kaveri bf4 good performance?

[mikk]

DDR4 wouldn't help really AMD since they are using DDR3-2133 already. Initial RAM speed targets for DDR4 are not much faster than DDR3-2133. Next year we see max DDR4-2133 with Haswell-E and from what we heard Skylake supports max DDR4-2400. Given that the GPU Raw performance increases every year a switch to DDR4 doesn't change the bandwidth limitation.

 

[ShintaiDK]

That also underlines that the only viable solution for AMD is to go the eDRAM route as well.

 

[mikk]

That also underlines that the only viable solution for AMD is to go the eDRAM route as well.

I think so and they wouldn't need 128MB like Iris Pro to make it as cheap as possible. 64MB is fine for Full HD render targets. A fast 64MB low power edram would be nice. DDR4 evolves slowly, they could only make it a triple or quad channel interface to really improve the bandwidth situation. There is also GDDR5M but no idea how much power it eats and the additional costs of such solution.

 

[R0H1T]

DDR4 wouldn't help really AMD since they are using DDR3-2133 already. Initial RAM speed targets for DDR4 are not much faster than DDR3-2133. Next year we see max DDR4-2133 with Haswell-E and from what we heard Skylake supports max DDR4-2400. Given that the GPU Raw performance increases every year a switch to DDR4 doesn't change the bandwidth limitation.

Well not really since AMD could make the memory channel 256bit wide, per DIMM slot, & so it could alleviate the bandwidth problem to a certain extent but as has been said before that they'll just be eating into their money making GPU division's profit & as the such the whole exercise would be self defeating. Thus AMD needs to maintain a certain amount of margins for their CPU division to get back in shape again & for that to happen they'll either have to vacate the low end GPU arena(HD7770 or lower) else artificially place their APU's as mid range offerings in this shrinking x86 CPU market :|

 

[ShintaiDK]

I think so and they wouldn't need 128MB like Iris Pro to make it as cheap as possible. 64MB is fine for Full HD render targets. A fast 64MB low power edram would be nice. DDR4 evolves slowly, they could only make it a triple or quad channel interface to really improve the bandwidth situation. There is also GDDR5M but no idea how much power it eats and the additional costs of such solution.

As far as I know, Intel also only used 128MB to test cases until they was sure they could go with smaller.

I agree, 64MB eDRAM for IGPs as a L4 cache looks to be the perfect option.

Main memory will simply not evolve fast enough unless they get a revolution there.

 

[ShintaiDK]

Well not really since AMD could make the memory channel 256bit wide, per DIMM slot, & so it could alleviate the bandwidth problem to a certain extent but as has been said before that they'll just be eating into their money making GPU division's profit & as the such the whole exercise would be self defeating. Thus AMD needs to maintain a certain amount of margins for their CPU division to get back in shape again & for that to happen they'll either have to vacate the low end GPU arena(HD7770 or lower) else artificially place their APU's as mid range offerings in this shrinking x86 CPU market :|

You would need new memory for 256bit per DIMM. More traces, more pins, cost cost cost. eDRAM is a cheap proven solution in multiple chips and over many years.

The only question is, when will AMD integrate eDRAM. They already proven that they can do it (with eSRAM) from the Xbox One.

 

[NTMBK]

You would need new memory for 256bit per DIMM. More traces, more pins, cost cost cost. eDRAM is a cheap proven solution in multiple chips and over many years.

The only question is, when will AMD integrate eDRAM. They already proven that they can do it (with eSRAM) from the Xbox One.

Remember that the eSRAM isn't implemented as a cache- it's a standalone buffer which requires developer implementation. It probably wouldn't get any use on a PC. AMD would need to make a proper L4 (or L3, given that their APUs don't have one) eSRAM cache, which they have yet to do.

Not that it's impossible, just not as easy as I would hope.

 

[ShintaiDK]

Remember that the eSRAM isn't implemented as a cache- it's a standalone buffer which requires developer implementation. It probably wouldn't get any use on a PC. AMD would need to make a proper L4 (or L3, given that their APUs don't have one) eSRAM cache, which they have yet to do.

Not that it's impossible, just not as easy as I would hope.

True, and AMD still havent connected the GPU to a ringbus with the CPU. But still, it shouldnt be impossible to add for the next CPU release if they already got it planned. It would be rather disasterous if they continue to rely on main memory as the only memory for the GPU part. because its only a question of time before eDRAM with its massive performance boost becomes a defacto standard from Intel.

Unless they perhaps hope for the Volta effect. Wait long enough till they can sell CPUs with stacked on preset memory amount.

 

[R0H1T]

You would need new memory for 256bit per DIMM. More traces, more pins, cost cost cost. eDRAM is a cheap proven solution in multiple chips and over many years.

The only question is, when will AMD integrate eDRAM. They already proven that they can do it (with eSRAM) from the Xbox One.

Sorry but I don't understand, do you mean to say that DDR4 doesn't support 256bit wide memory or does AMD need to redesign their MC from scratch to support such an implementation with DDR4 ?

 

[frozentundra123456]

For desktop gaming, the competition for Kaveri is not really another igp, but AMDs own discrete HD7750/7770 with a cheap cpu. Unless they can overcome the bandwidth and thermal restrictions somehow, Kaveri still will be a solution that is the best choice for only very limited gaming scenarios on the desktop (such a HTPC/SFF uses.)

In addition, eventually new 20nm dgpus will hit the market that will have much better performance at the price point of the current HD7750, making even tougher competition for apus.

 

[ShintaiDK]

Sorry but I don't understand, do you mean to say that DDR4 doesn't support 256bit wide memory or does AMD need to redesign their MC from scratch to support such an implementation with DDR4 ?

There are no 256bit DIMMs(Unfortunately). So you need new DIMMs for a minority segment. That would make them pricy unless they could get Intel on the boat. But it would be nice with 256bit DIMMs, I am getting silly flashbacks to the EDO/FP RAM times with 4 modules at a time.

 

[ShintaiDK]

There is also GDDR5M but no idea how much power it eats and the additional costs of such solution.

Just saw the IHS cost of the PS4 GDDR5(And thats not the mobile solution.). Thats 88$ in wholesale price for the chips only. And you can buy 8GB DDR3 1600Mhz DIMMs in retail for 50$(60$ for mobile). So I think a GDDR5 solution would simply be too costly.

 

[R0H1T]

There are no 256bit DIMMs(Unfortunately). So you need new DIMMs for a minority segment. That would make them pricy unless they could get Intel on the boat. But it would be nice with 256bit DIMMs, I am getting silly flashbacks to the EDO/FP RAM times with 4 modules at a time.

Well then all of this sounds like a worthless exercise to me cause cost escalation is one thing AMD cannot afford in their market segment, the only thing making it interesting now is HSA/HUMA but I have a bad feeling that it could disappoint a lot of unsuspecting folks unless there are some serious optimizations on the software front to take advantage of the two new APU features.

 

[Silverforce11]

Well then all of this sounds like a worthless exercise to me cause cost escalation is one thing AMD cannot afford in their market segment, the only thing making it interesting now is HSA/HUMA but I have a bad feeling that it could disappoint a lot of unsuspecting folks unless there are some serious optimizations on the software front to take advantage of the two new APU features.

IF Kaveri can reach 30% faster than Richland in games, that is good enough for a cheap HTPC light gaming market.

Mantle is the Wildcard.

 

[R0H1T]

IF Kaveri can reach 30% faster than Richland in games, that is good enough for a cheap HTPC light gaming market.

Mantle is the Wildcard.

I'll be glad if it does, happier if the whole APU performance went up ~30% over Richland with HSA/HUMA & Mantle contributing something or the other in the process.

 

[Atreidin]

I think if AMD can get Kaveri to efficiently hybrid crossfire with add-in GCN cards it would stop their APUs from competing with their low-end GPUs and instead promote add-in cards as a decent upgrade path. I don't know if that is in their plans or not but it would be a good idea for them to look into it anyway.

Imagine a really weird situation where Kaveri+290x and Mantle was the fastest (single add-in card) combination around for some Mantle-optimized game that mitigated the CPU bottlenecks. I'm not saying that is likely but it is an interesting thought.

 

[Erenhardt]

You mean this:

I don't think kaveri will have problems running 290X in reasonable settings (>1080p)

 

[Insert_Nickname]

IF Kaveri can reach 30% faster than Richland in games, that is good enough for a cheap HTPC light gaming market.

Mantle is the Wildcard.

I have a 6800K for that exact purpose. While 30% extra performance would be nice, Richland+2133MHz memory does perfectly fine for 720p and older titles.

 

[ViRGE]

Thread cleansed and posters infracted again. This isn't the "Iris Pro costs too much" thread or the "CMT sucks" thread. So I don't know why you guys keep bringing it up.

-ViRGE

 

[Blitzvogel]

Ok well it's obvious (now that I think about it) that a quad channel 768 SP Kaveri could eat into sales of Oland, Cape Verde, and even Bonaire products (to an extent), but especially when price and clocks are considered, it's a different kind of product that can be marketed across a much broader spectrum of users and products. Even if you can't sell an AMD APU, AMD can still attempt to sell an AMD GPU to an Intel user or anyone else applicable.

 

[Unoid]

You would need new memory for 256bit per DIMM. More traces, more pins, cost cost cost. eDRAM is a cheap proven solution in multiple chips and over many years.

The only question is, when will AMD integrate eDRAM. They already proven that they can do it (with eSRAM) from the Xbox One.

Thanks for your conjecture. What are your thoughts on the latest gen consoles using APU's? :biggrin:

 

[Piroko]

Well not really since AMD could make the memory channel 256bit wide, per DIMM slot, & so it could alleviate the bandwidth problem to a certain extent

Going for 256 Bit on a single DIMM doesn't seem feasible. As an example, you would need to up the pinouts of each DIMM from currently 240 (DDR3) to at the very least 240+128 = 368, probably closer to 400 with ECC and other pins. That also makes for double the amount of traces on your motherboard that need to have (very nearly) the same length, which would probably cost more than to simply go triple/quad channel in the first place.

Soldering parts of your memory to the motherboard could be a cost effective option to increase available bandwith, they already did that once with their IGPs with Sideport memory and more and more Notebooks opt for that solution as well. The press resonance would probably be bad though.

 

[ShintaiDK]

In terms of memory, isnt the parallel nature striking back yet again. Both on performance and traces? A serial 32 or 64bit bus would most likely outperform a 256bit parallel one. Its no different than the evolution in PCIe and SATA.

 

[R0H1T]

Thanks for your conjecture. What are your thoughts on the latest gen consoles using APU's? :biggrin:

Now that you mention it the new Xbox seems to have quad channel DDR3 memory support & that should put the onus back on AMD to deliver something similar for Kaveri in the future ~

 

[Piroko]

I think currently two ~1GHz 64 bit parallel links (DDR3 interface clock afaik) are more viable than a single ~4GHz 32 bit parallel link. At least as long as you have to bridge several connectors and a fairly large distance.