939 latest amd cpu have almost 100%

Not exactly but close. In synthetoic benches it gets up to 90%.

12:13 PM, August 01, 2006  
Anonymous nosredna_ekim said...

"I wonder who AMD was waiting for."



I wonder who Intel is waiting for to release the core 2 duo in full capacity... Oh wait... I almost forgot Duh.. its INTEL...

This Extreme AMD fanboism on this website must be getting to my head....

12:13 PM, August 01, 2006  
Blogger Phillip Fayers said...

Who or what are AMD waiting for?

I don't know but I reckon Sun will be there with AMD on launch day to talk about new systems. One of the first rumoured launch dates for Socket F systems turned out to be the date of the last Sun quarterly conference where they launched the x4600 et al. I don't think that was coincidence.

12:24 PM, August 01, 2006  
Anonymous Ho Ho said...

I wonder who Intel is waiting for to release the core 2 duo in full capacity... Oh wait... I almost forgot Duh.. its INTEL...

What do you think is simplier, reconfiguring labs to produce entirely new core or produce old core with new memory controller and new socket type?

12:26 PM, August 01, 2006  
Blogger Sharikou, Ph. D said...

What do you think is simplier, reconfiguring labs to produce entirely new core or produce old core with new memory controller and new socket type?

Depends on how smart your processing technologies are. AMD's APm3.0 is fully automated. The same FAB is making 100 kinds of CPUs at the same time. Intel is dumb copy exact: first a development FAB works out the recipe, then the production FAB copies exactly, including the size of Windows.

12:35 PM, August 01, 2006  
Anonymous Edward said...

"What do you think is simplier, reconfiguring labs to produce entirely new core or produce old core with new memory controller and new socket type?"

First, Core 2 is NOT entirely new core. It's a wider Core Duo with some twicks (memory disambiguation, branch prediction, etc), which is basically two Pentium-M's with a shared L2 cache. Core 2 Duo to Pentium-M is about the same as Athlon64 X2 to AthlonXP.

Second, Core 2 uses the same fabrication technologies as the 65nm Pentium-D. AFAIK it doesn't require extra layers, nor particular more complex. The design efforts could be higher, but the production efforts should be the same.

Third, nobody at AMD is waiting. I'd say the design team at AMD did much more per person than any tream at Intel. The Opteron went production in 2003, then the Athlon64 and Turion64, then dual cores in 2005, then AM2 and Turion X2. Keep in mind that AMD is less than 1/10 of Intel's size between 1/4 to 1/5 to Intel's x86 alone.

Now, what is Intel waiting for with its world largest x86 task force? It can't produce enough of its new chips that can compete with AMD?

1:05 PM, August 01, 2006  
Anonymous Anonymous said...

I wonder who Intel is waiting for to release the core 2 duo in full capacity... Oh wait... I almost forgot Duh.. its INTEL...
And we are still waiting on Intel for that. The difference is that AMD may be waiting on themselves, or on their partners, or both, for their release of socket F Opterons. At launch the volume of products from all vendors will give the answer though.

1:26 PM, August 01, 2006  
Anonymous Ho Ho said...

First, Core 2 is NOT entirely new core. It's a wider Core Duo with some twicks (memory disambiguation, branch prediction, etc), which is basically two Pentium-M's with a shared L2 cache.

If we start to generalize like that then I can say that every current x86 CPU is just a little upgrade from 368.

If you want to see the differences of core architectures between Netburst, Core and Core2 you can read this article:
http://www.realworldtech.com/includes/templates/articles.cfm?ArticleID=RWT030906143144&mode=print

1:32 PM, August 01, 2006  
Anonymous Anonymous said...

"What do you think is simplier, reconfiguring labs to produce entirely new core or produce old core with new memory controller and new socket type?"

it does not matter. it also does not matter is it copy exact or apm3.0

what matters is tech process (which is the same in both cases, 65 for intel, 90 for amd) and lith masks, which are new in both cases.

so theoretical speed of ramp should be equal.

2:58 PM, August 01, 2006  
Blogger Sharikou, Ph. D said...

First, Core 2 is NOT entirely new core.

Put it simply, CORE2 is just a faster Pentium III, designed by Bob Colwell many years ago. CORE2 is missing all the important technologies you found on K8

1) CORE-CORE communication
2) CORE-MEMORY communication
3) CORE-IO communication
4) CPU-CPU communication

On Core2 and Pentium 3, all you got is a crappy FSB, and your only hope is adding more cache.

3:01 PM, August 01, 2006  
Anonymous enumae said...

Sharikou, Ph. D said...

2) CORE-MEMORY communication

I am not sure so maybe someone could elaborate a little, but wouldn't a shared L2 cache be considered "CORE-MEMORY communication"?

3:28 PM, August 01, 2006  
Anonymous Edward said...

"If you want to see the differences of core architectures between Netburst, Core and Core2 you can read this article:"

Right... and exactly as I said, Core 2 (compared to Core) is wider, and with better memory disambiguation, and Micro op fusion WAS in Core Duo. Just look at how similar these two are.

5:02 PM, August 01, 2006  
Anonymous Anonymous said...

I am not sure so maybe someone could elaborate a little, but wouldn't a shared L2 cache be considered "CORE-MEMORY communication"?

Core-Memory as he speaks of is a direct link from a CPU to a bank of memory. That CPU gets all the bandwith of the memory, no more, no less, although some will be communicated externally if another core needs to access that memory bank. On Intel CPUs, because they use a FSB architecture, all memory accesses route through one bus and basically get timeshares of the bandwidth of the memory bank.

The shared L2 cache is basically a way to optimize how much cache each gets, where one core can use all,part,or none of the L2 cache. However, being shared. I am not 100% sure, but depending on the application, for cache coherency to work, if a core modifies a value in the shared cache, it may or may not have to write-back to memory before the other core can access it. If it didn't need to write back to memory, the shared L2 couuld in some applications give performance increases in effective memory latency, although I would label that more as core-core communication rather than core-memory.

5:03 PM, August 01, 2006  
Blogger Sharikou, Ph. D said...

I am not sure so maybe someone could elaborate a little, but wouldn't a shared L2 cache

No. That's cache-FSB-chipset-mem controller-memory communication. Just like what Pentium 3 works

5:05 PM, August 01, 2006  
Anonymous Edward said...

"wouldn't a shared L2 cache be considered "CORE-MEMORY communication"?"

Cache helps reducing memory access latency but not increasing bandwidth. Large caches are good for benchmarks that have lots of computations per byte data.

On the other hand, if you just perform simple transformation on large (in size of in number) photos, having a big cache really doesn't help you at all.

Those benchmarks that show Core2's 'superiority' ALL fall into the first category.

5:14 PM, August 01, 2006  
Anonymous Anonymous said...

"Intel is dumb copy exact: first a development FAB works out the recipe, then the production FAB copies exactly, including the size of Windows."

Copy exact is configure the equipment (and facilities) the same from one fab to another for a technology node; it has nothing to do with switching between product where there are different production tool recipes and litho mask sets to switch from product to product.

There are feed forward systems at every fab in the world to enable fast switching between product A and product B. Most tools at Intel and every other place in the world switch seemlessly between products with a lot cascade - this is doen by factory automation systems feeding the correct recipe and at time metrology info to the production tools in the fab.

By most industry benchmarks Renassas is probably world class at this(better than either AMD or Intel). AMD/Intel are considered (HVLM) high volume, low mix manufacturers.

As an earlier poster pointed out most products are produce on the same set of equipment with zCin a given technology node. Copy exact just manages equipment differences between different geographic locations (and differences between equipment of the same type within a fab) to ensure yield/performance at one site is matched to others.

5:18 PM, August 01, 2006  
Anonymous Edward said...

"If we start to generalize like that then I can say that every current x86 CPU is just a little upgrade from 368."

You must know better than this. 386 is just a pipelined version of 286. it's doesn't even do superscalar (Pentium) or OoO (Pentium-II).

When I say Conroe/Meron are just enhanced Pentium-M with larged shared cache, I am NOT generalizing. The basic datapath structure are the same; Core 2 has (again) wider datapaths, better memory disambiguation etc., and large shared L2 cache. No, you cannot same the same thing on Pentium-4, nor any other CPU on the planet!

5:19 PM, August 01, 2006  
Anonymous Anonymous said...

To use your words the "rational" person would not take the words "up to 21%" to mean an average of 21% more performance they would take it to mean a maximum.

It could mean one application was 21% faster and the rest were far less. As you have parsed the Conroe vs FX-62 benchmarks in a similar way I would be careful or you might end up saying "only 21% faster on this particular benchmark all others are within X%".

5:23 PM, August 01, 2006  
Anonymous Anonymous said...

And yet, despite all those 'disadvantages' C2D manages to outperform A64s by 20% clock for clock.

5:29 PM, August 01, 2006  
Blogger Sharikou, Ph. D said...

Conroe 2.93GHZ is about 10% faster than FX62 (2.8GHZ) see discussion at


this link
. There is no doubt that Intel lied when they hyped 40% lead during IDF.

If you look at number crunching performance, a 2.6GHZ k8 is faster than a 3GHZ Woodcrest.

5:33 PM, August 01, 2006  
Anonymous Anonymous said...

Sharikou - stop spreading misinformation. Intel NEVER stated that Conroe would be 40% better than AMD, they said 40% better than previous Intel product (P4). Most Intel fans and websites convenientlty re-interperted this to mean 40% better than AMD.

Please provide information where INTEL stated at IDF that Core would be 40% better than AMD. (Hint - there is none)

5:44 PM, August 01, 2006  
Anonymous Anonymous said...

Read the link a little better Sharikou:

1) It says up to 21%. I can say that Conroe is up to 80% faster than an FX-62 by pulling out the right benchmark but that doesn't mean it is 80% better. Conversely you could say Fx-62 is up to X% (don't recall the eact amount) better than Conroe based purely on Sciencemark benchmarks.

2) IBM is comparing this to previous products - there is no way of knowing what other components on the system have been changed which might also impact performance.

This post seems to be heavily grasping at straws compared to some of your others...

6:11 PM, August 01, 2006  
Anonymous Anonymous said...

Well, you may interpret that way, but I think no rational person will read it like that.

What do you think does "can" mean?
It obviously doesn't mean that it always performs 21% faster but in some cases.

Sharikou, you remain a complete idiot.

6:36 PM, August 01, 2006  
Anonymous Anonymous said...

Well, you may interpret that way, but I think no rational person will read it like that.

What do you think does "can" mean?
It obviously doesn't mean that it always performs 21% faster but in some cases.

Sharikou, you remain a complete idiot.

6:37 PM, August 01, 2006  
Anonymous Anonymous said...

Yeah, I'm pretty sure Intel said 40% faster compared to Netburst, not AMD64.

Dr Sharky, I'm sure you could gain more credibility if you stopped cherrypicking your results.

Heck, why don't you try HardOCPs gaming article, it seems FX-62 is just as fast as X6800 in GPU limited cases!

In pure CPU grunt Core2Duo is by far stronger than AMD.

To put it into perspective, enthusiasts generally agree that a 2.4GHz E6600 is slightly faster than a 2.8GHz FX-62.

This suggests that clock for clock Conroe is about 20% faster.

6:48 PM, August 01, 2006  
Blogger Sharikou, Ph. D said...

I'm pretty sure Intel said 40% faster compared to Netburst, not AMD64.

I was refering to Intel's fraud at IDF which showed Con 6700 40% faster than FX62.

7:14 PM, August 01, 2006  
Anonymous Wirmish said...

Rev F Opterons have been shipping for several weeks, but have not been formally launched.

http://www.thechannelinsider.com/article/AMD+Intel+Brace+for+ThirdQuarter+Showdown/184942_1.aspx

8:46 PM, August 01, 2006  
Blogger Sharikou, Ph. D said...

It says up to 21%.

When IBM says 21% faster, it is on something meaningful. In fact, IBM had a footnote on what it reached the 21% number: "For compute-intensive high performance environments, based on integer and floating point results of internal lab measurements with IBM products."

Intel fanboys dance on SuperPI scores. In real applications, a Con XE 6800+ is only 10% faster than FX62. In server performance, a 3GHZ Woodcrest is only 2% faster than a 2.6GHZ Opteron (see GamePC Apache bench). On FP performance, a 3GHZ woodcrest is slower than 2.6GHZ Opteron on both SpecFP and SpecFP_rate. Intel is a fraud.

9:20 PM, August 01, 2006  
Anonymous enumae said...

You have probably already seen this...

New IBM Servers Based on Old Opterons

10:08 PM, August 01, 2006  
Blogger Azary Omega said...

'''Sharikou, Ph. D said...

First, Core 2 is NOT entirely new core.

Put it simply, CORE2 is just a faster Pentium III, designed by Bob Colwell many years ago. CORE2 is missing all the important technologies you found on K8'''


I think pentium 3 was 'almost' a copy of pentium pro. Also, the difference between core 2 duo and core duo is larger (on some models) cache, faster FSB, wider pipeline (goin from 3 instruction to 4 per clock, only translates in to 10-15% performance increase, because of the often found branch instructions, and goin from 4 to 5 instructions per tick is even less than 10-15%). Still wanna tell me core2 is that much better than core1 just because of all above? I dont think so. Chances are, a multi billion dollar corporation (intel) paid big bucks for advertising a product on which they whole existence depends = 99.99989%

10:32 PM, August 01, 2006  
Anonymous Anonymous said...

Even IF we assume the far-fetched situation that Revision F is 21% faster than E on ALL benchmarks, how come AMD is so slient about it? Especially in the face of the torturous Conroe hype it would definitely help them if they told everyone about Socket-F. In fact at the analyst meeting from AMD, I distinctly remember Hector being asked what they will do about core2, he said "I fully expect the performance crown to exchange hands every few years..." Gee, I would have thought he would say- "just wait for socket F".

Anyway, more interesting, Socket-F launch is only a few days away, now tell us Shakira, what will you do if AMD releases Socket F and does not claim your 21% all-round performance advantage over Rev-E?

10:50 PM, August 01, 2006  
Anonymous Anonymous said...

Anyway, more interesting, Socket-F launch is only a few days away, now tell us Shakira, what will you do if AMD releases Socket F and does not claim your 21% all-round performance advantage over Rev-E?


it was ibm who said socket f is 21% faster, dr. sharikou never claimed such.

reading comprehension problems i surmise.

2:52 AM, August 02, 2006  
Anonymous Edward said...

"I think pentium 3 was 'almost' a copy of pentium pro."

Nope. P-II was an improved P-pro. Well, but P-III was an improved P-II though.

But none of them was almost a copy of its predecessor.

7:56 AM, August 02, 2006  
Anonymous Anonymous said...

"it was ibm who said socket f is 21% faster, dr. sharikou never claimed such.

reading comprehension problems i surmise."

Maybe you should heed your own comments, IBM stated that it CAN PERFORM UP TO 21% BETTER. This means:

A) It is at least 21% better on ALL applications
B) It is 21% better on average
C) It is 21% better on SOME applications, but possibly less on others.
D) Sharikou is yet again creatively re-interperting things
E) We don't really know because there was no actual data published

While there was no actual data, IBM is respectable company so my vote is on "C".

1:31 PM, August 02, 2006  
Anonymous Anonymous said...

"I was refering to Intel's fraud at IDF which showed Con 6700 40% faster than FX62."

Show me any link or article on this and I will apologize.

Madmodmike claimed the same thing and l I kept pushing him on it until he finally relented saying something to the effect "Intel said it would crush (or outperform?) an FX62" (I don't remember his exact words) I have seen no published claims by Intel that they stated a 40% improvement over FX62.

1:35 PM, August 02, 2006  
Anonymous Anonymous said...

Why do people spend so much time debating on whether an architecture is an evolution or re-design of a previous one. It is just semantics.

If the chip delivers the performance who cares if it was an evolution or a ground up design. Nor do I care who copied whom.

That's like saying my Toyota is better than your Ford because it's got this new cool Ceramic engine design and gets 35 mpg and goes 0-60 in 6 sec. Even though your car gets 38mpg and goes 0-60 in 5 sec, it is only an evolution of a much older, archaic metal engine design. If reliability and cost of ownership are similar do you really care what your engine is made of or do you about the performance output?

This is similar to the process techology debates - if you look at Si processing Intel and AMD rarely invent new process technology (whether it be SOI, strained Si, high K gate oxides, low K, Cu, etc...). This is all generally based on work done by universities, consortia, equipment suppliers, etc. What you should care about is what the final performance of the integrated process is (and how that translates into chip performance), not who has the better strain technology or SOI.

1:54 PM, August 02, 2006  
Anonymous Anonymous said...

"first a development FAB works out the recipe, then the production FAB copies exactly, including the size of Windows."

This is not true - this is only true for the initial process technology transfers to the fab. Think about what you stated - Intel's development facility is now working on 45nm development. When new products come out on a 90nm or 65nm process any recipe development for that product is done by the fab producing that will be product. Do you honestly think Intel's development fab does recipes for new 90nm chipsets? other 65nm chips? 0.18um embedded products?

2:00 PM, August 02, 2006  
Blogger Sharikou, Ph. D said...

Show me any link or article on this and I will apologize.

1) Find the IDF pumper sites

2) Note the 40% performance claims

3) Note that the benchmarks were pre-arranged by Intel

4) Note that Intel later quoted the paid pumpers to Wall Street

Conclusion: Intel = fraud

2:23 PM, August 02, 2006  
Anonymous Anonymous said...

"1) Find the IDF pumper sites

2) Note the 40% performance claims"

You have continually stated Intel lied, not just the so-called Intel paid pumper-sites.

I maintain that Intel has only stated 40% better than previous Intel product and you provide absolutely no evidence that INTEL has said anything to the contrary.

I'll keep my challenge open for you to find a single article where Intel (Intel, not THG, not anand, etc...) stated that Conroe would be 40% better than FX-62.

I'm awaiting another one of your non-answers "it was on the web somewhere"! Or the paid-pumper sites are the same as Intel....blah, blah, blah... I can't back up my accusations with actual fact so I'll just keep making stuff...

11:36 PM, August 02, 2006  
Blogger Mad Mod Mike said...

"I'm awaiting another one of your non-answers "it was on the web somewhere"! Or the paid-pumper sites are the same as Intel....blah, blah, blah... I can't back up my accusations with actual fact so I'll just keep making stuff..."

So at IDF when Intel published false benchmarks of a 2.67 Conroe beating a "FX-62", did they just sit around with a George Bush look on their face and go "OH hey, look at that!"? Are you really that stupid? Please....

9:18 AM, August 03, 2006  
Anonymous Anonymous said...

So when they said it beats an FX-62, which I believe based on the data out right now it does, you naturally assume beats = 40% better?

What are the false benchmarks? Are you telling everyone reading your post that Intel lied when they said a 6700 beats an FX62? Did they say it beats it by 40% at IDF or just that it beats it? (If 40%, please post your refernce link or article)

Yet another good non-answer! The 40% metric INTEL has stated was always in reference to past Intel product - a point which you have been unable to refute!

At least Sharikou posts the counter points of view on his blog. He may or may not agree or listen (which is completely his perogative )but at least he posts them and lets readers decide for themselves - a practice I've noticed you've stopped doing on your blog MMM...

11:57 PM, August 03, 2006  
Blogger Mad Mod Mike said...

"At least Sharikou posts the counter points of view on his blog. He may or may not agree or listen (which is completely his perogative )but at least he posts them and lets readers decide for themselves - a practice I've noticed you've stopped doing on your blog MMM..."

I don't post irrelevant, idiotic posts made by morons that don't substantiate themselves, kinda like you.

7:49 AM, August 05, 2006  

Post a Comment

Links to this post:

Create a Link

<< Home