You're right. I'm sorry for being confused between Presler and Smithfield. I guess I was a little carried away with the cost and yield arguments.
I've been advocating AMD to have a better top range product, because by that, they get to set the price ladder, as opposed to responding it. Back in the old days, an X2 3800 was selling at 300USD, and its not hard to imagine the margin AMD gained from that. People were shelling out money to purchase X2s, despite significantly higher cost, when Pentium Ds were forced to go onto the price ladder in order to be competitive.
Even with their lower prices, the Pentium Ds were not better performance-for-the-dollar chips vs. the X2s, possibly except the 805, which was just by its virtue of being less than 40% as expensive as the slowest X2 was. I bought my computer during this time period and paid $360 for an X2 4200+, while the Pentium D 930 was a scant $30 less expensive but far slower. The X2 3800+ was about $300, but it competed with the Pentium D 940 the closest and the D 940 was over $400. So I wouldn't say that AMD specifically was overcharging people for the performance their chips gave with the X2 3800+ as *everybody* was charging that much.
The current Pentium D generation was never price/performance competitive with the X2s when it was being sold. The closest the Pentium Ds came to being price/performance competitive was when the old 820 and the newly-crippled 820 called the 805 were brought out at low prices when the 900 series was selling for high prices. These chips were not even competitive with the lowest of the low end of the X2 line but their low price got them a lot of business.
However, now the roles have changed. AMD now is being pushed to the low / value end, where margin is virtually nonexistent. With monolithic quad core like Barcelona, the cost of that is significantly higher than Intel's MCM part, with potentially lower yield, and lower performance.
I'll agree with the higher cost and lower yield, but the performance bit is up for grabs and yet to be determined on the desktop. The server Barcelonas haven't been benched a whole lot outside of SPEC benches and the results are all over the map. Sometimes they aren't quite clock-for-clock with the Xeons (int) and sometimes they take it all (rate_fp.) We'll get a better idea when the parts ship.
I agree. I really wonder if monolithic die is a better approach than MCM.
There are advantages and disadvantages. The monolithic core allows for easier faster core-to-core communication than an MCM does, especially as the number of dies goes up. A monolithic core also makes a typical integrated memory controller easier to implement than if there are two cores. Lastly, a monolithic core allows for finer-grained power control. MCMs' advantages are pretty much limited to the fact that they are simply two existing dies placed on one substrate. They are less expensive due to higher yields and can be rolled out quicker and easier than drawing up a new dis mask. So basically a monolithic multi-core chip is technically superior but the MCM one has the cost advantage.
It definitely takes a lot more time to come up with a monolithic CPU than MCM. However, as you said, it takes little to no effort to glue two cores together, and brand them as quad cores. My argument was that AMD should have done something about Intel's quad core.
I agree. AMD had technical reasons why they would not to reasonably do this. They had a few approaches to make an MCM available to them:
1. Put two Opteron 22xx dies in a package and slave Die 1 off Die 1's IMC via HT 2.x.
2. Put two Opteron 22xx dies in a package and wire one RAM channel to each chip's IMC and then use HT 2.x for inter-die communication.
3. Redesign the Opteron 22xx die to use HT 3.0 and slave Die 1 off Die 0's IMC.
4. Redesign the Opteron 22xx die to use HT 3.0 and wire one RAM channel to each die's IMC and use HT 3.0 for inter-die communication.
5. Make a new socket to handle four DDR2 RAM channels and put two Opteron 22xx dies in it, wiring a full two channels of RAM to each IMC and using HT 2.x for inter-die communication.
Option 1 would be absolutely terrible as cores 2 and 3 on die 1 would be terribly starved for RAM bandwidth. Gigabyte had an dual-socket Opteron setup like this in the past and it was awful.
Option 2 would work okay, but now you are using NUMA, which Windows doesn't like much (look at the QuadFX vs. X2 6000+ benches.) Much of why Windows hates it is because remote RAM accesses over HT 2.x are not all that stellar. It would also pretty much require memory interleaving, so installing RAM in pairs is mandatory.
Option 3 would be better than Option 1 as HT 3.0 is much faster than HT 2.x and you are using an UMA setup like a monolithic quad or an Intel FSB setup uses. I don't know how it would compare to any other options, though. However, this would require a redesign of the core, something that AMD wants to do while also putting a new architecture in place. That takes extra time.
Option 4 is pretty much what's planned for the future 8-core MCMs AMD has planned. It will need to use NUMA, so WIndows performance won't be stellar, but HT 3.0 should mitigate that some. Not a very good current option as it requires a core redesign. It should also pretty much require memory bank interleaving.
Option 5 would be basically a full dual-Opteron-in-a-single-socket setup. Bandwidth ought to be good, but four RAM channels is 960 data pins just for RAM, so you're looking at 1500 pins or so for the socket. This isn't a good option as it requires a new socket (I can hear the groans) and uses NUMA. Also, you'd want to install RAM in sets of four, which can be a pain in the butt.
The lack of response from the green team led me to believe that they pretty much put the fate of the company on Barcelona.
I don't see AMD's fate resting any more or less on 10h than Intel's is resting on Core. AMD sells chipsets for Intel processors as well as a graphics cards for a bunch of different makers' units. So if the 10h tanks, yeah, AMD will take it on the chin, but they will still get revenue from somewhere. Intel pretty much just sells Core-based CPUs, chipsets, and motherboards that only work with Core-based CPUs. They make a few NICs and disk controllers that will work on other platforms, as well as a few external storage devices. AMD has far less of their revenue depending on the success of their CPUs than Intel does.
When Barcelona flopped (as of now, it might get better later), AMD is left with nothing but continuing price cuts. This is not helping them. IMO, AMD's 65nm is not mature enough to tackle monolithic approach. If AMD really yield quad cores at 30%~40%, and can't get their clockspeed up, its pretty much unacceptable.
We haven't seen much of Barcelona, so we can't say if it's a flop, how up to the task the 65 nm architecture is, or what their yields are. Time will tell that one.
Thanks for the information. I'm curious though, that did Intel have a parallel research effort when they knew AMD was working on x86_64, before they dumped the project and sued AMD?
I couldn't find really anything about that. AFAIK the only 64-bit Intel was working on before AMD's x86_64 came out was Itanium IA64.
I would respectfully disagree with you on this one. People install 4Gb of RAM, but how many of them actually utilize that many stick? I do realize that it has increasingly become a concern, but for most people, 32bit is sufficient.
Most people don't do much intensive and a five-year-old machine, properly cleaned up, will do just fine. I know people who just upgraded from Pentium MMX 200s and PII-350s, largely because they had hardware break on the old machines and replacements are hard to get today. It does not take a QX8650 and 8 GB RAM just to type a simple document and use a Web browser. We had graphical word processors and Web browsers pretty darn near identical in as what's new today running a dozen years ago on machines with 32 MB RAM and 486s and Pentiums. There is no reason that we should need processors that are 100 times faster and 100 times more RAM to do essentially the same tasks. <old-school joke>What, did the Emacs guys write all of the software out there today or something?</old-school joke>
Only people that run memory intensive applications, such as video editing, need that much RAM. Even for gamers, I believe 4Gb is a little overkill. As for AMD64, I believe its a little early for its time. It was released back in 2003, and we still have yet to see a massive X64 user base.
Eh, some games are getting close to exceeding the 2 GB/process limit that 32-bit OSes commonly have. You'll need a 64-bit OS to get around that and have things be stable- yes, I know you can adjust the kernel split to 3 GB/1 GB, but that sometimes isn't stable.
Someone mentioned that in SC he/she will crash to desktop if 2Gb of RAM were installed. I encountered the same problem in Vista, but everything ran fine in XP. I'm leaning more towards the inefficient coding in Vista that caused this problem. For most, 2Gb is sufficient.
2 GB is sufficient for general usage. If you have a new game that needs 2 GB of its own RAM, you'll want at least 3 GB or more. Every day more people are hitting the 32-bit memory limit and need to get around it, which means a 64-bit OS. You're correct in stating that most haven't hit it yet, but a combination of ever-more-detailed games, HD video, higher-resolution photographs, but mostly the ever-increasing software bloat (look at Vista) pushes people over it every day at an increasing rate. I'll bet most people will be on a 64-bit OS by the time the next Windows ships.
Facebook
Twitter
Instagram