What is PH3? The Phenom 2's with DDR3 support have a release date in February. IDontCare was saying that if there is going to be a Phenom 3, it would likely be the Bulldozer architecture in 2011.
VirtualLarry
No Lifer
- Aug 25, 2001
- 56,587
- 10,227
- 126
Do you have any actual data to support that conclusion, or are you just talking out your ass becauase you like the Core i7.
Because IDC posted some very interesting multiprocessor scaling benchmarks, and the Core i7, frankly, sucks at MP.
IDC, correct me if I have the wrong impression of your results.
- Sep 28, 2005
- 21,131
- 3,665
- 126
while you guys are arguing, Im hearing 32nm intel samples are starting to distribute.
Catchup game? Amd will be playing catchup for a while now.
The company themself gave up trying to go for the crown, dont try to inflat the company please.
Is AMD bad? Definitely no, the PH2 looks like a great chip. Most likely most of you wont even use it to its full potential.
Server sector? Sorry, a Neha-EP with 2 processors HT 16 threads will gobble anything AMD has to offer on the server side for a while..
i can answer this...
When testing, and trust me a lot of testing went on, we racked 2 systems. A QX9650 and a i965.
We clocked both systems at the same clock. The QX9650 was indeed slightly faster when looking at single or double threaded programs.
When we compared it on a full scale multi core program, the i965 almost did 2x the work the QX9650 did.
The real numbers was about 1.45-1.5x more work. So nemisis's statement isnt off.
Catchup game? Amd will be playing catchup for a while now.
The company themself gave up trying to go for the crown, dont try to inflat the company please.
Is AMD bad? Definitely no, the PH2 looks like a great chip. Most likely most of you wont even use it to its full potential.
Server sector? Sorry, a Neha-EP with 2 processors HT 16 threads will gobble anything AMD has to offer on the server side for a while..
i can answer this...
When testing, and trust me a lot of testing went on, we racked 2 systems. A QX9650 and a i965.
We clocked both systems at the same clock. The QX9650 was indeed slightly faster when looking at single or double threaded programs.
When we compared it on a full scale multi core program, the i965 almost did 2x the work the QX9650 did.
The real numbers was about 1.45-1.5x more work. So nemisis's statement isnt off.
VirtualLarry, yeah the scaling was rather abysmal for two benchmarks that have been published comparing 2P shanghai vs. 2P barcelona vs. 2P harpertown vs 1P Nehalem.
Euler3DBenchmarkScaling.gif
MyriMatchBenchmarkScaling.gif
However scaling isn't the whole story as it doesn't speak to absolute/raw performance. In the sheer performance evaluation Nehalem decimates the competition (even these 2P rigs) in these two benchmarks. Also don't forget these two benches are nuimber-crunching benches and not your typical server programs and benches.
Nevertheless they did provide an interesting view into the scaling of 1-8 threads on Nehalem, as the scaling was pretty lackluster (even for 1-4 threads w/o HT) compared to the other chips. Only time and more review data will determine whether something was screwy with these initial benches or if something wasn't configured correctly in the memory sub-system on the i7 rig that was benched.
But there can be no doubt that a 2P nehalem xeon rig is just going to cream the majority of mid-iron servers currently selling for $30k-$40k. The raw performance is there for it to happen. (I know nothing, just extrapolating benches on a backdrop of personal expectation)
Phynaz, no doubt it creams a niagaraII system but is that what you are alluding to or are you perhaps hinting at some power6 carnage? The spec Cpu2006 results hinted at so far would suggest 2P nehalem will threaten absolutely everything that isn't 8P or more including itanium. Are the 32nm samples x86? They can't be poulson already, can they?
- Sep 28, 2005
- 21,131
- 3,665
- 126
why do you think im using neha-ep on a i7 platform?
Remember when i told ya i would can' my i7 entirely? and go straight to gainestown.
*kicking the can*
Well, why do you think im brining it back?
Take this one at a time...
"On workstation and desktop apps deneb will be comparable to penryn" - you are making the assumption that Deneb vs Shanghai perform the same on desktop apps.
Regardless, the Techreport review didn't do the usual desktop benchmarks...they rightly focused more on server and workstation benchmarks.
In addition, they didn't do the benchmark series that Shanghai has been designed to focus on (because of the high demand)...namely VM benches. They stayed mainly with Nehalem's strength which is rendering and encoding.
"It will get crushed on server workloads by much cheaper nehalems" - since Nehalems will be the same size, what makes you think they will be cheaper?? I would bet that the opposite is true...
Also (as I said above), it depends on what you're doing with the server. The word is that the Opterons will be significantly better for servers running Virtual Machines, which is a very large portion of today's needs.
"their loss for Q4 seems to be even bigger than Intel's" - Don't think so...Again, you need to look at the 2 warnings in perspective. Intel made their warning in Oct, and AMD in Dec. Since Intel's estimate is so old, it's accuracy is quite doubtful considering what's been happening to the economy.
Markfw
Moderator Emeritus, Elite Member
- May 16, 2002
- 27,403
- 16,253
- 136
Well, speaking for my company, we have 15,000 Wintel servers, that we are trying to reduce to 1/4th that using virtual machines, and I am not sure the unix count, be we are trying to "compact" those also, using bigger more effcient boxes with more lpars.
Cookie Monster
Diamond Member
- May 7, 2005
- 5,161
- 32
- 86
Report: 3GHz Phenom II to launch at $275
However I dont know how credible EE Times, so take it with a grain of salt.
However I dont know how credible EE Times, so take it with a grain of salt.
If you mean who is saying that the Opteron will perform better? I agree that we still have no reliable benches of the 2 compared...
However, the latest VMMark benches give a hint to how far the Shanghai has moved forward...
http://www.vmware.com/products/vmmark/results.html
Notice that AMD Shanghai is well atop both the 8 and 16 core categories...
It is not assumption. It is fact, it's the same architecture, only bandwidth numbers and some uncore clocks will change from Shanghai -> Deneb. Rumour has it that uncore will be clocked lower btw.
Techreport used some benchmarks that will make their way into desktop testing and I've seen lots of unofficial benchmarks (leaks, previews, reviews, etc) each and every one hinted in the same direction.
Then there is common sense, no one manages huge gains with a modified die shrink. Don't expect miracles.
It's just a guessing game whose costs per die will be bigger. They are more or less the same size. Some of the differences include (EDIT: list updated to reflect some more ideas):
Nehalem: More logic (-), double patterning (-), mature process (+), tool reuse(+), bulk silicon (+), tricores? (+-) cache redundancy? (+-)
Deneb: process (-), immersion expensive AFAIK (-), expensive SOI (-) less logic (+), tricores possible (+)
Going by the official SPEC numbers and official techreport review we can be sure that Nehalem will be up to 50% faster in certain applications, I just extrapolated what this could mean. I don't believe similar performing Nehalems will be more expensive than Shanghai, at least as a new platform, upgrades are a different ballgame.
And if Idontcare is right and upgrades become more important throughout 2009, this could help AMD.
Regarding virtualisation, we only know that Shanghai performs great, but we don't have any numbers from Intel. IIRC nehalem also sports some virtualisation improvements.
I've already considered this in one of my posts and am well aware of the fact that Intel could lower their expectations once again.
On the other hand AMD sells a lot of high-end graphics cards which only recently started to be profitable, so I think it is to be expected that they'd lose more than Intel.
EDIT: updated to reflect a more complete, albeit still limited, list
It's the same core architecture, but not the same chip design...(less logic circuitry, different cache, no RVI, etc... on Deneb). Also, since scaling is so efficient on K10, the clock changes will be quite significant to performance.
The only desktop bench I saw at TechReport was the encoding one...what else are you talking about?
There are certainly many other differences than a die shrink...remember that AMD has stated that there will be an up to 20% improvement even before the faster clockspeed.
I can only say that "hints" and "leaks" don't say (to me at least) that AMD will be "crushed" in any way at all...
But it isn't a guessing game as to die size...they are the ~same, depending on which configuration you're talking about (remember that cache can be VERY big).
BTW, Nehalem isn't a mature process at all! Whenever you have a radically new design (like Nehalem), the chances for error increase dramatically. That's the reason for the long waiting period on Nehalem servers...it takes much longer for a completely new design to be qualified. Obviously you have a statistical increase in rejected parts, so the early Nehalems will probably be much lower yield than those a year from now.
As for virtualization, the Nehalem will (supposedly) finally support RVI, however they will not include Nested Pages I believe...
Ummm...those graphic cards started being profitable a year ago when the 3850 was first released...
The Nehalem core size is 24.4mm^2, while Shanghai is only 15.3mm^2 link, so the Shanghai yield should be greater (errors in cache are less likely to prduce a "bad" chip). Plus AMD's manufacturing costs should be much lower than Intel's, because they are using Immersion Lithography, and Intel is using twice as many passes. So Shanghai would be cheaper to produce than Nehalem in every way that I can see (Higher yield, cheaper manufacturing process.) I really don't see where you are getting that Nehalem is cheaper to produce.
[/quote][/quote]
Thanks Martimus...I knew in the back of my mind that Shanghai was going to be cheaper, but I had forgotten about the double passes on Nehalem. Shanghai should be significantly cheaper to produce...
Are you proposing Deneb is a different die than Shanghai? AMD actually runs two mask sets to produce these otherwise remarkebly similiar chips?
I always assumed it was one chip (one mask set) with features fused off at packaging to determine the ultimate SKU based on market demand and parametric yields for the varying features.
You are talking about architecture design while jackyp is talking process technology.
Nehalem is introduced on an existing mature 45nm process tech. Shanghai is not.
Nehalem is a new architecture, Shanghai is not considered to be a new architecture in the traditional sense that this phrase is meant to convey.
Guys easy on the waaaay oversimplification of cost structure here.
If AMD had Intel's volume scale to assist in reducing manufacturing costs then we'd be a skosh closer to being able to compare line-item process technology items in a pro vs. con fashion.
As it stands now we can't in good faith (or intelligence) argue that even Intel would have cheaper 45nm parts were they to implement immersion litho to replace their existing (and cost/performance optimized) double-pattern process.
We have no way of knowing how optimized and low-cost Intel's double-pattern is relative to the higher equipment cost approach of implementing immersion litho. Conversely we have no way of knowing just how efficiently AMD is implementing their immersion litho.
Both sides are going to do their best to give you (the potential shareholder) the impression that they have a management team delivering superior decisions relative to the competition.
No. As I said...same die, different chip. But to suggest that the different features don't make a difference in performance I think is incorrect...
With that said, I really don't know the answer (I am assuming the same way you are as it makes sense).
I actually did know what he meant...
I was pointing out (obtusely, of course) that while Intel has been using 45nm for awhile now, it doesn't mean they will garner mature yields on a radicall new architectual design.
Caution noted and well said...
However, the lions share of the costs for immersion should already have been paid for by AMD's Capex budget over the last year. I can't imagine that they spent all that money on new equipment if they could be anywhere close to as efficient using double patterning...
It is true that there can never be certainty in having mature yields translate from one product to another at the same process node, but you usually take confidence in knowing that problems encountered and fixed in the past will translate to fewer problems to uncover as you move forward with new designs on the same process tech.
I.e. of the things that can and will go wrong in the remaining lifetime of Intel's 45nm process tech, there are fewer things in that bucket now versus one year ago when it was freshly introduced.
The same can't be said of AMD's bucket, it is full and brimming with gotcha's and murphy's law items just waiting to be discovered and uncovered as volume ramp is attempted.
If you mean to compare AMD's existing 45nm cost structure with immersion to the cost structure of a hypothetical version of their 45nm node (AMD's that is) which uses an optimized double-pattern process then I agree you are justified to spend the energy contemplating those parallel universes.
Like many things in the industry, AMD migrating to immersion at 45nm (albeit introducing it a year later than Intel's 45nm double pattern and a year earlier than Intel's 32nm immersion) represents risk management and investment in lowering the cost of developing future tech.
The value of AMD going with immersion won't be evident until we see how quickly they transition to 32nm. The other value metric would be an actual cost structure analysis, which we (the common shareholder) will never get our hands on.
The non-common shareholder (mutual fund companies, big investors, etc) will get to see the cost structure numbers but hopefully people will know to assume they represent a silly amount of the best case scenario as management MUST justify their compensation packages to these shareholders and going immersion was their decision so it HAD to be the best one (to justify those bonuses). I've been in those meetings before. Caveat emptor.
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads
- Started by Tigerick
- Replies: 24K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
-
Facebook
Twitter