Dell puts AMD opteron in server

[nyker96]

WHat do you guys think? Even in light of recent C2D roll out, seems AMD is mking some headwaves in server market.

hxxp://www.dailytech.com/article.aspx?newsid=4667

 

[DrMrLordX]

First off, you might want to correct the many spelling errors in your post. Try changing "pus" to "puts", and also fix that link so it doesn't read "hxxp". Please. I'm not trying to be a nitpicker here, but those two errors are problematic.

Secondly, this kinda makes sense, especially for 4p servers and up. In 1p and 2p servers, I'd rather use Woodcrests to be honest.

 

[Zap]

Well, they're just playing catchup to the other companies that have had Opteron servers for a while already. Opteron is still a great choice in 4P+ and versus any Netburst Xeon. The 2P Woodcrest platforms are a different beast though.

 

[nyker96]

Lorx, I corrected the pus. the hxxp is due to the fact that I don't want to direct link things. Some forum has such limitations, just to be safe I always use hxxp as recommanded by several forums, easy enough for you to find the link from there.

I also got the same impression reading this article, 4P+ AMD still seem to have an edge on performance. This will allow Dell to better compete with company like HP which has been offering AMD opteron for a while.

 

[dexvx]

Originally posted by: Zap
Well, they're just playing catchup to the other companies that have had Opteron servers for a while already. Opteron is still a great choice in 4P+ and versus any Netburst Xeon. The 2P Woodcrest platforms are a different beast though.

Tell me the difference between:

4x DC Opterons and
2x QC Xeons

 

[Aluvus]

Originally posted by: dexvx
Tell me the difference between:

4x DC Opterons and
2x QC Xeons

You can buy 4 dual core Opterons.

 

[customcoms]

Quad core woodcrests (essentially Kentsfield for Server) is a performance animal; nothing AMD has can compete, period.

 

[The-Noid]

Cloverton is the name. Clock for clock they are faster. Possibly 4x DC Opterons provides more performance/watt though which is more important in the server market. Dell has seemed to make mistakes on everything lately, I wouldn't doubt that now that they take AMD, AMD starts doing worse then Intel after it being the other way for years. Seems that everything Dell touches goes to pot.

 

[nyker96]

I have a feeling AMD's quade will be fairly good for the server market in terms of power. So it will be competetive even if per core performance will be worse.

 

[Regs]

To be honest I invested in AMD 2 years ago. I have no reason to sell my stocks now.

 

[myocardia]

Originally posted by: Aluvus

Originally posted by: dexvx
Tell me the difference between:

4x DC Opterons and
2x QC Xeons

You can buy 4 dual core Opterons.

Exactly. Although, I have the feeling that within a year of Kentsfield/Cloverton being released, AMD will start losing back market share in the server arena. It serves them right, to tell you the truth. They said a good while back that they were going to be neglecting the desktop arena, so they could spend their time/money on the server and mobile arenas.

edit: Link to the article that the OP didn't link.

 

[Viditor]

Originally posted by: myocardia

Originally posted by: Aluvus

Originally posted by: dexvx
Tell me the difference between:

4x DC Opterons and
2x QC Xeons

You can buy 4 dual core Opterons.

Exactly. Although, I have the feeling that within a year of Kentsfield/Cloverton being released, AMD will start losing back market share in the server arena. It serves them right, to tell you the truth. They said a good while back that they were going to be neglecting the desktop arena, so they could spend their time/money on the server and mobile arenas.

edit: Link to the article that the OP didn't link.

I disagree about Cloverton...
1. Expecting 2 Clovertons on a single FSB to perform on par with 1 Kentsfield (especially in a server environment) is a mistake. Remember that you are greatly increasing the traffic, especially as half of the cache coherency must use the FSB as well.
2. A dual Cloverton platform is not upgradable, whereas a 4P Opteron platform can not only be upgraded to 16 cores, it can also be upgraded to the K8L (Rev B) cores at the same time, which will arguably (even with an equal number of cores) be faster than the Clovertons.

 

[myocardia]

Originally posted by: Viditor
I disagree about Cloverton...
1. Expecting 2 Clovertons on a single FSB to perform on par with 1 Kentsfield (especially in a server environment) is a mistake. Remember that you are greatly increasing the traffic, especially as half of the cache coherency must use the FSB as well.
2. A dual Cloverton platform is not upgradable, whereas a 4P Opteron platform can not only be upgraded to 16 cores, it can also be upgraded to the K8L (Rev B) cores at the same time, which will arguably (even with an equal number of cores) be faster than the Clovertons.

You don't really think that SuperMicro/Tyan are going to be producing dual Cloverton motherboards with a single fsb, do you? I can assure you, they have more sense than that. And since K8L is still roughly 12 months away, neither of us knows whether or not a 4P Opteron board that's bought today will support K8L's at all.

 

[Furen]

Originally posted by: myocardia
You don't really think that SuperMicro/Tyan are going to be producing dual Cloverton motherboards with a single fsb, do you? I can assure you, they have more sense than that. And since K8L is still roughly 12 months away, neither of us knows whether or not a 4P Opteron board that's bought today will support K8L's at all.

I believe the server-side Rev B will be coming in Q2, the desktop part is what we'll see in Q3 (that's still 8-10 months, by the way). Yes, Cloverton will indeed use dual-FSB and quad-channel FB, though having so many cores will increase cache-coherency traffic massively, no questions about that. Even Opterons with their direct-connect architecture suffer quite a bit with more than 4 cores.

 

[DrMrLordX]

Originally posted by: myocardia
And since K8L is still roughly 12 months away, neither of us knows whether or not a 4P Opteron board that's bought today will support K8L's at all.

For what it's worth K8L should launch on Socket F. At least according to a quote here:

http://www.digitimes.com/bits_chips/a20060314PR200.html

"Well, if we?re going to be more open about 2007 plans, I think Computex would be a good time to start saying a little more about it since that?s where we are introducing our AM2 new infrastructure. That's not to say we're going to present K8L at Computex?don't get me wrong?but I think that that would be a good time to start to disclose more about the future because one of the strong attributes of our roadmap, both in 2006 and 2007, is socket compatibility. The nice thing we're going to do is to deliver to customers. Whatever improvements K8L will provide, they will be applicable to some of the sockets we will be introducing. Therefore, there's a certain logic, to my mind, in disclosing more at that time."

- Henri Richard, AMD Executive Vice President and Chief Officer for Marketing and Sales

Kind of an old-ish quote there, but he's making it sound like K8L should work on current Socket F boards. Should.

 

[myocardia]

Originally posted by: DrMrLordX

Originally posted by: myocardia
And since K8L is still roughly 12 months away, neither of us knows whether or not a 4P Opteron board that's bought today will support K8L's at all.

For what it's worth K8L should launch on Socket F. At least according to a quote here:

http://www.digitimes.com/bits_chips/a20060314PR200.html

"Well, if we?re going to be more open about 2007 plans, I think Computex would be a good time to start saying a little more about it since that?s where we are introducing our AM2 new infrastructure. That's not to say we're going to present K8L at Computex?don't get me wrong?but I think that that would be a good time to start to disclose more about the future because one of the strong attributes of our roadmap, both in 2006 and 2007, is socket compatibility. The nice thing we're going to do is to deliver to customers. Whatever improvements K8L will provide, they will be applicable to some of the sockets we will be introducing. Therefore, there's a certain logic, to my mind, in disclosing more at that time."

- Henri Richard, AMD Executive Vice President and Chief Officer for Marketing and Sales

Kind of an old-ish quote there, but he's making it sound like K8L should work on current Socket F boards. Should.

Yeah, I agree that it should, but, in the computer world, 15-18 months is a lifetime. Don't hold your breath, in other words. And notice that he didn't say which socket.

I believe the server-side Rev B will be coming in Q2, the desktop part is what we'll see in Q3 (that's still 8-10 months, by the way).

Yeah, you're right about servers. I forgot that the desktop parts were coming after the server K8L's. But, you do realize that September 30th is >11 months away, don't you? To me, roughly 12 months, and >11 months are pretty much identical.

 

[Furen]

Originally posted by: myocardia
But, you do realize that September 30th is >11 months away, don't you? To me, roughly 12 months, and >11 months are pretty much identical.

It's 8 months and 1 week to the 1st of July, the first day of Q3 and 10 months and a week 'till the begining of September.

It's pretty much confirmed that all Rev Bs will work with current DDR2 motherboards but they'd be using 1GHz Hypertransport instead of HT 3.0. This is particularly important for server parts and more so in the 4-way space.

 

[myocardia]

Originally posted by: Furen
It's 8 months and 1 week to the 1st of July, the first day of Q3 and 10 months and a week 'till the begining of September.

It's pretty much confirmed that all Rev Bs will work with current DDR2 motherboards but they'd be using 1GHz Hypertransport instead of HT 3.0. This is particularly important for server parts and more so in the 4-way space.

Hah, quarters are three months long, not two. So, it's >11 months until the end of Q3 '07, like I said to begin with. I agree that there's a chance that K8L's will function in a handicapped way on Socket F. Notice that not one single person from AMD has said so. That means that none of us know yet.

If you want me to start assuming, then I would have to guess that there will be some Socket F K8L's, but the reason that AMD doesn't want to say that there will be is that they would then be asked about how well they'll perform. My best guess would be that they won't be performing much differently from a Socket F Opteron, since they'll not only be hindered by the 1Ghz Hypertransport, but will also be slowed considerably by the slower DDR2 RAM.

And I don't think that I'm the only person who has realized all of this. That was why I said in my first post that I think that AMD will be losing a vast amount of the marketshare that they've gained in the server market. I have to assume that anyone who builds a server in the next 6 months will be going with the Xeon 3000 and 5000 series, because they not only outperform the Socket F Opterons, but they'll be a drop-in upgrade, once Cloverton is released, and with no performance hit, like a Socket F K8L will have.

 

[Hard Ball]

Originally posted by: myocardia
You don't really think that SuperMicro/Tyan are going to be producing dual Cloverton motherboards with a single fsb, do you?

Expect Clovertowns to be all introduced on Blackford with DIB, so it should perform roughly equivalent; or fall short a little of 2 X Kentsfield's throughput, due to coherency issues; depending on the workload.

I can assure you, they have more sense than that. And since K8L is still roughly 12 months away, neither of us knows whether or not a 4P Opteron board that's bought today will support K8L's at all.

K8L will be here in about 6-7 months. virtually every source, public or private, that I have seen, say the same thing.

 

[Viditor]

Originally posted by: myocardia

Originally posted by: Viditor
I disagree about Cloverton...
1. Expecting 2 Clovertons on a single FSB to perform on par with 1 Kentsfield (especially in a server environment) is a mistake. Remember that you are greatly increasing the traffic, especially as half of the cache coherency must use the FSB as well.
2. A dual Cloverton platform is not upgradable, whereas a 4P Opteron platform can not only be upgraded to 16 cores, it can also be upgraded to the K8L (Rev B) cores at the same time, which will arguably (even with an equal number of cores) be faster than the Clovertons.

You don't really think that SuperMicro/Tyan are going to be producing dual Cloverton motherboards with a single fsb, do you? I can assure you, they have more sense than that. And since K8L is still roughly 12 months away, neither of us knows whether or not a 4P Opteron board that's bought today will support K8L's at all.

As has been pointed out, K8L quad core for servers (which is what we were discussing) is due in 5-8 months (closer to 8), so let's say you were half right there and leave it at that. (and I will try to find the page, but I remember distinctly that the K8L will be compatable with current boards...you might be thinking of the AM2+ socket info that's floating around, not the servers though)

As to the dual FSBs for Cloverton, you need to look at exactly what's meant by that...
The signal path for the Cloverton is split into 2 DPs per chip...so for cache coherency the path is either DP1A->NB->DP1B using FSB1, or DP1A->NB->DP2A or 2B using FSB 1 and 2...notice that the NB is a common bottleneck.
Secondly, data from all 8 cores goes through a single memory controller to the system ram (even with 2 FSBs).
With the 4P Opteron boards, each DP has it's own Ram which is directly connected to the others, and cache coherency doesn't share a common bottleneck...

This is the scenario that gives HT such a big advantage (and the reason why Intel is striving mightily to get CSI completed).

 

[Zap]

Originally posted by: myocardia
I think that AMD will be losing a vast amount of the marketshare that they've gained in the server market.

Servers (and generally speaking computers marketed towards business) don't go through the quick product refreshes that desktop computers go through. It takes a while to build momentum and a while to lose it. For the years that Opterons were the "better" choice, why didn't AMD overtake Intel in weeks or months? IS managers do not think like enthusiasts. They want proven platforms and long term support in terms of service, parts and replacement availability. Do you think companies are going to "upgrade" their ½ year old Dell Xeon servers to Dell Opteron servers now, and Woodcrest (or whatever) servers ½ year from now? No. They're going to use whatever servers they have until they come to their planned upgrade cycle a few years from now and no amount of manufacturer benchmarks or chest beating by fanboys will change that.

 

[Furen]

Originally posted by: myocardia
Hah, quarters are three months long, not two. So, it's >11 months until the end of Q3 '07, like I said to begin with. I agree that there's a chance that K8L's will function in a handicapped way on Socket F. Notice that not one single person from AMD has said so. That means that none of us know yet.

If you want me to start assuming, then I would have to guess that there will be some Socket F K8L's, but the reason that AMD doesn't want to say that there will be is that they would then be asked about how well they'll perform. My best guess would be that they won't be performing much differently from a Socket F Opteron, since they'll not only be hindered by the 1Ghz Hypertransport, but will also be slowed considerably by the slower DDR2 RAM.

The desktop Rev Bs will be released in Q3, not the end of Q3. It's 11 to the end of Q3 but the chips can be released anywhere between month 8 and month 11. Usually we see stuff released during the July/August period and I'd assume it'll be the same for Rev B but we won't know for sure until we get closer to the release time (remember that by this time Opteron Rev Bs will already be a few months in production, so it's not like it'd be a brand-new architecture being introduced).

Now that's a big assumption. K8 doesn't really benefit much from DDR2. It can achieve some crazy bandwidth numbers but it cannot use the ram to the max. Kentsfield doesn't seem to be bottlenecked by dual-channel DDR2 either so what would make you assume that Rev B will? If you want to talk about slow ram all you have to do is look at FB dimms which Intel uses pretty well with its Woodcrests (it does seem to be a bit of a bottleneck but nothing too extreme). Intel plans to keep using these 4 FB channels to drive 8 cores (with Cloverton) and it still doesn't look like they'll be a huge bottleneck. In 2-way systems Hypertransport (even current 1GHz HT) will not be a bottleneck. Why, you ask? The shared L3 cache which, as I understand, will reduce cache coherency traffic massively. I still don't get why you say Rev B won't perform too differently from current K8s... it'd like saying that Merom does not perform too differently from Yonah since it has the same FSB...


Viditor: As I understand there will also be a socket F+ late next year. The difference being that it includes HT3. All socket F CPUs should be able to drop into it, however, because the only "link" between the CPU and the motherboard is the HT link and that is fully backwards-compatible. Heck, the AM3 CPUs could also be backwards compatible since DDR2 and DDR3 are similar enough that you could get away with getting a single memory controller to drive both types.