Bulldozers Weak/Strong points?

[formulav8]

You would think by now the typical consumer space would be threaded. Even the simplier things like virus scanning (although probably more disk bound) would be multi-threaded. Video encoding has finally taken the mt approach and it makes a huge difference. 3D games should definitely be highly threaded but they still appear to be lagging. It does look like some of the brand new gaming engines takes advantage of multi-threading. :thumbsup: Although this is something that should have started Years ago. :whiste:

 

[lol123]

Server - give us a CPU that can service/retire a bajillion threads in a high throughput environment because not only do we have multi-threaded apps but we are multi-tasking those multi-threaded apps thanks to virtualization.

Isn't that what Sun did with Niagara/UltraSPARC T1? And judging by the reviews I've seen, that approach hasn't really paid off in performance as opposed to more conservative x86 processors like the Opteron.

For example here: http://tweakers.net/reviews/649/2/d...nt-amd-opteron-the-ultrasparc-t1-niagara.html

Maybe the new Bulldozer architecture will yield a better combination of a large amount of cores that still have high single-threaded performance though.

 

[Idontcare]

Isn't that what Sun did with Niagara/UltraSPARC T1? And judging by the reviews I've seen, that approach hasn't really paid off in performance as opposed to more conservative x86 processors like the Opteron.

For example here: http://tweakers.net/reviews/649/2/d...nt-amd-opteron-the-ultrasparc-t1-niagara.html

Maybe the new Bulldozer architecture will yield a better combination of a large amount of cores that still have high single-threaded performance though.

Yes, SUN is the extreme example with their T1 and T2 Niagara products.

But I wouldn't look to those products from SUN to give us an indication of where future products like those are going to perform because the T1 is a 90nm product meaning the xtor budget was really tiny compared to today's and tomorrow's design budgets. (even the T2 is a now aging 65nm part)

SUN had to make too many trade-offs to shoehorn their design into the xtor budget and power-envelopes afforded by the 90nm and 65nm nodes.

A lot more can be done with a modern 32nm node, fewer trade-offs need be made in the design.

 

[Cerb]

You would think by now the typical consumer space would be threaded. Even the simplier things like virus scanning (although probably more disk bound) would be multi-threaded. Video encoding has finally taken the mt approach and it makes a huge difference. 3D games should definitely be highly threaded but they still appear to be lagging. It does look like some of the brand new gaming engines takes advantage of multi-threading. :thumbsup: Although this is something that should have started Years ago. :whiste:

Software has been lagging. CPUs were able to go much faster every time somebody tried to make it mainstream, before (such as Be), and so now we're behind the ball. That inertia will take awhile to deal with. In a game, FI, you will not just have to give it more threads, if you want it to scale very well: you will have to take a new look at what you want to do, and change your problem into being a more parallel-friendly one, and work at threading out from the beginning, with your important world/entity data structures. You might get a little bit by tacking on some physics that can run on its own, but that won't make too much good use of the extra cores we will keep getting.

 

[formulav8]

I was actually thinking physics would be good to use for multi-thread purposes. Why would it not be worth it?

 

[Cerb]

I was actually thinking physics would be good to use for multi-thread purposes. Why would it not be worth it?

Yes, but how far can you go? If the main game engine is single-threaded, physics becomes one of a very few things you can add, and you can only add so much before it bogs performance down (taking more work to keep resources synchronized than what the extra threads can do for you), and/or stops scaling (this is a problem, regardless, but game engines represent such a varied workload, that I can't imagine we are anywhere near practical limits). At some point, whatever you might add that scales out, needs to enabled to do so from the very beginning, to make good use of more and more threads. Threads need to each do different work, communicate their state with each other, not step all over each other's memory, yet all end up at the same point, synchronized, when the next thread or tick (if frames are asynchronous wrt the game world) comes around.

 

[Phynaz]

Isn't that what Sun did with Niagara/UltraSPARC T1? And judging by the reviews I've seen, that approach hasn't really paid off in performance as opposed to more conservative x86 processors like the Opteron.

For example here: http://tweakers.net/reviews/649/2/d...nt-amd-opteron-the-ultrasparc-t1-niagara.html

Maybe the new Bulldozer architecture will yield a better combination of a large amount of cores that still have high single-threaded performance though.

Sun currently holds the top spot in TPC-C benchmarks with a machine that processes 8 threads per core.

Power 7 runs four threads per core, and is one killer of a CPU. A less than fully configured system held the top spot when it was released.

Edit:
Just to be clear, these are big machines, and I'm not saying they are indicative of the performance of a machine most people would buy.

 

[Phynaz]

Yet there is no indication of intel doing that this time around, they are relying on IB being able to beat BD which is what John said was risky.


But that's not the whole story, because MC is superior when the software actually makes use of the extra cores.

When the software is in place the Xeon's 12 faster cores + HT can't compete with 24 "real" cores. There are obviously other factors to consider but on a pure hardware level MC is every bit as good as Xeon on perf/watt.

Johan just benched the Westmere-EX. Picture kinda changes when you actually compare Intel's best against the Opteron.

 

[Abwx]

Johan just benched the Westmere-EX. Picture kinda changes when you actually compare Intel's best against the Opteron.

Westmere 10C/20T is meant to compete with 16C/16T Interlagos
according to Intel , so the comparison with Magny Cours is close
to being irrelevant...

 

[JFAMD]

More importantly, can someone post the costs of all of those systems?

Last I checked there was ~44% higher price between a beckton system and a magny cours system. Westmere EX prices actually went up, so I would love to see a comparison on these two systems at the cost level.

 

[Phynaz]

Westmere 10C/20T is meant to compete with 16C/16T Interlagos
according to Intel , so the comparison with Magny Cours is close
to being irrelevant...

Oh, so it's okay to compare a 12 core Opteron to a 4 core Xeon, but not to a 10 core.
Fanboys are silly.

Anyway, just add 1/3 onto the Opteron scores to make up for the "missing" four cores. How do they do then? Go ahead and add 50% to make up for any thread scaling I may have missed. How do they do then?

 

[Phynaz]

More importantly, can someone post the costs of all of those systems?

Last I checked there was ~44% higher price between a beckton system and a magny cours system. Westmere EX prices actually went up, so I would love to see a comparison on these two systems at the cost level.

Do we get to include software prices in the system cost?

How many more Opteron cores do I need to get equivalent performance to the Xeon system? Let's say I have to license a few more cores of say Oracle, or SAP, or WebSphere or maybe even PeopleSoft on the Opteron. What does that do to the price equation? Over the lifespan of the system?

Or, let's say I'm building out a VDI infrastructure. How many more servers and software licenses do I need to service an equivalent number of users with the Opteron system compared to the Xeon?

That's the dirty little secret that doesn't seem to get much conversation.

 

[Abwx]

Oh, so it's okay to compare a 12 core Opteron to a 4 core Xeon, but not to a 10 core?
Fanboys are silly.

Anyway, just add 1/3 onto the Opteron scores to make up for the "missing" four cores. How do they do then? Go ahead and add 50% to make up for any thread scaling I may have missed. How do they do then?

Core count has nothing to do since each manufacturer has its
own strategy, be it a lot of cores or less cores but with HT,
but the fact is that MC12C is 45nm old history while Westmere10C
is a brand new 32nm high end offering.

So it s basically a soon to be EOL CPU comparison with a product
that will be Intel s best performing CPU for the coming months..

As for performance, AMD said that Interlagos will have 50% better
performance in server apps that the higher binned MC wich is 2.5ghz ,
while the one on AT review is at 2.3ghz, so you can confidently
raise those score by 50% to have an idea about Interlagos perfs..

 

[krumme]

More importantly, can someone post the costs of all of those systems?

Last I checked there was ~44% higher price between a beckton system and a magny cours system. Westmere EX prices actually went up, so I would love to see a comparison on these two systems at the cost level.

This is B2B market, and all that matters is TCO. We can asume the customers is not idiots, and that AMD server share is a mirror of cost / benefit. I remember, even in the days when otellini had to call Michael (or was it opposite), AMD server share was pretty high, and so was profit. That BD better move those integers big time, if you want gold and not silver. 10% extra performance can be 1000% extra profit in the end.

 

[Phynaz]

Core count has nothing to do since each manufacturer has its
own strategy, be it a lot of cores or less cores but with HT,
but the fact is that MC12C is 45nm old history while Westmere10C
is a brand new 32nm high end offering.

Core count has nothing to do with it? You just said this:
"Westmere 10C/20T is meant to compete with 16C/16T Interlagos"

So, are we counting cores or not? You seem to change your mind on every post.


Just for kicks let's say we aren't.

Bench the fastest, highest throughput CPU that is shipping from each manufacturer. That's what Johan did.

BTW, can you please post the link where Intel said "Westmere 10C/20T is meant to compete with 16C/16T Interlagos according to Intel"? Thanks.

 

[Abwx]

BTW, can you please post the link where Intel said "Westmere 10C/20T is meant to compete with 16C/16T Interlagos according to Intel"? Thanks.

It was on an Intel internal note that was published on the web
last year ,though with some lines expurged.

Too lazy to google for it , i ve got it in my HD, but we cant post jpeg
pictures in this site , unless i did miss something , but in short it is said :

"... depending upon the performance of Bulldozer we have the capability
to release a 10C..."

Exactly what they are doing 2 months before BD launch....

 

[Soleron]

"... depending upon the performance of Bulldozer we have the capability
to release a 10C..."

Exactly what they are doing 2 months before BD launch....

Well, they pulled in the launch a little so it would be before BD, or benches would exist before BD.

There is no way they would change whether they would launch a product just because of BD's performance. The chip was in planning three years ago, they've likely sunk tens of millions into the development, so it was just when it would launch.

 

[JFAMD]

Do we get to include software prices in the system cost?

How many more Opteron cores do I need to get equivalent performance to the Xeon system? Let's say I have to license a few more cores of say Oracle, or SAP, or WebSphere or maybe even PeopleSoft on the Opteron. What does that do to the price equation? Over the lifespan of the system?

Or, let's say I'm building out a VDI infrastructure. How many more servers and software licenses do I need to service an equivalent number of users with the Opteron system compared to the Xeon?

That's the dirty little secret that doesn't seem to get much conversation.

The dirty little secret is that the majority of those software licenses are sold as site licenses, so core counts are not relevant. But thanks for trying.

 

[Abwx]

Well, they pulled in the launch a little so it would be before BD, or benches would exist before BD.

Intel know how BD perform since the first samples were sent
to manufacturers and OEMs..

There is no way they would change whether they would launch a product just because of BD's performance. The chip was in planning three years ago, they've likely sunk tens of millions into the development, so it was just when it would launch.

Although WS EX10C was planned, its launch can be made
to keep competitivity in a market, and if a competing product that
is to be laubched is competitive, it make sense to release
an update few time before the competition...

Otherwise, if BD was not in sight , it would have been
way more profitable to simply release a 32nm shrinked
8C with close to 350mm2 rather than the 513mm2 10C.

 

[Phynaz]

The dirty little secret is that the majority of those software licenses are sold as site licenses, so core counts are not relevant. But thanks for trying.

Not so John.

IBM does not site license. They license at the core level based upon architecture ( a point system).

Oracle used to be site licensed, they are moving their customers to the IBM pricing model at renewal.

SUN no longer offers site licenses, and they are now unbundling their software licenses - they no longer license suites.

How do I know this? Because right now I'm orchestrating a move off IBM Power onto x86 that is going to cost millions of dollars up front, and probably an equal amount in FTE dollars. But we more than make up for it in software costs.



Even if the software vendors weren't moving away from site licenses, only the largest companies can afford them. Didn't you say the average server an AMD customer sells is a couple of cores and a couple of $K? Companies that buy that size of a server don't site license enterprise applications.


"But thanks for trying"

 

[purefun1965]

Not so John.

IBM does not site license. They license at the core level based upon architecture ( a point system).

Oracle used to be site licensed, they are moving their customers to the IBM pricing model at renewal.

SUN no longer offers site licenses, and they are now unbundling their software licenses - they no longer license suites.

How do I know this? Because right now I'm orchestrating a move off IBM Power onto x86 that is going to cost millions of dollars up front, and probably an equal amount in FTE dollars. But we more than make up for it in software costs.


"But thanks for trying"

I am seeing this also. we have bulldozer to test. Main reason is our contract is due for renewal and we are being courted.

 

[lol123]

Not so John.

IBM does not site license. They license at the core level based upon architecture ( a point system).

Oracle used to be site licensed, they are moving their customers to the IBM pricing model at renewal.

SUN no longer offers site licenses, and they are now unbundling their software licenses - they no longer license suites.

How do I know this? Because right now I'm orchestrating a move off IBM Power onto x86 that is going to cost millions of dollars up front, and probably an equal amount in FTE dollars. But we more than make up for it in software costs.


"But thanks for trying"

Lol that comes off as highly ironic. IBM pursuing a licensing model that is causing you to migrate away from their hardware platform.

 

[itsmydamnation]

Not so John.

IBM does not site license. They license at the core level based upon architecture ( a point system).

Oracle used to be site licensed, they are moving their customers to the IBM pricing model at renewal.

SUN no longer offers site licenses, and they are now unbundling their software licenses - they no longer license suites.

How do I know this? Because right now I'm orchestrating a move off IBM Power onto x86 that is going to cost millions of dollars up front, and probably an equal amount in FTE dollars. But we more than make up for it in software costs.


"But thanks for trying"

the place i used to work for (left a year ago) is running an oracle licence that hasn't existed for a decade, when you know what buttons to push and you stick to your guns its more then possible to tell these companies to takes there new pricing models and shove it.

 

[Phynaz]

Yes, we have many applications that we are grandfathered into old licensing models. The problem is you are locked into that license. Need to expand your contract because of more users? You are now under the new license and pricing structure. This applies to any contractual change. Want to add a product? Change your support level? Bam, all your licenses just got priced under the new model.

One of our customer service applications was licensed in the mid 90's. Total cost was about $250K for 120 users, with maintanance of something around $50K a year. Due to an increase in needs we had to buy another 60 licenses this year. Because we had to break our grandfathered contract, our maintenance went up to $250K starting next year.

 

[purefun1965]

Yes, we have many applications that we are grandfathered into old licensing models. The problem is you are locked into that license. Need to expand your contract because of more users? You are now under the new license and pricing structure. This applies to any contractual change. Want to add a product? Change your support level? Bam, all your licenses just got priced under the new model.

One of our customer service applications was licensed in the mid 90's. Total cost was about $250K for 120 users, with maintanance of something around $50K a year. Do to an increase in needs we had to buy another 60 licenses this year. Because we had to break our grandfathered contract, our maintenance went up to $250K starting next year.

We are in a similar situation. Imho this is a true statement.