Report: AMD Kaveri APUs Will Work in Upcoming FM2+ Socket

[sykozis]

Kaveri benefits from GDDR 5 and considering its not a socket change or SoC there is not enough space for an embed on the die itself so FM2+ will likely have the GDDR5 embeded on the motherboard itself with extra controller in the NB to allocate it exclusively to the iGPU, as for the rest it will be like Z87 is from Z77 thunderbolt, faster SATA, better VRM's, Faster RAM speed support and of course allowes mobo manufacturers to remarket and redesign.

Can we drop the "Kaveri will have GDDR5 memory" rumor?

1) its unsubstantiated
2) it's contrary to AMD's own roadmap for how they are planning to update their APUs.

AMD is planning, with Kaveri to do what they did with Jaguar, and eliminate the "graphic ram/system ram" wall that currently exists in APUs/iGPUs everywhere. This is a key redesign which allows the cpu/gpu to work on a task simultaneously inside the same ram.

Adding gddr5 ram to the Kaveri motherboard would invalidate this whole structural redesign. It won't happen. A llano/trinity/richland APU would benefit from an inbedded chunk of gddr5 ram... Kaveri would not.

HUMA doesn't explain untapped bandwidth and since bandwidth is performance (in gaming circles) where will performance derive from. HD6670 1GB GDDR3 (~40GB/s) vs a HD6670 1GB GDDR5 (~54GB/s) turns into around 15% difference between the cards and the APU's function on the same basis. While HUMA allows the CPU and GPU to access memory independantly that doesn't result in bandwidth, without it no matter the GCN core its going to result in the same outcome a starved iGPU.

I'm failing to see where any of this actually matters on an entry-level/budget system. AMD isn't targeting the high-end market with APUs. They're targeting the entry-level and budget markets where most users are looking for something cheap that will let them browse the web, watch youtube videos and do word processing with light gaming.

 

[alidan]

Sorry, can someone explain how FM2+ vs FM2 is a "more convenient" upgrade than say FM2 to FM3?

It's nomenclature, to indicate that it's an update, not redesign, of socket FM2. it's also named thusly because it RETAINS COMPATIBILITY with current FM2 chips. of course, few would buy a new mobo without a new APU to go with it. but still, the backwards compatibility is nice

personally, a motherboard failure but leaving all the other parts useable is a great thing for these motherboards. my little brother had a capacitor on his motherboard burn out, but everything on the board was left untouched, so all he had to do was swap the board out and a fresh win 7 install for the thing to work again, no other parts besides the motherboard were needed.

 

[sarinaide]

Kaveri benefits from GDDR 5 and considering its not a socket change or SoC there is not enough space for an embed on the die itself so FM2+ will likely have the GDDR5 embeded on the motherboard itself with extra controller in the NB to allocate it exclusively to the iGPU, as for the rest it will be like Z87 is from Z77 thunderbolt, faster SATA, better VRM's, Faster RAM speed support and of course allowes mobo manufacturers to remarket and redesign.

Can we drop the "Kaveri will have GDDR5 memory" rumor?

1) its unsubstantiated
2) it's contrary to AMD's own roadmap for how they are planning to update their APUs.

AMD is planning, with Kaveri to do what they did with Jaguar, and eliminate the "graphic ram/system ram" wall that currently exists in APUs/iGPUs everywhere. This is a key redesign which allows the cpu/gpu to work on a task simultaneously inside the same ram.

Adding gddr5 ram to the Kaveri motherboard would invalidate this whole structural redesign. It won't happen. A llano/trinity/richland APU would benefit from an inbedded chunk of gddr5 ram... Kaveri would not.

HUMA doesn't explain untapped bandwidth and since bandwidth is performance (in gaming circles) where will performance derive from. HD6670 1GB GDDR3 (~40GB/s) vs a HD6670 1GB GDDR5 (~54GB/s) turns into around 15% difference between the cards and the APU's function on the same basis. While HUMA allows the CPU and GPU to access memory independantly that doesn't result in bandwidth, without it no matter the GCN core its going to result in the same outcome a starved iGPU.

I'm failing to see where any of this actually matters on an entry-level/budget system. AMD isn't targeting the high-end market with APUs. They're targeting the entry-level and budget markets where most users are looking for something cheap that will let them browse the web, watch youtube videos and do word processing with light gaming.

In the AMD demo's at Computex they demonstrated Kaveri with emphasis on gaming performance, they want to transition from just a fuddy duddy iGPU to entry level mainstream, so yes they need to find the bandwidth to make that transistion. This will need some form of embedded eDRAM or motherboard embedded GDDR5 or some form of exclusive SO-DIMM for GDDR5 as bandwidth is king when it comes to iGPU performance. Then to the second point I can't remember the head interviewed at the time but top brass in AMD's design basically said that APU's will start entry level and evolve into high end systems, then the rumours that AMD are wanting to unify sockets, all in all these APU's are not destined to remain low end, they are destined to be the evolution of AMD's journey from traditional design to hybrid future systems architecture.

 

[Olov Hansson]

Kaveri benefits from GDDR 5 and considering its not a socket change or SoC there is not enough space for an embed on the die itself so FM2+ will likely have the GDDR5 embeded on the motherboard itself with extra controller in the NB to allocate it exclusively to the iGPU, as for the rest it will be like Z87 is from Z77 thunderbolt, faster SATA, better VRM's, Faster RAM speed support and of course allowes mobo manufacturers to remarket and redesign.

Can we drop the "Kaveri will have GDDR5 memory" rumor?

1) its unsubstantiated
2) it's contrary to AMD's own roadmap for how they are planning to update their APUs.

AMD is planning, with Kaveri to do what they did with Jaguar, and eliminate the "graphic ram/system ram" wall that currently exists in APUs/iGPUs everywhere. This is a key redesign which allows the cpu/gpu to work on a task simultaneously inside the same ram.

Adding gddr5 ram to the Kaveri motherboard would invalidate this whole structural redesign. It won't happen. A llano/trinity/richland APU would benefit from an inbedded chunk of gddr5 ram... Kaveri would not.

HUMA doesn't explain untapped bandwidth and since bandwidth is performance (in gaming circles) where will performance derive from. HD6670 1GB GDDR3 (~40GB/s) vs a HD6670 1GB GDDR5 (~54GB/s) turns into around 15% difference between the cards and the APU's function on the same basis. While HUMA allows the CPU and GPU to access memory independantly that doesn't result in bandwidth, without it no matter the GCN core its going to result in the same outcome a starved iGPU.

I'm failing to see where any of this actually matters on an entry-level/budget system. AMD isn't targeting the high-end market with APUs. They're targeting the entry-level and budget markets where most users are looking for something cheap that will let them browse the web, watch youtube videos and do word processing with light gaming.

In the AMD demo's at Computex they demonstrated Kaveri with emphasis on gaming performance, they want to transition from just a fuddy duddy iGPU to entry level mainstream, so yes they need to find the bandwidth to make that transistion. This will need some form of embedded eDRAM or motherboard embedded GDDR5 or some form of exclusive SO-DIMM for GDDR5 as bandwidth is king when it comes to iGPU performance. Then to the second point I can't remember the head interviewed at the time but top brass in AMD's design basically said that APU's will start entry level and evolve into high end systems, then the rumours that AMD are wanting to unify sockets, all in all these APU's are not destined to remain low end, they are destined to be the evolution of AMD's journey from traditional design to hybrid future systems architecture.

I agree with most of your statement, but this memory detail is a bit off, i think because:

1. Yes, the Kaveri iGPU would oversaturate DDR3@2400MHz/128bit, so i agree there IS a bandwidth problem.
2. Yes, GDDR5 is king, but also expensive. Too expensive for this application, unless
3. They use a small amount as a large and "fast enough" cache.
4. In future APUs, newer memorystandards will definitely be used, but
5. Not in this one, because DDR4 is too immature and not yet properly on the market, and
6. GDDR5, again, is too expensive to be a good investment.
7. Bandwidth can be increased in other ways, for example
8. A wider bus and/or more channels would be an easier to implement and more customer-friendly solution, and
9. Will be necessary in future generations anyway, so
10. It is the most likely solution to the bandwidth bottleneck.

 

[sarinaide]

Kaveri benefits from GDDR 5 and considering its not a socket change or SoC there is not enough space for an embed on the die itself so FM2+ will likely have the GDDR5 embeded on the motherboard itself with extra controller in the NB to allocate it exclusively to the iGPU, as for the rest it will be like Z87 is from Z77 thunderbolt, faster SATA, better VRM's, Faster RAM speed support and of course allowes mobo manufacturers to remarket and redesign.

Can we drop the "Kaveri will have GDDR5 memory" rumor?

1) its unsubstantiated
2) it's contrary to AMD's own roadmap for how they are planning to update their APUs.

AMD is planning, with Kaveri to do what they did with Jaguar, and eliminate the "graphic ram/system ram" wall that currently exists in APUs/iGPUs everywhere. This is a key redesign which allows the cpu/gpu to work on a task simultaneously inside the same ram.

Adding gddr5 ram to the Kaveri motherboard would invalidate this whole structural redesign. It won't happen. A llano/trinity/richland APU would benefit from an inbedded chunk of gddr5 ram... Kaveri would not.

HUMA doesn't explain untapped bandwidth and since bandwidth is performance (in gaming circles) where will performance derive from. HD6670 1GB GDDR3 (~40GB/s) vs a HD6670 1GB GDDR5 (~54GB/s) turns into around 15% difference between the cards and the APU's function on the same basis. While HUMA allows the CPU and GPU to access memory independantly that doesn't result in bandwidth, without it no matter the GCN core its going to result in the same outcome a starved iGPU.

I'm failing to see where any of this actually matters on an entry-level/budget system. AMD isn't targeting the high-end market with APUs. They're targeting the entry-level and budget markets where most users are looking for something cheap that will let them browse the web, watch youtube videos and do word processing with light gaming.

In the AMD demo's at Computex they demonstrated Kaveri with emphasis on gaming performance, they want to transition from just a fuddy duddy iGPU to entry level mainstream, so yes they need to find the bandwidth to make that transistion. This will need some form of embedded eDRAM or motherboard embedded GDDR5 or some form of exclusive SO-DIMM for GDDR5 as bandwidth is king when it comes to iGPU performance. Then to the second point I can't remember the head interviewed at the time but top brass in AMD's design basically said that APU's will start entry level and evolve into high end systems, then the rumours that AMD are wanting to unify sockets, all in all these APU's are not destined to remain low end, they are destined to be the evolution of AMD's journey from traditional design to hybrid future systems architecture.

I agree with most of your statement, but this memory detail is a bit off, i think because:

1. Yes, the Kaveri iGPU would oversaturate DDR3@2400MHz/128bit, so i agree there IS a bandwidth problem.
2. Yes, GDDR5 is king, but also expensive. Too expensive for this application, unless
3. They use a small amount as a large and "fast enough" cache.
4. In future APUs, newer memorystandards will definitely be used, but
5. Not in this one, because DDR4 is too immature and not yet properly on the market, and
6. GDDR5, again, is too expensive to be a good investment.
7. Bandwidth can be increased in other ways, for example
8. A wider bus and/or more channels would be an easier to implement and more customer-friendly solution, and
9. Will be necessary in future generations anyway, so
10. It is the most likely solution to the bandwidth bottleneck.

Since I know Kaveri LGA is not SoC based it rules out embeds and any SO-DIMMS or bolt ons are not apparent from FM2+ leaks so I am inclined to believe the tweaks are Bus and Channel related and maybe down to the improved architecture on both the x86 and GCN iGPU SIMD. Could AMD use a big chunk say 64MB L3 space, highly doubt it due to space. For Kaveri to be a success they will need North of 50GB/s which is double what existing parts are capable of and should see the punted 50% gains. Kaveri at Computex was touted as "focused on gaming" so hopefully the people at AMD that are smarter than us are in fact paying attention to this and/or maybe they actually have a fix and its just secret to us not in the know.

 

[Rob2222]

I stupidly(?) JUST upgraded now from old core duo e8400 config to ATI A10 6800K with new motherboard (gigabyte F2A85X-UP4) & 8GB 1866 RAM but I've continued using my existing ATI 6950 card.

This was all instead of spending a bit more on a Intel 3750k setup. I'm a gamer so want decent gaming performance realise that IS NOT what this chip is for. Its more targetted at low spec, no GPU gaming. What a mistake. I should have just taken it back and replaced before opening ...

Anyway possibly a saving grace for future gaming is Kavari but it will run on FM2+ socket.

Will I be able to upgrade to a Kaveri in the future without replacing my FM2 motherboard? I've read it will work with some limitations?

 

[slomo4sho]

I stupidly(?) JUST upgraded now from old core duo e8400 config to ATI A10 6800K with new motherboard (gigabyte F2A85X-UP4) & 8GB 1866 RAM but I've continued using my existing ATI 6950 card.

This was all instead of spending a bit more on a Intel 3750k setup. I'm a gamer so want decent gaming performance realise that IS NOT what this chip is for. Its more targetted at low spec, no GPU gaming. What a mistake. I should have just taken it back and replaced before opening ...

Anyway possibly a saving grace for future gaming is Kavari but it will run on FM2+ socket.

Will I be able to upgrade to a Kaveri in the future without replacing my FM2 motherboard? I've read it will work with some limitations?

Although you may have been better off with an Athlon X4 760K, I am not sure what you are worried about.

 

[sarinaide]

With a Athlon X4 750/760 he would have forefeited any future HSA benefits and get no existing performance benefits. The only pro to that is price $40 saved is $40 saved.

 

[Rob2222]

Tonight I've been testing how it overclocks I have a cheap Deepcool GAMMAXX 200 cooling it and it reached 4.9Ghz stable, but 5.0Ghz was unstable i'm fairly sure i could have got it running at 5Ghz but didn't want to increase past 1.5V it was still quite cool idle ~33C and under load was between 45 and 65C depending on the fan speed. What i noticed is that as the CPU overclocked i was seeing direct increase in Bioshock infinite benchmark FPS under ultra settings at 1080p. whereas if i overclock my ATI 6950 i see no increase at all. Would this indicate I'm bottlenecked by the CPU, surely not with my GPU right?

 

[slomo4sho]

With a Athlon X4 750/760 he would have forefeited any future HSA benefits and get no existing performance benefits. The only pro to that is price $40 saved is $40 saved.

I am unsure which future benefits you may be speaking of. HSA is only useful for integrated graphics. Since a discrete graphics card is already being used in place of the integrated graphics, there are no benefits of HSA when utilizing a dedicated GPU that can't engage in hybrid crossfire.

 

[sarinaide]

With a Athlon X4 750/760 he would have forefeited any future HSA benefits and get no existing performance benefits. The only pro to that is price $40 saved is $40 saved.

I am unsure which future benefits you may be speaking of. HSA is only useful for integrated graphics. Since a discrete graphics card is already being used in place of the integrated graphics, there are no benefits of HSA when utilizing a dedicated GPU that can't engage in hybrid crossfire.

You can still have the iGP active despite using discrete, HSA will operate the same as though there was no discrete, save Lucid MVP where you can shut down the discrete card. The only benefit of the 750/760 is price other than that its slower mostly due to clock speed (stock frequencies)

 

[clonazepam]

I made a linux box out of llano, for the kid, back in 2011. Lovin' it. Hoping all the best for FM2+. Eventually, I'll get the itch to start tinkering and overclocking again, and this little llano pretty much sucks for that. Still, for what it is, it'd have to burst into flames to give me reason to upgrade it.