AMD: 4-Way Servers Should Be More Affordable.

[god_43]

No problem, you owe me a beer when I am in your town

Facts are your friends, opinions are difficult to back up.

np ;p.

[JF-AMD]

I still firmly believe though that you need Bulldozer in the market as fast as possible, because you're an entire generation behind in per-Core performance and power consumption efficiency, when comparing a K10 vs Nehalem. What I think that it will be interesing to see is a Bulldozer drop in replacement for Magny Cours, that will be pretty easy to do considering that you will have the platform established for then.

Funny, the folks that push the "per core performance" message tend to sidestep the "per core price" and the "per core power."

I find it is better to evaluate things at the processor level vs. the core level, because that is how you actually buy the processors. At the processor level, even though we are at 45nm, you find that performance, power and price are all better with AMD.

True, in single threaded applications I may not have the top speed, but people buying 12-core processor for single threaded applications are a pretty small group.

[zir_blazer]

Funny, the folks that push the "per core performance" message tend to sidestep the "per core price" and the "per core power."

I find it is better to evaluate things at the processor level vs. the core level, because that is how you actually buy the processors. At the processor level, even though we are at 45nm, you find that performance, power and price are all better with AMD.

True, in single threaded applications I may not have the top speed, but people buying 12-core processor for single threaded applications are a pretty small group.

While as someone that dedicates to Server related stuff, Single Threaded performance isn't very important for you because most if not all applications scales with more Cores (In the same way that for a Desktop user, massive Multithreaded performance isn't useful if there are little applications that could do use of it), the fact that AMD uses Core designs that are for the most the same silicon shared for the consumer and server market at the same time, means that looking at what you do, it can also give some tips about how the consumer market may evolve or what products may be released for it.
Magny Cours isn't precisely an example of this because it is obviously a product specifically aimed at the Server market due to its low Frequency and high Core count, that is something that requieres exploiting the fact that Server applications scales well enough with more Threads. But, as good examples, a six Cores K10 (Phenom II X6) isn't precisely news for anyone that may have been following a bit of AMD server market, in the same way that early K8 results made in Opterons also gave expectations on the Athlon 64 performance before launch.


The problem with looking at just the Processor is that it may extend up to the platform, or actually even the Software applications that are to be compared, then back to the Processor or Core level. I favour than the Xeon got a better general purpose performance because it got stronger per Core performance, and while it may not have a high Core count as the new Opterons, and depends on SMT to archieve a higher Thread count, it still gives very competitive performance on heavy Multithreaded applications while beating the Opteron in anything that doesn't scales well with the amount of Threads, and that is where per Core performance wins. In your Server market that doesn't happens very often because applications are designed to scale with more Threads, but in the Desktop market that isn't that common and usually is the other way around, so per Core performance is highly favoured to have tons of Cores. On the other hand, even if you wanted to, there is also a limit on how much wider a Core may be due to instruction level parallelism, so there is also another limit there that depends on how well the application itself is designed. And indeed the universal cure for performance could be raising the MHzs to the maximum that a design may be capable of, but the MHzs plus power consumption and cooling requiered for doing so also got a limit. Well, obviously, here my head is hitting the fact of "Design decitions" that engineerers must made to seek such a balance bewthem all the latter.

If anything, the "per core price" or "per core power" become meaningless under some circunstances, because if we're talking about the Processor as a whole unit, then no matter how many Cores, Threads, or power/price per Core, then everything just resumes in actual performance in the target applications, total power consumption, and price. Yes, you are overally competitive and even best option for value oriented, specifically because the price of both the Processor and the platform. And you depend on the fact that the applications in your target market do decent uses of the high amount of Threads (Something that in the Desktop market doesn't happens too often, and where the Xeon main advantage relies at with any application that doesn't), but the Xeons are there and are comparable even with all the disadvantages that you see on them.
However, I don't like the fact that you need more than twice the silicon to just be competitive with a Xeon, though that is not something that impacts me in any way as a consumer because is not something that provides a positive or negative impact on me, just on your manufacturing cost. But for that amount of silicon, instead of just "being competitive" I would have prefered that you knock out Xeons like the K8 did with Prescott. You aren't helping to feed my AMD fanboism if you can't provide a better Processor in EVERY sense.

My fingers hurts.

[JF-AMD]

Well I am a server guy, I will never admit to understanding the client market.

The fact that we use more silicon "to be competitive" is really immaterial. What matters to the customer is "what am I paying, how does it perform and how much power does it consume?"

Outside of that, Paul Otellini put it best when he said customers don't care what is inside the processor. I couldn't agree more. What matters is what you get out of it, not what is inside it.

[Behemot]

Outside of that, Paul Otellini put it best when he said customers don't care what is inside the processor. I couldn't agree more. What matters is what you get out of it, not what is inside it.

From the customer point of view, I could not disagree either.

But from yours point of view, as a manufacturer, we know you're pretty pissed you have to waste so much of silicon to be at least somehow competitive...you just won't say it publicly of course.

You know, you can say AMD has nice margins and so, but compared to your competitor, Intel has insane margins when pricing smaller silicon higher. That's suddenly what AMD needs as nice margins are not enough to bring you to black numbers. You need some insane margins too Hopefully Bulldozer will push you forward enough to catch up what you missed back at K8 times...

[generics_user]

From the customer point of view, I could not disagree either.

But from yours point of view, as a manufacturer, we know you're pretty pissed you have to waste so much of silicon to be at least somehow competitive...you just won't say it publicly of course.

You know, you can say AMD has nice margins and so, but compared to your competitor, Intel has insane margins when pricing smaller silicon higher. That's suddenly what AMD needs as nice margins are not enough to bring you to black numbers. You need some insane margins too Hopefully Bulldozer will push you forward enough to catch up what you missed back at K8 times...

umm core i 3/5/7 are larger than the competing amd counterparts

[Behemot]

umm core i 3/5/7 are larger than the competing amd counterparts

And what? Have you seen i3/i5 blade somewhere? We're talking about server business here. I am not denying the fact AMD parts are competitive thanks to smaller size in desktop market (more cores for the same price). But it hurts too much in case of Magny Course when selling much bigger die(s) for the same price. But again will you use it? (I personally would love 4core now, I'm optimizing my HDD-usage in last weeks, it means compression of tens of GB's, Brisbane is not best for this you know. But average users don't do such a things.)

In the case of MC you have to admitt average results do not tell a thing. But that's a problem since no reviewer will install 20 separate virtual systems (VS) and than test the performance. MC has potencial for running such a number of separate VS (which is great part of market right now as everybody virtualize). But nowhere else it's better than 2,6-3GHz clocked Nehalem.

MC is very usage-dependent piece of HW. I do not know an application out there which could use 12 cores, so MC is not meant for running single OS. Virtual servers, virtual desktops, cloud computing is the application for MC. Or supercomputers, most of scientists is working hard to make their computing as parallel as possible.

[SocketMan]

I never understood the reasons behind 4p pricing,it felt like it was
priced that high so that people won't buy it.Why would someone get 4P when
you can get 2x2P for a lot less.

Glad it's finally changed.

[charged3800z24]

Maybe now my company will use an AMD based 4p server for our clients. The price can't be ignored.

[hyc]

OpenLDAP has been tested to scale with 90% efficiency across up to 64 sockets (SGI Altix, Itanium 2, dual cores, 1TB of RAM). There's plenty of database workloads that will love Magny-Cours, and these things are the workhorse of commercial enterprises.

[JF-AMD]

I never understood the reasons behind 4p pricing,it felt like it was
priced that high so that people won't buy it.Why would someone get 4P when
you can get 2x2P for a lot less.

Glad it's finally changed.

Well, the secret is that when you bring out products, you establish pricing and a business model. When the first Opteron hit the market, we chose the established pricing model (i.e. expensive 4P), but we brought the price down from ~$3500 to ~$2500. All of the business modeling reflects a mix based on that pricing, which is why it was always difficult to make this change when you are creating 2 different products.

But when we combined the 2000 and 8000 together, it was the right time to finally break the mold. With a 2000 and 8000 you have the same silicon and the same test flow (generally) but you have different inventory costs, planning costs, validation and marketing costs. So they physical cost is the same but there is some overhead costs that are different.

With the 6000 you compress the costs by allowing everything to run through one process and the only difference is some of the validation (2P vs. 4P), but that becomes minor as all of the other aspects are equalized.

Then, on the business side you start to see the benefit for growing share in 2 areas: high end 2P business and expanding 4P business.

If I can sell 4 $744 processors for $2976 and they perform 50% faster in aggregate than 2 x5680's at $3326, then I can capture some of that business. And even if beckton is 10% faster, but it is 250% more expensive, how can you, as a business owner, justify that expense? So I capture a disproportionate amount of the 4P business.

Now, I won't go into margins (they are fine, trust me) but if you look at pure revenue, you see that as an aggregate, if you shift your 2000/8000 mix, you get far more revenue.

Let's take a $750 ASP since that is a nice midpoint for 4P and assume that me sell 1000 processors in a quarter. With ~6% of the mix being 4P, you are selling 60 4P procs for every 940 2P. We will assume that we have a 6% total gain in units. Not unrealistic based on these new prices:

Model Units ASP Revenue
2000 2P 940 350 $329,000
8000 4P 60 1500 $90,000
$419,000

6000 2P 940 350 $329,000
6000 4P 120 750 $90,000
$419,000

You start to see that even though the ASP dropped in half, your revenue to the company stayed the same. AND you picked up market share.

Now, imagine that you can really take this "Value 4P" model to town and really grow share. You are losing those $1500 ASP processors, but you are gaining absolute revenue and margin dollars at a faster rate.

You could always argue that the competition could just match our prices. Have at it. They have a seperate chipset (more money) and they have proprietary memory buffers on each memory channel (more money), so their infrastructure cost cannot be below ours, unless they want to drop their processor pricing below ours.

So, when you start to do the math that argument of "AMD can't make money doing this" starts to fall apart very quickly.

If anyone remembers 1996, when Dell got into the server business in a big way, many argued that with the industry average of $7500 for a 2P server, the new Dell $5000 2P server would not sell because a.) people will perceive it as "cheap" and unreliable and b.) at those prices nobody can make money.

Fast forward a year, and they had gone from ~#8 in the IDC ranking to #3 worldwide, they had grown server revenue and margin, and the overall 2P market grew because people had the same budgets and they could get more servers for their money.

Fast forward 10 years, and you see that prices are even lower than in 1997 AND the market is significantly higher.

So, apparently, the economics hold. At lower prices, people buy more stuff.

This should bode well for AMD in the 4P space.

[JF-AMD]

sorry, the columns didn't import well from the spreadsheet.

[skycrane]

i love low priced 4P systems ive got 3 sitting in the basement myself

i really hope AMD comes through with the price point on these systems, i know a few on the DC teams that woulld replace thier whole pharms with them, i know i would.

[Sn0wm@n]

From the customer point of view, I could not disagree either.

But from yours point of view, as a manufacturer, we know you're pretty pissed you have to waste so much of silicon to be at least somehow competitive...you just won't say it publicly of course.

You know, you can say AMD has nice margins and so, but compared to your competitor, Intel has insane margins when pricing smaller silicon higher. That's suddenly what AMD needs as nice margins are not enough to bring you to black numbers. You need some insane margins too Hopefully Bulldozer will push you forward enough to catch up what you missed back at K8 times...

when say a client orders a big cluster from cray or whatever what they want to have is the highest performing supercomputer.... that draws the less juice from the wall while taking less space for less $ possible ..... and if it brings lots of proffit they dont care how the silicon was used ..... proffit is proffit