AMD CPU speculation... and expert conjecture

[palladin9479]

Search by CPU for CPU2006. T5 for the Oracle and Power 7 for the IBM. A single T5 is a 16 core chip where each core can handle 8 threads and is clocked at 3.6Ghz, has eight memory channels (per chip) and four 10Gbe circuits. You usually buy a dual socket solution but they offer quad and eight socket solutions along with single socket blades. No published M series data which is the big iron mainframe class systems.

Crazy HPC performance.

Always cool to see what the big boys are doing. Wonder how it will compare to the 16 core Ivy-E's coming out.

I found a cheat though in the fine print for T5.
"up to two threads running in each core simultaneously".

After recently reading an IBM power 7 article that's less than what IBM is doing. Their 4way SMT cores are running all 4 threads simultaneously.
http://www.dac.com/App_Content/files/49/DAC_Friedrich_20120603_slides_on_web.pdf
Slide 11

Anyhow that's just 1 aspect of the design. Sun must be making up the performance somewhere else.

Remember it's a RISC design, meaning no need for a front end instruction decoder and needing different schedulers (external and internal). Your binary code is run on bare metal, this is how they can add in all those features. You also gotta understand the nature of HPC. DT / cheap server is mostly about how many integer ops you can perform per second. HPC on the other hand is running so much stuff that chip I/O becomes incredibly important. Feeding the CPU is a greater task then actually transacting the data. This is why Oracle went with eight BoBs (their idea of a memory channel is kinda weird for us) per chip with lower speed memory vs high memory in dual channel mode. The CPU can communicate with up to eight others at 153.6 Gb/sec, this is kinda critical due to NUMA. It can execute two integer instructions per core simultaneously but it can track eight sets of code per core. This is important because of how Solaris handles threading, it tends to not move threads around unless it absolutely has to.


Anyhow a solid T5-2 (dual socket) solution looks to be $55K and includes pretty much everything you'd need except any FC components. Oracle support is stupid expensive though, we hate having to cost for that crap.

 

[Cazalan]

nobody wants to have to stick 4 sticks of ram into an APU system in a laptop. Which is what AMD's main target is. The cost for a larger memory bus would be quite expensive too and probably drop yields. Its an expensive solution that probably won't get much use.

I agree they likely won't go that route. They have years of experience now with eDRAM (WiiU and XBox One). They probably can't afford to go to 128MB like Intel did but even their internal reports said 32MB was sufficient. At any rate it sounds like they have been testing several strategies and they'll pick the one most cost effective at the time. I doubt we'll see that until the chip after Kaveri though.

DDR2400 is about the same price as DDR2133. That's likely the best we'll get for Kaveri.
Some early FM2+ motherboards are marked for 1866

 

[MU_Engineer]

A quad channel memory controller would drive up production costs for both APUs and MBs. It wouldn't be economically viable in their current business model. Could they do it cheaper than SB-E setups? Sure...but that segment isn't huge anyway. Most people don't want to shill out $700 for CPU/MB before they spend more money on the rest of the system. Granted there are some that do...but even if you got that down to $500 it would still be a radical up tick in market for AMD.

AMD does make a four channel socket with Socket G34. You can get a single socket G34 motherboard (Supermicro H8SGL) and an 8-core Opteron (6320) for about $500. The baord is $230 and the CPU is $270. You can get an LGA2011 board and CPU for about $500 as well with an entry-level single-socket board from ASUS, Supermicro, or TYAN (about $300) and the heavily crippled quad-core Xeon E5-2603 for $215. You wouldn't want to actually use the E5-2603 setup as it is a 1.8 GHz quad-core SB-E with no Turbo Boost, no HyperThreading, and a maximum RAM speed of DDR3-1066. And you can't overclock it either.

The reason why the AMD setups can be less expensive is because the G34 CPUs are essentially two separate CPUs in one die. There are two 8-core dies on board and each has its own 2-channel memory controller. The dies communicate via an unusually wide and low-latency on-die HyperTransport link. Each die is much smaller than Intel's giant monolithic 6+ core LGA2011 die so it yields much better and is thus less expensive to make. The two independent sets of memory controller channels on G34 Opterons are easier to route on a board and can be run at DDR3-1866, versus Intel's single four-channel-wide controller traces which top out at DDR3-1600. This makes the G34 Opteron boards less expensive for an otherwise equivalent board.

The benefits from quad channel memory are also not the same as DDR4 or GDDR5...you get marginally more bandwidth for roughly the same cost as adding DDR4 compatability. DDR4 is orders of magnitude more bandwidth, and thus far more effective than quad channel DDR3.

DDR4 appears to top out at something like DDR4-4800 speeds. That would be a little more than doubling of bandwidth compared to DDR3-2133 which is likely going to be the fastest officially supported DDR3 speed. I bet their "10x" comparison uses DDR3 at its slowest official spec of DDR3-800 in a dual-channel arrangement compared to a quad-channel DDR4-4800 setup. That would be 12x as fast...but a completely idiotic comparison since nothing besides some heavily crippled early Intel Xeons ran DDR3 any slower than DDR3-1066, and those setups still had more than 2 memory channels.

Any kind of memory controller other than dual channel DDR3 would drive up production costs. And AMD already have the technology available in their Opteron parts, but would have to create a DDR4 controller from scratch.

Not only that, but DDR4 quite simply isn't available at the moment, and until it is as ubiquitous as DDR3, will be far more expensive.

AMD will have a make a DDR4 controller from scratch, but so will everybody else including Intel. The rumor is that Intel will debut DDR4 in 2014 on Haswell-EX and AMD will debut it in 2014 as well with the replacement to the C32/G34 Opterons. It will take 18-24 months before DDR4 becomes less per GB at a similar performance than DDR3 after DDR4 is debuted. That being said, I'd still got for an early DDR4 setup over a very late DDR3 setup as DDR4 modules will eventually be available in much larger sizes than DDR3. Older desktop computers have a nasty habit of becoming obsolete due to not being able to add more RAM.

 

[Cazalan]

1st kaveri chips "already delivered" per AMD. I am assuming ES's...but launch slated for early Q4 2013.

http://www.hardwareluxx.com/index.php/news/hardware/cpu/26584-computex-2013-amd-shows-qkaveriq-in-action.html

BGA Kaveri chips to have GDDR5 on board memory:

http://www.xbitlabs.com/news/cpu/display/20130530232155_AMD_s_Next_Gen_Kaveri_APUs_Will_Require_New_Mainboards.html

So we can start the timer for "already delivered". Richland was already delivered the first week of Jan and showed up on the shelves in June. So that puts Kaveri at June -> November?


Regarding GDDR5 you did see the note where they'd be limited to 4GB RAM due to current densities? That's rather low for a 2014 computer.

 

[8350rocks]

Yes, but opteron solutions aren't really aimed at consumers...I suppose you could build one. Though, I really don't think the quad channel opteron controller would work in the FX series. Even though the 2 are of the same architecture, the 16 core opterons are a different animal. Integrating them into the desktop FX line would likely present engineering challenges and R&D costs of it's own. The benefits of which would have to be weighed versus the cost of losing business they might have gained.

Considering their current position, I think it would probably cost them less in the long run to just keep dual channel memory, because the percentage of consumer grade PCs with quad channel memory demands is likely so small that what little margin there is in it wouldn't recoup the costs from the R&D.

 

[8350rocks]

1st kaveri chips "already delivered" per AMD. I am assuming ES's...but launch slated for early Q4 2013.

http://www.hardwareluxx.com/index.php/news/hardware/cpu/26584-computex-2013-amd-shows-qkaveriq-in-action.html

BGA Kaveri chips to have GDDR5 on board memory:

http://www.xbitlabs.com/news/cpu/display/20130530232155_AMD_s_Next_Gen_Kaveri_APUs_Will_Require_New_Mainboards.html

So we can start the timer for "already delivered". Richland was already delivered the first week of Jan and showed up on the shelves in June. So that puts Kaveri at June -> November?


Regarding GDDR5 you did see the note where they'd be limited to 4GB RAM due to current densities? That's rather low for a 2014 computer.

It is, but it's much higher bandwidth as well...so processes could be dumped more quickly from the GDDR5 than if you were on DDR3 since the data will transfer much more quickly.

 

[Cazalan]

It is, but it's much higher bandwidth as well...so processes could be dumped more quickly from the GDDR5 than if you were on DDR3 since the data will transfer much more quickly.

BSoN recanted their GDDR5 source saying it's been removed from the NDA documentation. Same with the possible 6 cores number.

"A newer version of the NDA documents seen by us no longer include the references to a GDDR5 interface. The document detailing the GDDR5 interface was from last year, while the newer one we have seen is only a few months old. It seems AMDs engineers for some reason dropped GDDR5 in Kaveri. The possibility of models with 3 compute units (i.e. 6 cores) was removed as well. "

http://www.brightsideofnews.com/news/2013/6/1/amd-updates-roadmaps2c-shows-28nm-kaveri-for-socket-fm22b.aspx

 

[8350rocks]

That's a bit disappointing...I was looking forward to a 6 core Kaveri! That would have been a beast!

 

[Cataclysm_ZA]

Anyhow a solid T5-2 (dual socket) solution looks to be $55K and includes pretty much everything you'd need except any FC components. Oracle support is stupid expensive though, we hate having to cost for that crap.

And they won't help you run Crysis on it either

 

[8350rocks]

Anyhow a solid T5-2 (dual socket) solution looks to be $55K and includes pretty much everything you'd need except any FC components. Oracle support is stupid expensive though, we hate having to cost for that crap.

And they won't help you run Crysis on it either

LOL!!!

 

[Cataclysm_ZA]

LOL!!!

Never gets old, does it? ^_^

 

[The Q6660 Inside]

LOL!!!

Never gets old, does it? ^_^

PUT MY WATERBLOCK BACK ON!

 

[MU_Engineer]

Yes, but opteron solutions aren't really aimed at consumers...I suppose you could build one. Though, I really don't think the quad channel opteron controller would work in the FX series. Even though the 2 are of the same architecture, the 16 core opterons are a different animal. Integrating them into the desktop FX line would likely present engineering challenges and R&D costs of it's own. The benefits of which would have to be weighed versus the cost of losing business they might have gained.

AMD and Intel have both adapted out-and-out server chips for desktop use before. AMD did it with the first Socket 940 FXes and then again with the Socket 1207/L1 QuadFX. Intel did it with the "Skulltrail" LGA771 QX9775 and then unofficially with the EVGA SR-2. The Socket 940 FXes required registered server DDR rather than "desktop" unbuffered DDR. Skulltrail went one further and required FB-DIMMs and was a dual-socket setup to boot. AMD did a little extra work on the Second Generation Opterons to create the QuadFX by removing the need for registered memory, but it still used NUMA (which Windows XP hated) and was a dual-socket setup.

AMD would not need to do a whole lot to make a G34 Opteron into a desktop setup. The only thing that would need to be done is to steer people towards Noctua for an aftermarket heatsink since the 3.5" AM*/FM* heatsinks won't fit. There are already water cooling blocks available for 4.1" G34 socket already. G34 already can use unbuffered DDR3 out of the box. A single socket G34 setup easily fits on an ATX board and a single G34 board is even less expensive than a single LGA2011 board. The chipsets used with the G34 Opterons are virtually identical to the current non-IGP AM3+ chipsets- or you could use an actual desktop 9xx series chipset since the CPU can talk to any chipset that uses HyperTransport. The G34 Opterons are a little lower clocked compared to their desktop FX equivalents due to the higher core count, but they otherwise have identical features such as Turbo Boost. All AMD would really need to do is offer an unlocked G34 Opteron and persuade somebody to make essentially a Supermicro H8SGL with the 990FX/SR5690 chipset and beefier VRMs to handle overclocking. The only negative to this is that I doubt that an overclocked G34 Opteron is really what people are after. Its main advantage over AM3+ FXes is that it would have better multithreaded performance. The 4-module FXes have plenty of multithreaded performance- it's the single-threaded performance that people are panning AMD for lacking (since a lot of older benchmarks and games are poorly threaded.) Plus the G34 Opterons won't clock as high as the AM3+ FXes simply due to their added complexity...and as a result have a lower SuperPi score.

Considering their current position, I think it would probably cost them less in the long run to just keep dual channel memory, because the percentage of consumer grade PCs with quad channel memory demands is likely so small that what little margin there is in it wouldn't recoup the costs from the R&D.

I think that offering G34 Opterons as performance chips won't gain AMD any real money because they would lose the poorly-threaded, ICC-compiled benchmark suites widely used by lesser review sites even to the current FXes. AMD is probably better off in offering special tweaked chips like the FX-9000 Centuriion than in trying one more hurrah in bringing more larger-server hardware to the desktop.

 

[8350rocks]

yes, the Centurion is an interesting piece of silicon...I am watching that one pretty closely to see where the money falls on it. If I can buy one for close to $300.00 I might just do it...

 

[The Q6660 Inside]

Yes, but opteron solutions aren't really aimed at consumers...I suppose you could build one. Though, I really don't think the quad channel opteron controller would work in the FX series. Even though the 2 are of the same architecture, the 16 core opterons are a different animal. Integrating them into the desktop FX line would likely present engineering challenges and R&D costs of it's own. The benefits of which would have to be weighed versus the cost of losing business they might have gained.

AMD and Intel have both adapted out-and-out server chips for desktop use before. AMD did it with the first Socket 940 FXes and then again with the Socket 1207/L1 QuadFX. Intel did it with the "Skulltrail" LGA771 QX9775 and then unofficially with the EVGA SR-2. The Socket 940 FXes required registered server DDR rather than "desktop" unbuffered DDR. Skulltrail went one further and required FB-DIMMs and was a dual-socket setup to boot. AMD did a little extra work on the Second Generation Opterons to create the QuadFX by removing the need for registered memory, but it still used NUMA (which Windows XP hated) and was a dual-socket setup.

AMD would not need to do a whole lot to make a G34 Opteron into a desktop setup. The only thing that would need to be done is to steer people towards Noctua for an aftermarket heatsink since the 3.5" AM*/FM* heatsinks won't fit. There are already water cooling blocks available for 4.1" G34 socket already. G34 already can use unbuffered DDR3 out of the box. A single socket G34 setup easily fits on an ATX board and a single G34 board is even less expensive than a single LGA2011 board. The chipsets used with the G34 Opterons are virtually identical to the current non-IGP AM3+ chipsets- or you could use an actual desktop 9xx series chipset since the CPU can talk to any chipset that uses HyperTransport. The G34 Opterons are a little lower clocked compared to their desktop FX equivalents due to the higher core count, but they otherwise have identical features such as Turbo Boost. All AMD would really need to do is offer an unlocked G34 Opteron and persuade somebody to make essentially a Supermicro H8SGL with the 990FX/SR5690 chipset and beefier VRMs to handle overclocking. The only negative to this is that I doubt that an overclocked G34 Opteron is really what people are after. Its main advantage over AM3+ FXes is that it would have better multithreaded performance. The 4-module FXes have plenty of multithreaded performance- it's the single-threaded performance that people are panning AMD for lacking (since a lot of older benchmarks and games are poorly threaded.) Plus the G34 Opterons won't clock as high as the AM3+ FXes simply due to their added complexity...and as a result have a lower SuperPi score.

Considering their current position, I think it would probably cost them less in the long run to just keep dual channel memory, because the percentage of consumer grade PCs with quad channel memory demands is likely so small that what little margin there is in it wouldn't recoup the costs from the R&D.

I think that offering G34 Opterons as performance chips won't gain AMD any real money because they would lose the poorly-threaded, ICC-compiled benchmark suites widely used by lesser review sites even to the current FXes. AMD is probably better off in offering special tweaked chips like the FX-9000 Centuriion than in trying one more hurrah in bringing more larger-server hardware to the desktop.

Ah, Skulltrail, ahead of its time, although it was just two QX9770s stuck onto a massive SLI board and required FB-DIMMS.

 

[palladin9479]

Anyhow a solid T5-2 (dual socket) solution looks to be $55K and includes pretty much everything you'd need except any FC components. Oracle support is stupid expensive though, we hate having to cost for that crap.

And they won't help you run Crysis on it either

You guys joke, I've recompiled opened source gaming engines for SparcV9 before. Actually it's not that hard to do, the hard part is getting SDL to place nice with the Solaris OpenGL implementation and the framebuffer drivers. I've done Quake 3 on my XVR-1200 (dual modified WildCat IV's on the same board). Took lots of work but I've actually turned a SunBlade 2500 into a "gaming box", was hilarious for a few weeks. So few opensource games worth a damn on Linux made me eventually stop.

http://www.oracle.com/us/corporate/innovation/sparc-t5-deep-dive/index.html

http://www.theregister.co.uk/2012/09/04/oracle_sparc_t5_processor/

 

[8350rocks]

Anyhow a solid T5-2 (dual socket) solution looks to be $55K and includes pretty much everything you'd need except any FC components. Oracle support is stupid expensive though, we hate having to cost for that crap.

And they won't help you run Crysis on it either

You guys joke, I've recompiled opened source gaming engines for SparcV9 before. Actually it's not that hard to do, the hard part is getting SDL to place nice with the Solaris OpenGL implementation and the framebuffer drivers. I've done Quake 3 on my XVR-1200 (dual modified WildCat IV's on the same board). Took lots of work but I've actually turned a SunBlade 2500 into a "gaming box", was hilarious for a few weeks. So few opensource games worth a damn on Linux made me eventually stop.

http://www.oracle.com/us/corporate/innovation/sparc-t5-deep-dive/index.html

http://www.theregister.co.uk/2012/09/04/oracle_sparc_t5_processor/

Yes, there should be more games ported to Linux!!

 

[griptwister]

No, You didn't do your homework. PhysX could run better on a CPU than it would CUDA cores. PhysX is a marketing scheme. I like Nvidia, I'm just sick of all this forced ideas on how gaming should be. If I could afford it, I'd buy a 780 right now, not for PhysX, but because of the performance of the GPU and the capabilities it has with Adobe. Imo, Once Radeon fixes their drivers and makes stronger GPU cores, we're going to see massive improvements across the board and everyone will be neck and neck.

No it wouldn't; Physics, especially multi-object dynamics, KILLS CPU performance even worse then rendering does. Same reason AMD is doing the SAME EXACT THING now with OpenCL: Because Physics runs faster on the GPU.

Then again, lets ignore the guy whos actually worked with the API. Thats cool too.

I could buy that too My information was from when Nvidia 1st introduced physx, so lots could have changed by then lol!

It is, but it's much higher bandwidth as well...so processes could be dumped more quickly from the GDDR5 than if you were on DDR3 since the data will transfer much more quickly.

BSoN recanted their GDDR5 source saying it's been removed from the NDA documentation. Same with the possible 6 cores number.

"A newer version of the NDA documents seen by us no longer include the references to a GDDR5 interface. The document detailing the GDDR5 interface was from last year, while the newer one we have seen is only a few months old. It seems AMDs engineers for some reason dropped GDDR5 in Kaveri. The possibility of models with 3 compute units (i.e. 6 cores) was removed as well. "

http://www.brightsideofnews.com/news/2013/6/1/amd-updates-roadmaps2c-shows-28nm-kaveri-for-socket-fm22b.aspx

Lol, well then... it looks like I'm not touching the FM2+ socket. I really wanted a 6 core power house Guess I'll have to settle for an 8 core AM3+ Steamroller instead.

 

[de5_Roy]

too little too late, gigabyte
http://vr-zone.com/articles/upcoming-gigabyte-fm2-itx-board-to-feature-ir-powerstages/35495.html
then again, better late than never...

AMD Showcases Kaveri APU Performance at Computex 2013
http://www.tomshardware.com/news/AMD-Kaveri-APU-Gaming,22947.html
haven't watched the clip yet.

 

[noob2222]

It is, but it's much higher bandwidth as well...so processes could be dumped more quickly from the GDDR5 than if you were on DDR3 since the data will transfer much more quickly.

BSoN recanted their GDDR5 source saying it's been removed from the NDA documentation. Same with the possible 6 cores number.

"A newer version of the NDA documents seen by us no longer include the references to a GDDR5 interface. The document detailing the GDDR5 interface was from last year, while the newer one we have seen is only a few months old. It seems AMDs engineers for some reason dropped GDDR5 in Kaveri. The possibility of models with 3 compute units (i.e. 6 cores) was removed as well. "

http://www.brightsideofnews.com/news/2013/6/1/amd-updates-roadmaps2c-shows-28nm-kaveri-for-socket-fm22b.aspx

The gddr5 could be a move by sony requesting AMD not drain the market of chips for the PS4 since kaveri and ps4 will be released about the same time. Not good for pricing for either company since sony's profits are based on component costs, wich would translate to fewer chips being purchased by sony.

Then again, the problem could be convincing oems to go gddr5 builds, or its just not working properly.

one thing is for certain, Kaveri is going to be an interesting release.
1: SR Cores
2: GCN APU
3: bulk silicon vs soi
4: 28nm
5: FM2+, not sure if its good or bad due to wich way compatibility will swing

 

[8350rocks]

too little too late, gigabyte
http://vr-zone.com/articles/upcoming-gigabyte-fm2-itx-board-to-feature-ir-powerstages/35495.html
then again, better late than never...

AMD Showcases Kaveri APU Performance at Computex 2013
http://www.tomshardware.com/news/AMD-Kaveri-APU-Gaming,22947.html
haven't watched the clip yet.

The demo is running devil may cry at a nice frame rate on the APU. It's pretty impressive actually.

@noob2222:

FM2+ will take any FM2 APU, but FM2 cannot take a FM2+ APU, much like AM3/AM3+ right now.

I agree the nix on the GDDR5 was likely bowing to a partner (or partners). They clearly have the technology to do it already.

I wasn't aware they were changing APUs/CPUs to bulk silicon @ 28nm. I thought they were still running on PD-SOI from GloFo, but I could be wrong. I haven't been paying as much attention to 28nm as I have 20nm.

According to what I am seeing here, with links to the sites talking about it...28nm will be FD-SOI FinFETs in Gate First configuration:

http://semiaccurate.com/forums/showthread.php?p=177720

There is a lot of speculation in the thread, but if you read down most of the way, a guy posted some links to semimd and a few other places showing GloFo's process plans.

 

[jdwii]

Anybody see Tom's A10-6800k review? The power consumption is actually a bit lower with better performance. This bodes well for future efficiency improvements...though kabini is already a very efficient architecture in mobile space.

Wish we could see a Kaveri release date leaked sometime soon...my kids want a new PC for school/games...so I am eyeballing a kaveri setup. I would do Richland, but what they have now is ok, and with Kaveri 3-6 months out...I would rather wait and build on FM2+.

Intel told us that's impossible.

 

[jdwii]

OK toms i have a hangover and you make a article like this now i'm going to get sick, this is enough proof to stay with something else besides Intel.

http://www.tomshardware.com/news/CPU-Silvermont-Celeron-Pentium-Intel,22979.html

 

[8350rocks]

Anybody see Tom's A10-6800k review? The power consumption is actually a bit lower with better performance. This bodes well for future efficiency improvements...though kabini is already a very efficient architecture in mobile space.

Wish we could see a Kaveri release date leaked sometime soon...my kids want a new PC for school/games...so I am eyeballing a kaveri setup. I would do Richland, but what they have now is ok, and with Kaveri 3-6 months out...I would rather wait and build on FM2+.

Intel told us that's impossible.

Well, it obviously was for them...look at how hasfail turned out.

Though Richland got about 10-15% better across the board and uses about 10% less power.

 

[The Q6660 Inside]

too little too late, gigabyte
http://vr-zone.com/articles/upcoming-gigabyte-fm2-itx-board-to-feature-ir-powerstages/35495.html
then again, better late than never...

AMD Showcases Kaveri APU Performance at Computex 2013
http://www.tomshardware.com/news/AMD-Kaveri-APU-Gaming,22947.html
haven't watched the clip yet.

The demo is running devil may cry at a nice frame rate on the APU. It's pretty impressive actually.

@noob2222:

FM2+ will take any FM2 APU, but FM2 cannot take a FM2+ APU, much like AM3/AM3+ right now.

I agree the nix on the GDDR5 was likely bowing to a partner (or partners). They clearly have the technology to do it already.

I wasn't aware they were changing APUs/CPUs to bulk silicon @ 28nm. I thought they were still running on PD-SOI from GloFo, but I could be wrong. I haven't been paying as much attention to 28nm as I have 20nm.

According to what I am seeing here, with links to the sites talking about it...28nm will be FD-SOI FinFETs in Gate First configuration:

http://semiaccurate.com/forums/showthread.php?p=177720

There is a lot of speculation in the thread, but if you read down most of the way, a guy posted some links to semimd and a few other places showing GloFo's process plans.

AM3 can take an AM3+ CPU and AM3+ can take any AM3 CPU, therefore incorrect. As for 28nm, I have never seen it as that much of an manufacturing process upgrade. GDDR5 will likely only be for BGA, which kind of defeats the upgrading purpose.