Too early for DDR4 in a mainstream part, GDDR5 also unlikely. If they did bump the top end desktop parts to 13CU it will mainly be to push their combined compute performance marketing. Notice all the tests run are related to GPU compute.
Oh .. well ... lets clarify:The bandwidth numbers are really low for quad channel. Given the 3GB listing it could be tri-channel, it's 50% better than current AMD A series dual channel. I'd hope they could hit 15GB/sec with just improved dual-channel, though.
d) The memory bandwidth is benched single threaded. Hence the enhancements is most probably due to the previously leaked information from BSN about wider interal busses (256bit per module, instead of 128bit).
Don't expect breakthrough iGPU performance (esp. on desktop) until completly new platform (Excavator) is made and DDR4 are on the field.
Kaveri may be small performance bump, but nothing groundbreaking. Maybe laptop Kaveri will be nice improvement vs. laptop Richland and that's it.
And how much do you expecting from Kaveri over Ritchland(6800K) ??
a) The datasheets were speaking of a 3module/6core Kaveri. Maybe they ditched the 6cores and oomped up the GPU instead? Anybody knows how much die area 5 more CUs would need?Personally I doubt 832 SP for kaveri (I could be wrong). Its simply too much die area (7790 is 160mm^2 so expecting around 140 mm^2 for igp alone) to stick on a chip that is selling for ~$150.
No offense, but believing in samples that run at full launch-clocks is reaaly silly ()Also, it seems silly to run 832 sp at 600 mhz.
Personally I doubt 832 SP for kaveri (I could be wrong). Its simply too much die area (7790 is 160mm^2 so expecting around 140 mm^2 for igp alone) to stick on a chip that is selling for ~$150. Add in CPU and cache and its a big chip.
Personally I don't have concrete expectations aside of that it won't be huge leap forward, at least not on desktop.And how much do you expecting from Kaveri over Ritchland(6800K) ??
Personally I doubt 832 SP for kaveri (I could be wrong). Its simply too much die area (7790 is 160mm^2 so expecting around 140 mm^2 for igp alone) to stick on a chip that is selling for ~$150. Add in CPU and cache and its a big chip.
Also, it seems silly to run 832 sp at 600 mhz. Why not run 512 sp at 800-850 mhz, same speed as trinity and richland and similar to the 7750? Especially on an extremely bandwidth bound chip?
Trinity/Richland gets around 15 GB/sec according to sandra and 12 GB/sec according to ADIA64. The 15 GB/sec numbers seen do not seem as much of an improvement (the platform getting 11-12GB/sec seems to be a notebook and that is going to get lower numbers).
Those numbers seem way off compared to my 6800K with dual 1333MHz, it gets somewhere around 18GB/s read and 10.5GB/s write. (Tested Aida64 v3.20, timings 9-9-9-24)
I don't have access to 2133MHz memory right now, but based on what I have seen I'd expect around 21GB/s read for 2133MHz...
Will post screenshot, but don't have time tonight.
As you said, HD7790(896) is 160mm at 28nm. Each Bulldozer module is 30.9mm2 including the L2 Cache at 32nm. It should be a little smaller at 28nm. Add two Modules (Quad Core) is another ~55mm2 at 28nm. Add the NorthBridge and you almost have 240mm2 at 28nm.
Trinity/Ritchland at 32nm is 246mm2 and they are selling for $150 max.
It is better(higher performance/power efficient) to have more Radeon Cores at lower frequency than fewer cores at Higher frequency. Also, Kaveri memory controller + HuMA + GCN will have better memory allocation making it perform better with the same memory Bandwidth than Trinity/Ritchland.
a) The datasheets were speaking of a 3module/6core Kaveri. Maybe they ditched the 6cores and oomped up the GPU instead? Anybody knows how much die area 5 more CUs would need?
b) They probably want to sell that baby for a bit more than $150?
Think about the Opteron version (Berlin) and then think about Seamicro, then check the prices for a FirePro-card with an 7770/7790 GPU-chip. Hint: It is more expensive then a 7770/7790 ;-)
No offense, but believing in samples that run at full launch-clocks is reaaly silly ()
Check the Sandra page for example, that user also has one score for a Trinity chip (7660D) and that one also clocks at 600 Mhz.
Sample speeds are never ever equal to the top launch speed grades.
To the memory benches:
Apples and oranges, you cannot compare one mem bench with another. If the one is single-threaded and the other is multithreaded .. then you have way different results.
Check these instead:
http://www.pcper.com/reviews/Proces...chland-Finally-Lands/Test-Setup-and-SiSoft-Sa
As you see, Richland maxed out at 14 GB/s, however with DDR3-2133. But the Kaveri system just uses DDR3-1600. Thus these 15 GB/s with DDR3-1600 are ways better then the current setup. As I already mentioned before that was expected. It will help single-thread IPC a bit, but neither multithreaded apps nor graphics.
Selling a 246mm^2 chip for $150 max. Look at intel's margins. They need to cut the die size down, especially for mobile where they will sell 240mm^2 cut down chips against intel's chips which are roughly half the size. Power depends completely on what speed the architecture is designed for. At low speeds its pretty linear.
Perhaps, but as of now their first priority is to sell more than having higher margins. And if Kaveri has 832 Radeon Cores it will be a beast.
Perhaps, but as of now their first priority is to sell more than having higher margins. And if Kaveri has 832 Radeon Cores it will be a beast.
It'll be a starving beast chained to the wall if they don't give it enough bandwidth.
Edram or GDDR5 is pretty much needed to drive 832 sp.
Kaveri will be interesting.
Perhaps, but as of now their first priority is to sell more than having higher margins. And if Kaveri has 832 Radeon Cores it will be a beast.
Considering what ShintaiDK said visa vi WSA, seems like designing a larger die with at least three memory channels available for the desktop variants would have made sense - then there would have been plenty of bandwidth. But then FM+ would need to be redesigned to accommodate those who populated the the DIMMs for the third channel. So, Kaveri will be a compromise compared to Intel's embarrassment of riches.
Carrizo, with eDRAM and on 20nm (if GF can pull off 20nm in time for a 2015 launch) would again be a nice APU. With AMD's graphics chops, it might even be able to play today's games in 1080p. Sadly, either AMD, GF or both always seems to come up a little bit short of expectations. DiGiTimes is showing 45W and 65W for Carrizo, which almost certainly means 20nm is the release target, at last look GF's 20nm process looked even more geared toward LP than their 28nm process - so Carrizo will need a big boost in IPC and SPs to hit reasonable desktop APU performance numbers (worse than Intel, obviously on the CPU side, but potentially very good on graphics). As far as laptops go, I think Skylake is just going to mop the floor with anything AMD can produce.
I always want to root for AMD, and they have some interesting products - but they are too constrained process wise and monetarily in terms of design resources to claw their way back. It would be nice if their current plans work out before AMD suffers another massive draw down in their workforce and product line. Financially, while some things are better, AMD still looks to be slowly spiraling down towards VIA-ville.
They cant just add GDDR5 sideport?
AMD have been on the path of no return to "VIAville" for quite some time.
(this is likely not interesting to Americans)
What I have noticed looking at "reklamer" (Danish word, translates as commercials, but also has a specific meaning over here), AMD has quite a presence at the low-end and value segment. From what I can gather Brazos/Kabini systems sell like hot cakes.
Wait... so can some one clarify to me what the current rumours are hinting at with regards to AMD and desktop CPUs? By that I mean are they shifting their emphasis on APUs for the desktop while killing off their FX line? So no more steamroller based CPUs (not APUs) in the nearby future?