AMD server roadmap 2014

[podspi]

Very surprising to me there are no big die Steamroller parts on that roadmap. With the WSA, it makes sense to me to make those big die chips.

 

[SammichPG]

My sig limit is reached, so I'm out of luck 🙂.

Oh wow that signature is amazing. 😀

 

[NTMBK]

Very surprising to me there are no big die Steamroller parts on that roadmap. With the WSA, it makes sense to me to make those big die chips.

Hopefully the consoles will take a big chunk out of that WSA- so long as they manage to get at least one of them fabbed at GloFo, of course.

 

[Vesku]

What I get from that is "GF 28nm volume production issues", why else would they be making only 1 version of Steamroller (2 module) and refreshing Piledriver (32nm) for multisocket. Perhaps limited production is aggravated by having at least one of the console chips produced on GF 28nm. XBox chip is rumored to be having problems with yield...

As to their bog standard ARM A57 chips, 2-4x the cores of Jaguar X series and slide says 2-4x performance so the A57 cores will be close to current Jaguar but probably more power efficient. Depending on the sales of the current Jaguar version AMD may maintain both lines depending on whether they will continue to develop successors of the cat family.

 

[Ibra]

AMD must be desperate if they are trying to compete in market segment where they have 0% market share. 😀

 

[mrmt]

Very surprising to me there are no big die Steamroller parts on that roadmap. With the WSA, it makes sense to me to make those big die chips.

Nobody wants them. Server is a fairly low volume market compared to desktops and mobile, and AMD is becoming a rounding error here. It's hard to justify developing and validating a part like that when you have no one to buy it.

 

[mrmt]

What I get from that is "GF 28nm volume production issues", why else would they be making only 1 version of Steamroller (2 module) and refreshing Piledriver (32nm) for multisocket. Perhaps limited production is aggravated by having at least one of the console chips produced on GF 28nm. XBox chip is rumored to be having problems with yield...

Server is very low volume, not high volume. If they are launching Kaveri for mobile, they would be able to launch it for servers without too much trouble even with a problematic node. We already saw AMD doing this before. With 32nm in a subpar state, they launched Llano and customers got Bulldozer server chips well before the launch date. Unless 28nm is in a worse state than 32nm in the beginning of their production run, this is not a probable hypothesis.

The main problem I see here is designing, testing and validating a product to server standards, which are far higher than consumer standards. AMD should have measured the opportunity and saw that the money wasn't there. Given that FX line is a fused off Opteron die, this also tells much about the future of the FX line as we know it. We shouldn't be seeing server chips as FX in the future, if we're going to see FX chips at all.

As to their bog standard ARM A57 chips, 2-4x the cores of Jaguar X series and slide says 2-4x performance so the A57 cores will be close to current Jaguar but probably more power efficient. Depending on the sales of the current Jaguar version AMD may maintain both lines depending on whether they will continue to develop successors of the cat family.

The slide states that ARM will supersede Jaguar in their server line up, not coexist with it.

 

[Khato]

It's more likely to be due to the huge improvement of the ARM cores. Note that the slide says 2-4x performance, not performance per Watt.

Pssst... It's a marketing slide. Do you really expect marketing to not latch onto the perfectly threaded scenario when advertising performance of their 8-16 core CPU? In which case going from a 4 core chip to 8-16 core one at the exact same single threaded performance you end up with 2-4x the performance.

Edit: Guess Vesku beat me to it, that's what I get for just glossing over the thread.

 

[cytg111]

Why dont they make one? They are still working on SR now, so if they wanted to, they could make a 4m SR.

Die size perhaps? If, and thats a big if, the rumored IPC gains are accurate for SR they must have added a great deal of transistors/circuit logic to make it happen.

 

[mrmt]

Question for the AMD crowd:

- If steamroller is everything Bulldozer was supposed to be, if we're going to see smaller die together with 20% higher IPC and same power consumption, why we're not going to see Steamroller server chips as soon as it is launched?

Because if there is one market that needs this wonderchip is servers/workstations, and it would be profitable enough to make up for the incurred costs.

 

[Ancalagon44]

Die size perhaps? If, and thats a big if, the rumored IPC gains are accurate for SR they must have added a great deal of transistors/circuit logic to make it happen.

That could only be a reason if they knew that they would not be able to find a price point for 4m SR that could make them money.

That is to say, larger die size means more expensive to manufacture. They can pass that cost onto business consumers (talking about the server market), however, it might be that they will be unable to compete with Intel on price for this particular market segment.

 

[Vesku]

Question for the AMD crowd:

- If steamroller is everything Bulldozer was supposed to be, if we're going to see smaller die together with 20% higher IPC and same power consumption, why we're not going to see Steamroller server chips as soon as it is launched?

Because if there is one market that needs this wonderchip is servers/workstations, and it would be profitable enough to make up for the incurred costs.

I said why I think that is, production volume issues. Would explain why they are only producing one chip layout of 2 modules + iGP on GF 28nm. If the capacity was there why wouldn't they have 4+ module Steamroller for both consumer and server?

 

[cytg111]

That could only be a reason if they knew that they would not be able to find a price point for 4m SR that could make them money.

That is to say, larger die size means more expensive to manufacture. They can pass that cost onto business consumers (talking about the server market), however, it might be that they will be unable to compete with Intel on price for this particular market segment.

Or thermals get out of hand with a 4 module SR ?

 

[mrmt]

Die size perhaps? If, and thats a big if, the rumored IPC gains are accurate for SR they must have added a great deal of transistors/circuit logic to make it happen.

Not an issue on the server market. In servers you are not counting pennies to be able to sell dies for $40, you can some times sell for 100x as much. Performance and performance/watt are issues.

 

[mrmt]

I said why I think that is, production volume issues. Would explain why they are only producing one chip layout of 2 modules + iGP on GF 28nm. If the capacity was there why wouldn't they have 4+ module Steamroller for both consumer and server?

That's what I'm asking. Couldn't AMD spare a few wafers from the mobile chips just to have a token quantity of halo Steamroller parts if the thing was so great for servers? It would make them far more money than mobile chip, so why not?

Ed: Prices in the server market are so high that you can swallow low yields and still make a descent amount of money.

 

[krumme]

Question for the AMD crowd:

- If steamroller is everything Bulldozer was supposed to be, if we're going to see smaller die together with 20% higher IPC and same power consumption, why we're not going to see Steamroller server chips as soon as it is launched?

Because if there is one market that needs this wonderchip is servers/workstations, and it would be profitable enough to make up for the incurred costs.

TCO on serverside is primarily electricity and especially cooling cost. Even if SR had 40% better efficiency and 20% better ipc it would not be able to compete with HW - neither on ipc or running cost. Die size does only play a very minor role.
Add some software licens is build on number of cores favoring top ipc.

 

[mrmt]

TCO on serverside is primarily electricity and especially cooling cost. Even if SR had 40% better efficiency and 20% better ipc it would not be able to compete with HW - neither on ipc or running cost. Die size does only play a very minor role.
Add some software licens is build on number of cores favoring top ipc.

Agreed here, and wouldn't Steamroller close a bit of the gap with Intel on IPC and power efficiency? Couldn't they sell the thing at discount, as they have been doing since Clovertown? To me it seems that whatever improvements Steamroller brings to the market it won't be able to close any gap with Intel, quite the opposite, tick-tock will still wide the gap regardless of Steamroller.

And if tick-tock killed AMD on the server market, how long until it kills AMD on the mobile market?

 

[Vesku]

That's what I'm asking. Couldn't AMD spare a few wafers from the mobile chips just to have a token quantity of halo Steamroller parts if the thing was so great for servers? It would make them far more money than mobile chip, so why not?

Ed: Prices in the server market are so high that you can swallow low yields and still make a descent amount of money.

Considering they've been getting the cold shoulder from OEMs due to not being a reliable supplier, I'd say splitting limited production doesn't make good sense. Then factor in that even with a really solid improvement from Piledriver to Steamroller a halo part is just going to be somewhat competitive to Intel server chips. Server purchasers aren't as easily blinded by seeing "5 GHz" on the product specs.

You seem to have reasoned this out:

Agreed here, and wouldn't Steamroller close a bit of the gap with Intel on IPC and power efficiency? Couldn't they sell the thing at discount, as they have been doing since Clovertown?

Discount and halo high profit margin are mutually exclusive.

 

[mrmt]

Considering they've been getting the cold shoulder from OEMs due to not being a reliable supplier, I'd say splitting limited production doesn't make good sense. Then factor in that even with a really solid improvement from Piledriver to Steamroller a halo part is just going to be somewhat competitive to Intel server chips. Server purchasers aren't as easily blinded by seeing "5 GHz" on the product specs.

If you have volume big enough to run a mobile line up, how can you not have the volumes for running a server line up?

And yes, I'm expecting that whatever improvements Steamroller brings to the table won't be enough to offset Intel gains in performance and power efficiency in the last few generations. I'm expecting the CPU gap to widen even more.

 

[Vesku]

If you have volume big enough to run a mobile line up, how can you not have the volumes for running a server line up?

Because the volume is just enough (if that) to meet mobile demands? Crank out exactly one die layout- anything that doesn't pass mobile requirements gets kicked up to desktop and server. Why wouldn't they have at least a 4 module desktop Steamroller being produced if 28nm production was fine?

 

[Soleron]

The slide states that ARM will supersede Jaguar in their server line up, not coexist with it.

This needs highlighted a LOT more. Future "cat" cores dead?

 

[mrmt]

Because the volume is just enough (if that) to meet mobile demands? Crank out exactly one die layout- anything that doesn't pass mobile requirements gets kicked up to desktop and server.

Doesn't make sense here, because servers have bigger margins than mobile, so if you can funnel something to servers, you will, simple as that.

Why wouldn't they have at least a 4 module desktop Steamroller being produced if 28nm production was fine?

Demand. There wouldn't be enough volume to justify designing the given SKU, which is an indicative that whatever improvements Steamroller brings, it won't be enough to keep even status quo ante with Intel parts.

 

[mrmt]

This needs highlighted a LOT more. Future "cat" cores dead?

At least for servers, yes.

 

[krumme]

Agreed here, and wouldn't Steamroller close a bit of the gap with Intel on IPC and power efficiency? Couldn't they sell the thing at discount, as they have been doing since Clovertown? To me it seems that whatever improvements Steamroller brings to the market it won't be able to close any gap with Intel, quite the opposite, tick-tock will still wide the gap regardless of Steamroller.

And if tick-tock killed AMD on the server market, how long until it kills AMD on the mobile market?

They can continue to sell Piledriver facelift at a hefty discount, but they are still only considered for the very few cases where the load on the cpu is low most of the time, and the arch is fit for the task - a niche in a niche sold at rock bottom prices.

It really doesnt matter if AMD closes the gab a little or it expands with SR. Just a 10% perf./power advantage turns into massive TCO differences that cpu cost can hardly offset.

When Michael called Otellini in around/before 2004 it was triggered by the serverside and Dells problems here. Obviously because server market is where the huge profit is. The point is, even if you are the big dude doing all the tricks with the OEM, the electricity bill argument, and the system cost around the server, is difficult to cope with, without the best product. This is B2B market, they take no s..t - only money, and second best is of minor use.

What AMD is delegated to is then small niches where they have a competitive advantage with their APU tech. Its certainly there but the market must be very small. But at least - they have an advantage besides price for the cpu. I think that is part of RR strategy - not to compete directly with Intel.

The treat to Intel serverside is not AMD SR competition but how the server market evolves in the comming years in combination with how A57 wrapped in seamicro tech turns out.
HCA might become something but for sure AVX2 will take off big time.

Intel can always leverage on their fab tech that is years ahead of competition - instead of using it only on mobile. On the serverside that is most important and can pay for the cost any day of the week. As i read the recent interview with Otellini thats what he indicated too.

 

[mrmt]

They can continue to sell Piledriver facelift at a hefty discount, but they are still only considered for the very few cases where the load on the cpu is low most of the time, and the arch is fit for the task - a niche in a niche sold at rock bottom prices.

Apparently they can't sell Piledriver even at discount, as they are phasing out 4-8C parts out of the market. That's just the mainstream server market blowing up for them.