AMD server roadmap 2014

[krumme]

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.

Agree. But i dont think it really matters as there cant be any profit here for them anyway. Only cost. Why they targeted server market like they did the last years is crazy because they didnt have the product. Launching bd here was a disaster. Inexcusable management failure.

BD is directly broken. PD is not fixing the problem but just making it less aparent using fx the clock mesh. SR aparently fixes some problems but the arch is still far from the efficiency Haswell can provide. The roadmap is yes a proof of that.

 

[mrmt]

Agree. But i dont think it really matters as there cant be any profit here for them anyway. Only cost. Why they targeted server market like they did the last years is crazy because they didnt have the product. Launching bd here was a disaster. Inexcusable management failure.

Some current AMD execs would say "unmitigated failure".

I think the picture of the Bulldozer disaster is getting clearer and clearer: It cost them their entire server market share (they have to build a new business from the ground up), it killed whatever hopes they had on mobile and to add insult to the injury, they couldn't compete even on desktops, where power efficiency isn't as necessary as in mobile and in servers.

BD is directly broken. PD is not fixing the problem but just making it less aparent using fx the clock mesh. SR aparently fixes some problems but the arch is still far from the efficiency Haswell can provide. The roadmap is yes a proof of that.

I don't think broken is a right choice of words here. Broken implies a sound concept but somehow a botched implementation. There is no big bug in AMD implementation, and 32nm SOI is doing fine so far, which leaves us with the concept.

Bulldozer is a failed concept, no matter how much money and resources you put on that dog, it won't perform as good as the competition.

 

[Vesku]

They had something like 5 years and billions of R&D money to create a core with better IPC than the Athlon based line. They failed at that, even if they'd abandoned CMT and launched Bulldozer as individual cores it would still be a failure because it wouldn't have been much different in performance and performance/watt than a Phenom II at 32nm. Look at Llano and add in theoretical L3 performance gain.

When we get some reviews of Kaveri we will see if the Bulldozer core has any true potential, being the first major revision. That AMD doesn't seem to think there is a market for 4 module Steamroller parts, at least not enough to justify cutting into Kaveri production, is not a good sign.

 

[mrmt]

Interesting bit here. I hope for AMD's sake it's just a rumor:


http://www.theregister.co.uk/2013/06/18/amd_opteron_arm_server_chips/

AMD used Taiwan Semiconductor Manufacturing Corp to fab its Kyoto Opteron X-Series processors using its 28-nanometer processes, and the kicker Berlin Opteron X chips will also come out of the same wafer baker and use the same 28nm processes.

 

[ViRGE]

Interesting bit here. I hope for AMD's sake it's just a rumor:


http://www.theregister.co.uk/2013/06/18/amd_opteron_arm_server_chips/

Huh? Berlin is clearly Kaveri. What were you expecting it to be made on?

 

[Vesku]

Isn't Kaveri supposed to be on GF 28nm?

Seems like GF 28nm is at least in production mode:

http://vr-zone.com/articles/rockchi...168-28nm-socs-tailored-for-tablets/38677.html

 

[AtenRa]

What do you expect them to release on the desktop that they wont release in the 1P server segment?

A 4 Module 8-core FX CPU for AM3+, probably the last AM3+ processors.

They will not release an 8-core Server part because Kaveri based Server APUs(Berlin) will fill that spot. They believe that Berlin with the 512 GCN cores is better suited. But in order to have the 12 and 16 core Server parts, they need to manufacture the 4 Module 8 core dies. They will use that dies for the desktop.

So, just because they will not introduce an 8 core Server CPU doesnt mean they will not do it for the Desktop as well.

 

[AtenRa]

If they had a 4 module Steamroller part, they would be selling it in the server market. They aren't, so they don't.

I wasnt talking about 4 Module SR but 4 Module PD-2(Warsaw).

 

[AtenRa]

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.

The future is FUSION.

 

[Vesku]

A 4 Module 8-core FX CPU for AM3+, probably the last AM3+ processors.

They will not release an 8-core Server part because Kaveri based Server APUs(Berlin) will feel that spot. They believe that Berlin with the 512 GCN cores is better suited. But in order to have the 12 and 16 core Server parts, they need to manufacture the 4 Module 8 core dies. They will use that dies for the desktop.

So, just because they will not introduce an 8 core Server CPU doesnt mean they will not do it for the Desktop as well.

12 and 16 core for 2014 looks to be just a Richland like refresh. I guess Steamroller isn't very competitive vs conventional Intel Xeon so AMD is just providing drop in upgrades for those customers with G34 platforms.

 

[AtenRa]

12 and 16 core for 2014 looks to be just a Richland like refresh. I guess Steamroller isn't very competitive vs conventional Intel Xeon so AMD is just providing drop in upgrades for those customers with G34 platforms.

I dont believe it has to do with performance, even if SR would only give 5-10% over PD it would be even better product. I believe that SR will need a new socket and they will continue to support current Servers for the entire 2014 with PD based Warsaw (less time and resources for server evaluation).

Also i believe AMD have started doing the same thing Intel is doing for the last 2-3 years. They will introduce the mainstream parts first (FM2 APUs) and then introduce the Server and High Performance Desktop products.

Not only that, they trying to push the APUs in every category, from low power small die size all the way to High Performance servers. Intel is doing exactly the same (look Knights Landing).

 

[Vesku]

I dont believe it has to do with performance, even if SR would only give 5-10% over PD it would be even better product. I believe that SR will need a new socket and they will continue to support current Servers for the entire 2014 with PD based Warsaw (less time and resources for server evaluation).

Also i believe AMD have started doing the same thing Intel is doing for the last 2-3 years. They will introduce the mainstream parts first (FM2 APUs) and then introduce the Server and High Performance Desktop products.

Not only that, they trying to push the APUs in every category, from low power small die size all the way to High Performance servers. Intel is doing exactly the same (look Knights Landing).

OK, but why no 3 or 4 module Kaveri? The Berlin socket shouldn't have a problem handling a few extra cores. Either production volume limits or performance/watt are at play here.

 

[AtenRa]

OK, but why no 3 or 4 module Kaveri? The Berlin socket shouldn't have a problem handling a few extra cores. Either production volume limits or performance/watt are at play here.

The future is FUSION.

Except of the part that they trying to push the APUs and not CPUs, having a second die will increase the time to market, resources and budget in order to design, manufacture, server validate etc and that’s not what they need.

 

[Vesku]

Validation shouldn't be much of an issue if it was done in parallel with the 2 module die. There are only a few reasons to not have done a 2M and a 4M Kaveri. Only one I can think of that's at all positive is if console chip production at GF bumped the 4M die.

 

[pablo87]

Interesting bit here. I hope for AMD's sake it's just a rumor:


http://www.theregister.co.uk/2013/06/18/amd_opteron_arm_server_chips/

I feel this could be a big positive for AMD if it means all new products sample and start shipping from TSMC then shift to GF - in theory, it should streamline processes and lead to earlier TTM esp. If TSMC maintains a year or more node advantage.

 

[wlee15]

I think it's a sign of AMD becoming more of a maker of servers through their Seamicro products. After all it's clear that Xeon have a substantial advantage and even if AMD managed to hit a home run it's not gonna change overnight. Plus Seamicro got an excellent core density 256 cores per machine in a 10u space. That's competitive with anything except the densest blade servers despite only quad core uniprocessor Xeons. That will change with the arrival of 12 core Ivy Bridge Xeons, but I suspect that we will see the next Seamicro machine arrive with double the capacity and perhaps with Haswell Xeons.

 

[inf64]

A question for all of you who debate single die PD opteron's lack on roadmaps. What do you think warsaw is made of? It's two PD based 4M dice in MCM arrangement. It's not single huge monolithic die. So yes, they will be making single die 4M parts and they would be stupid to NOT use them on desktop if these show better perf./watt scaling due to tweaked process.

 

[SiliconWars]

Here is an interesting slide that hasn't been published elsewhere for some reason -

http://www.itproportal.com/2013/06/...game-changing-amd-slide-no-one-published-yet/

The interesting part is of course, "OEM's and ODM's want ARM to win"

I suspect everyone is lining up for a bite at Intel's server share and those who currently profit from it want a bigger bite than what Intel is giving them.

 

[Ancalagon44]

I would like to see what AMD could do if they were to develop a custom ARM CPU core.

Perhaps one is already in development and we will see the results in 3 years time?

 

[SiliconWars]

I would like to see what AMD could do if they were to develop a custom ARM CPU core.

Perhaps one is already in development and we will see the results in 3 years time?

I would be surprised if internal AMD roadmaps didn't already have an x86-free look at the end. Note what Feldman said about the faster development time with ARM as well.

 

[grimpr]

Here is an interesting slide that hasn't been published elsewhere for some reason -

http://www.itproportal.com/2013/06/...game-changing-amd-slide-no-one-published-yet/

The interesting part is of course, "OEM's and ODM's want ARM to win"

I suspect everyone is lining up for a bite at Intel's server share and those who currently profit from it want a bigger bite than what Intel is giving them.

Very interesting, thanks. :thumbsup:

 

[mrmt]

Huh? Berlin is clearly Kaveri. What were you expecting it to be made on?

Globalfoundries, but the Reg says TSMC.

I feel this could be a big positive for AMD if it means all new products sample and start shipping from TSMC then shift to GF - in theory, it should streamline processes and lead to earlier TTM esp. If TSMC maintains a year or more node advantage.

Not really Pablo. Those are not compatible processes, so if true instead of finishing a design and start working on an updated version or in a brand new design on the process, AMD engineers will have to backport everything from TSMC to Globalfoundries process just because of the WSA nuisance. That's really bad from a TTM POV.

 

[mrmt]

A question for all of you who debate single die PD opteron's lack on roadmaps. What do you think warsaw is made of? It's two PD based 4M dice in MCM arrangement. It's not single huge monolithic die. So yes, they will be making single die 4M parts and they would be stupid to NOT use them on desktop if these show better perf./watt scaling due to tweaked process.

32nm is already very mature. How much do you think they can extract from this process?

 

[mrmt]

The interesting part is of course, "OEM's and ODM's want ARM to win"

I suspect everyone is lining up for a bite at Intel's server share and those who currently profit from it want a bigger bite than what Intel is giving them.

Why does AMD market always write things they will regret later on?

Would you buy an AMD x86 server after reading this?

 

[SiliconWars]

Globalfoundries, but the Reg says TSMC.

I think they are just mixing up 28nm's. Kaveri is globalfloundries, Berlin will be too.