Haswell vs Kaveri

So 4GB (which is the maximum possible GDDR5 configuration of this design as far as I know) fits reasonably well with the performance profile/target market of Kaveri while the higher price is justified by extremely high performance for a given TDP.


Since Kaveri is a 2014 product, 4GB of system+video memory doesn't look appealing anymore next year for a next-gen APU. Maybe for real low-end Notebooks in the 300€ range. It is unlikely that Kaveri with GDDR5 is targeted for low-end only. They have Kabini.
 
I do completely agree with all the comments around 4GB being less than ideal for the market segment Kaveris aimed at. It could sit in that segment with only 4GB without looking completely out of place but admittedly it would place it in the consumers eyes at the lower end of that market where its performance and cost would probably require it to be at the higher end of that market in consumer eyes. And unfortunately most consumers only care about bigger numbers.
 
Now that I think about it, using sideport in dual-memory mode did put the IGP using GDDR3 and DDR2/3 at the same time.
AMD is no stranger to using two different types of memory in a GPU, so why not just use the ideal solution of using both the GDDR5 and DDR3 at the same time for much better performance?
Again, it would need to come in a MCM.
 
Now that I think about it, using sideport in dual-memory mode did put the IGP using GDDR3 and DDR2/3 at the same time.
AMD is no stranger to using two different types of memory in a GPU, so why not just use the ideal solution of using both the GDDR5 and DDR3 at the same time for much better performance?
Again, it would need to come in a MCM.

That would require a 256-bit physical interface with all the added cost and complexity that come with it. It would also make unified memory significantly harder.
 
I agree with higher complexity, but if memory capacity is a limit, I think this is one way to go around it. Higher grade versoins of Haswell will do the same. You could think also to put one or two Gbytes of GDDR5 directly on the package, alleviating the costs on the motherboard side. Or, simply, they will have higher capacity GDDR5 memory at launch. But it makes me wonder about the costs (and availability) of the memory modules. A sideport memory is still a cheaper alternative.
 
Not every platform Kaveri or Kabini addresses needs to run full-blown windows style games on a full-blow windows OS; 2GB of GDDR5 would be a fast option for a Tablet with a cut down OS. If they do plan to address laptops or convertibles that do run real windows, 8 Gigs of DDR3 on a 128 bit bus would be plenty for a consumer and it would be faster than existing solutions. The separate GPU / CPU memory architectures of old is not something we should expect or want in the future.
 
That would require a 256-bit physical interface with all the added cost and complexity that come with it. It would also make unified memory significantly harder.

Again: AMD did it already with the 780G IGPs and AFAIK the world didn't implode and the IGP's performance increased.

I could fathom a pair of 4Gbit GDDR5 chips in the same MCM as the APU, resulting in 128bit DDR3 + 64bit GDDR5. That would both increase performance and fit in all the current FM2 boards.
The iGPU would be able to use a total of ~80GB/s.
 
Now that I think about it, using sideport in dual-memory mode did put the IGP using GDDR3 and DDR2/3 at the same time.
AMD is no stranger to using two different types of memory in a GPU, so why not just use the ideal solution of using both the GDDR5 and DDR3 at the same time for much better performance?
Again, it would need to come in a MCM.

It's akin to a Geforce Turbocache or Radeon Hypermemory, including a very low capacity and bandwith, and crap performance. I think the sideport memory was 32bit wide and was said to bring a small power usage optimisation in some circumstances - Intel says Crystalwell is about optimizing power too, I think we should better understand that concept, I don't, except some vague notion about data locality and that it's said to be better for power saving states (at least the old 780G sideport).

An APU using both gddr5 and ddr3, this sounds like using e.g. both ddr2 and ddr3 on a AM3 CPU and a ddr2+ddr3 mobo (the mobo exists, and AM3 CPUs do have two 64bit controllers it seems) but enabling this would be too complex and would have served a very tiny portion of the market.
I think what you wish for will be done by AMD eventually, but rather than using raw gddr5 chips they will do something similar to Crystalwell, custom "memory on interposer" or "2.5D stacking" or whatever you call this.
 
Again: AMD did it already with the 780G IGPs and AFAIK the world didn't implode and the IGP's performance increased.
Not by much though as it was only a 16bit interface which had very low bandwidth on its own - that's very different to a full blown second interface.

I could fathom a pair of 4Gbit GDDR5 chips in the same MCM as the APU, resulting in 128bit DDR3 + 64bit GDDR5. That would both increase performance and fit in all the current FM2 boards.
The iGPU would be able to use a total of ~80GB/s.
Well I guess another possibility would be just split up into 64bit gddr5 and 64bit ddr3. That way you could still have 2 ddr3 dimms for memory expansion (up to 16GB) while having "decent" bandwidth (not really though with the claimed low gddr5 datarate). The asymmetric nature though might make this difficult.
 
So 4GB (which is the maximum possible GDDR5 configuration of this design as far as I know) fits reasonably well with the performance profile/target market of Kaveri while the higher price is justified by extremely high performance for a given TDP.

Where does this your 4 GiB limit come from?

Is it based on
1) Just an assumption/hunch
2) Size of currently available GDDR5 chip sizes
3) GDDR5 specification

If it's not based on GDDR5 specification, then that limit does not exist, and all the speculation about "limiting themselves to only 4 GiB" is meaningless.
 
Where does this your 4 GiB limit come from?

Is it based on
1) Just an assumption/hunch
2) Size of currently available GDDR5 chip sizes
3) GDDR5 specification

If it's not based on GDDR5 specification, then that limit does not exist, and all the speculation about "limiting themselves to only 4 GiB" is meaningless.

2. Assuming 4Gbit chips are the highest density available then the maximum possible memory configuration using clamshell would 4GB (8 chips) on a 128bit bus. 128bit is a given based on the low quoted bandwidth figures.
 
2. Assuming 4Gbit chips are the highest density available then the maximum possible memory configuration using clamshell would 4GB (8 chips) on a 128bit bus. 128bit is a given based on the low quoted bandwidth figures.

Is this 4Gbit the maximum _currently available_ or maximum that will be available inside one year?

Kaveri might even be also nice motivator for memory chip manufacturers to create bigger GDDR5 chips.

It will definetely be 128-bit, 256 bits would be too much for the pcb in that price range.
 
Maximum currently available, but it's scarce as it is and I don't think it's going to be doubling any time soon. Kaveri can't depend on technology a year out, it's going to have to be out at or shortly after launch..
 
The data rates BSN mentioned fit better to GDDR5M, which will be made available on SODIMMs (there is a common physical spec for DDR4/GDDR5M SODIMMs, they are just keyed differently) also using smaller device widths (x16, not only x32), which doubles the maximum amount of memory with a certain chip density. One can connect 16 x16 chips in clamshell mode to a 128bit controller. That makes 8GB possible (4GB per SODIMM with 4GBit chips).
 
The data rates BSN mentioned fit better to GDDR5M, which will be made available on SODIMMs (there is a common physical spec for DDR4/GDDR5M SODIMMs, they are just keyed differently) also using smaller device widths (x16, not only x32), which doubles the maximum amount of memory with a certain chip density. One can connect 16 x16 chips in clamshell mode to a 128bit controller. That makes 8GB possible (4GB per SODIMM with 4GBit chips).

Cheers, I wasnt fully aware of gddr5m but as you say the speeds look right so*it looks to be the obvious solution. That means 8GB Kaveri platforms should be a reality and make perfect sense for the low power/high performance laptop market segment.
 
Nice. OIT seems particularly interesting. Intel thinks it will be able to make devs use it and not just drop intel gpus on the floor.
 
http://vr-zone.com/articles/meet-haswell-s-graphics-thump-card-crystalwell-gt3-edram-/19604.html

8jpEeCw.jpg
 
Back
Top