Sometimes that isn't the most important metric in a server. For example, a VM server handling many mostly-idle virtual servers. The number of servers may mandate a certain number of cores, but if those cores are idle 95% of the time the impact of having higher power usage for 5% of the time can be less relevant than you would otherwise think.
This.
Rarely does a chip come onto the market where it roundly outperforms the competition but still manages to fall flat on its face in terms of marketshare.
But history is littered with the corpses of dead microarchitectures that came to the market needing a song and a prayer (new code, better compilers, etc) to have a chance at succeeding.
Why AMD decided to bring a "song and a prayer" type architecture to the market when they needed an "I can do ALL that and bring my own bag of chips!" type microarchitecture (i.e. a Conroe in its day) is beyond explanation.
What they needed is obvious, what they delivered is obvious, all this verbosity to explain the gap between the two is needless.
Woosh!
The point is demonstrating that the Sandy Bridge architecture delivers much higher performance/watt. That fact extends into Sandy Bridge-E and Interlagos, both of which use the Sandy Bridge and Bulldozer architecture, respectively.
/yawn
And yet it remains that none of your example parts are used in servers. The tests were done on Win7 Ultimate. Also, not a server OS. I don't know anyone that would take a data set from one thing and say it's more or less the same to something almost entirely different. You actually have test and prove what you are saying, which is not what you have done. You have shown what the desktop parts do and nothing more.
/yawn
And yet it remains that none of your example parts are used in servers. The tests were done on Win7 Ultimate. Also, not a server OS. I don't know anyone that would take a data set from one thing and say it's more or less the same to something almost entirely different. You actually have test and prove what you are saying, which is not what you have done. You have shown what the desktop parts do and nothing more.
You must be new here.You actually have test and prove what you are saying, which is not what you have done.
It is equally telling that you do not, can not, have a data set to disprove their assertions either.
Interlagos is MIA, no SKU specs, no performance results. Nothing.
Isn't the burden of proof on the person trumpeting that 'Bulldozer is teh sux on server!!!1111one!11!!'? I personally will wait to see server benchmarks done on the Interlagos SKU, if it's ever released. There are some server scenarios where Interlagos could put up a fight.
Isn't the burden of proof on the person trumpeting that 'Bulldozer is teh sux on server!!!1111one!11!!'? I personally will wait to see server benchmarks done on the Interlagos SKU, if it's ever released. There are some server scenarios where Interlagos could put up a fight.
You ran server load benchmarks on Interlagos ES?Because server workloads are multi-threaded, and we already know Bulldozer's multi-threaded performance. For servers the picture won't change at all.
You ran server load benchmarks on Interlagos ES?The same thing will play out as when we compared 4 core/4 thread and 4 core/8 thread Sandy Bridge to 8 core/8 thread Bulldozer, except now it's with double the cores for both plus some minus changes. They'll also be clocked in a similar low-to-mid 2GHz range.
There are patches going into the Linux 3.2 kernel with Bulldozer optimizations. or was it 3.1? Don't remember.I don't see how you're magically expecting this to change; it won't. It didn't for Magny Cours, it won't for Interlagos.
I guess you never developed for servers nor have any server knowledge/experience. You do realize that not everyone is doing FP stuff server side right? For example, we got this server farm at work that acts as proxy for mainframe data. All we need is a metric boatload of slow cores to handle TCP connections from one end to another while doing minor data conversions. For our app, speed doesn't matter, we process a LOT of parallel connections with very little CPU processing needed. This could be any [insert high core count cheap server here]. A server optimized Bulldozer could be perfect for the job if the price is right.Another thing to take into account is that a sixteen core Interlagos will actually have eight floating point units, and FP performance is more important for server and workstation workloads. Again, doesn't look any better in that scenario.
Remember people, it is not that Bulldozer is a bad chip. Instead, it is
A) GlobalFoundries's fault for producing a 2 billion transistor chip that uses more than 35 watts of power
B) GlobalFoundries's fault for only being able to get launch day chips on an immature process to 3.6GHz.
C) Microsoft's fault because they purposefully made Windows 7 reduce bulldozer performance by 10% or something
D) Microsoft's fault for allowing non-AVX software to run on Windows 7 SP 1 since this software gives bulldozer a 10~% boost relative to Sandy Bridge, and this is Microsoft's fault.
E) Asus's fault for forcing AMD to use their faulty motherboards in the launch day review kits that unfairly slowed AMD processors by 25% or so.
F) Intel's fault for bribing every software maker on the planet to purposefully make software that is non-AVX because this hurts AMD.
G) Most reviewers' fault for testing the processor's power usage.
H) Most reviewers' fault for actually testing the processor in situations it will be used in instead of AMD approved AVX friendly linux distros with only avx software.
I) Intel's fault for hiring people to post on tech forums to baselessly smear bulldozer's performance
J) Kill-o-Watt's fault for reporting bulldozer's power usage.
Anyone that disagrees with this is an Intel employee shilling for their employer
I'll never understood the scorn about that. Intel tested it internally, found it slow, canned the project. People getting worked up about an unreleased product boggles the mind. Had it been released and found lacking, ok, scorn warranted.With that logic, larrabee never deserved the scorn that was heaped upon it either, given that it was never tested, being the vaporware that it was.
Exactly. Bulldozer is going to have the same issues as Niagara (unless AMD can get B3 stepping out with good Linux kernel optimizations in time, and even then ...). Basically good in some niche markets, limited scenarios, unfortunately.I'm willing to read between the lines. Interlagos will do well in those same workloads where Niagara gives Power and IPF a run for the money. What is Niagara's marketshare?
Isn't the burden of proof on the person trumpeting that 'Bulldozer is teh sux on server!
You ran server load benchmarks on Interlagos ES?
You ran server load benchmarks on Interlagos ES?
There are patches going into the Linux 3.2 kernel with Bulldozer optimizations. or was it 3.1? Don't remember.
I guess you never developed for servers nor have any server knowledge/experience. You do realize that not everyone is doing FP stuff server side right? For example, we got this server farm at work that acts as proxy for mainframe data. All we need is a metric boatload of slow cores to handle TCP connections from one end to another while doing minor data conversions. For our app, speed doesn't matter, we process a LOT of parallel connections with very little CPU processing needed. This could be any [insert high core count cheap server here]. A server optimized Bulldozer could be perfect for the job if the price is right.
You make it sound very easy for a company with not much market share to ask others to implement new ideas new ways of doing things. If Microsoft can't get traction for WP7 with all their muscle, money paid to developers and ad dollars, you think a little company with much less cash can? MS supposedly blew 500+ million on just to advertise WP7. How many more millions went to partners, developers and telcos? Result? Not much! And you think AMD can do better?
Gaming companies don't care, they mostly port console games to PCs and not to mention that the majority of the gaming is done 1080p or below so no powerful CPU is required. Office software, browsers and other software used by the common mortal doesn't need anything more than a little dual core. So for consumer market, there isn't much AMD could have done, other than maybe beg MS to implement a better scheduler in Windows (maybe they did and MS refused, who knows).
For servers, it's back again to the traction problem. On the server side they have what? 5% market share. Big boys either don't care to modify their own programs for a possible gain even if AMD pays every single dime for it (which they can't afford). If I were a CEO of a big company, and AMD would pay for me to rewrite parts of my code to run better on Bulldozer, I would flat out refuse it. I would have to possible maintain two code bases, support more systems, train more of my own employees and whatnot (read: raise costs, lower margins). All this in the hope that maybe, MAYBE Bulldozer will conquer the world. But when I look at the competition, there is no chance for that, so I would refuse AMD's offer.
It's easy to dismiss AMD's problems and point fingers to management, Bulldozer and whatnot, but lot of people forget that AMD has no resources to compete with Intel. The only reason AMD still exists today is to keep Intel out of anti-trust/monopoly problems. It's Apple vs MS in the late 90s over again. The only difference is that AMD doesn't have its own Steve Jobs and it's going to be kept alive as a zombie to keep regulators out of Intel's profits.
Sometimes that isn't the most important metric in a server. For example, a VM server handling many mostly-idle virtual servers. The number of servers may mandate a certain number of cores, but if those cores are idle 95% of the time the impact of having higher power usage for 5% of the time can be less relevant than you would otherwise think.
Also, just a guess, the server version of bulldozer may be clocked lower and use significantly less power than the consumer CPU's we are familiar with.
LOL, Bulldozer apologists.
Yes, FP performance is largely irrelevant in desktop workloads, but definitely something to take into account in servers/workstations. Integer performance is still more important, but less FPU resources won't do it any favors.
Last two PC's I've build have been Intel.
My point is that it's often worth looking at things from a wider perspective and not limit ones opinion to "OLOLOLOL FAIL". The value of Bulldozer architecture will only be apparent once it has matured - like with Phenom. Agena definitely wasn't a good start, as isn't FX, yet Deneb made the Phenom the a very, very popular CPU series, which is why I'm excited about Pilediver.
The 1st plane flew only a couple of meters - EPIC FAIL, right?
Thus my zombie comment. AMD is kept artificially alive. They just don't have the resources to compete, never had. Athlon 64 was an Intel error. Catching Intel performance wise was only possible during a short time when Intel was managed by marketing (Gigahertz uber alles folk!).You just reinforced my argument for me, thanks. If AMD cannot change the industry, then why release a product that is so poor at existing software? It's not what a product 'can' do 2 years from now, it is what it can do NOW. The industry doesn't stand still, and they don't wait for anyone either.
Apples to oranges. NVidia was in a market that was pretty open. There wasn't a big competitor that utterly dominated and had a 2+ year advantage in manufacturing with 10 times the monetary resources. So AMD had no choice. Instead of trying something new like Bulldozer, they could have done what? Die shrinked Thuban/Deneb and try to catch up with Intel? Good luck. There is virtually no reason for someone to get an AMD processor on desktop unless some rare cases where they are limited by budgets. That would not have changed with a die shrink of Thuban either. They had to try something new and get traction with more cores. Did/will it work, who knows.The goal is to succeed with current needs and offer a 'niche' that can gain traction. Nvidia did this great with CUDA. Keep making good (performance-wise) GPUs that have a feature that people can adopt and use. Sure, that may lead you way off in a different direction years down the road, but it allows you to succeed in the current market AND be forward-thinking. That's smart business.
LOL, people who pull non existent Interlagos benchmarks out of their own asses. See what I did there?
Just shows how you got no clue about servers and server software. Again, not all server task is FPU ... I coded stuff that runs on 10+ servers in a farm, none has a fast CPU, we just need cores. The software receives a request from the web server farm, calls the mainframe, waits until mainframe did his job, does minor transformation on the data, sends it to web server. No FPU use at all.
So if you completely take away the FPU, good, less power consumption. For us having a lot of cores is golden because we run a metric boatload of threads who consume virtually no CPU power, but if we pile up too many of those threads on a low number core CPU, the CPU keeps context switching from thread to thread instead of processing requests and then ... hilarity ensues. You really think HP is venturing into ARM servers for shits and giggles? There are some workloads that just require cores, no high IPC, no kickass FPU, just cores.
Understand? Need me to draw a picture?
You have absolutely no clue of what you're talking about. We already know how the architecture performs on multi-threaded workloads, so a very accurate conclusion can be drawn from that. Same as SB and SB-E. It having a "server" denomination doesn't change the performance of the architecture. If you think it does, I have a golden bridge to sell you for $100. Bulldozer isn't made of magic pixie dust.
If you think FP performance in servers doesn't matter you're clueless about that, too. And a full core consists of both integer and floating point units. Like I said before, on integer workloads BD is an eight-core; on floating point it's a quad-core. Floating point units, like integer units, are also execution engines, which is why AMD is saying eight integer core Bulldozer is an "eight core".