Haswell vs Kaveri
- Thread starter AnarchX
- Start date
I don't think TSMC has any gate first 28 nm process(so it's either GloFo or Gate Last)
GloFo use Gate-First
http://www.globalfoundries.com/technology/28nm.aspx
So on paper the A10-7850K has a GPU similar to the 7750, just a bit lower clocked (and of course DDR3), correct? Assuming otherwise similar config, that should give it ~730 Gflops, ~11.5 Gpixels, ~23 Gtexels?
So compared to Iris Pro 5200's ~ 830/10.4/20.8 it should be slightly faster on the sorts of workloads that iGPUs typically do (at higher power, but still). And of course a good chunk faster with MSAA. Will be interesting to see how much the DDR3 bottlenecks it, if much at all. Looking forward to reviews!
So compared to Iris Pro 5200's ~ 830/10.4/20.8 it should be slightly faster on the sorts of workloads that iGPUs typically do (at higher power, but still). And of course a good chunk faster with MSAA. Will be interesting to see how much the DDR3 bottlenecks it, if much at all. Looking forward to reviews!
Gotcha. I recall they had some contract w/ Glo Fo's for required use of their capacity, but I don't remember where Kaveri will be made since I've heard conflicting reports. I would hope for them that this contract will be over at some point soon, and that they get the freedom to take advantage of their fab-free status. I seem to recall Glo Fo's 28nm process was delayed so it's conceivable that they might be using TSMC.
No, it's definitely GloFo: Kaveri engineering samples bear the mention "diffused in Germany". TSMC doesn't have fabs in Germany, only GloFo does.
Besides, TSMC has been mass producing 28nm chips since 2011, so a 2014 launch for Kaveri certainly doesn't contradict the notion that GloFo's 28nm is late. Finally, your recollection is correct: AMD does have a contract with GloFo (called the Wafer Supply Agreement) and Trinity, Richland, Vishera are the only AMD products currently made by GloFo. Vishera (server variant included) is barely selling, and Trinity/Richland will be replaced by Kaveri. Everything else is TSMC. If AMD were to move Kaveri to TSMC, there would be almost zero production at GloFo, which would result in severe penalties.
Well you can find HD7750 ddr3 reviews, I would certainly expect this to be in the same ballpark (so about 2/3 the performance of a gddr5 HD7750). I am not actually sure if the graphics really is higher power as the Iris Pro 5200 graphics, since the 7750 ddr3 seems to be around ~40W max (and Kaveri even has 10% lower gpu clocks), possibly there's still not any useful cpu/gpu power management. Or maybe the GloFo process is really bad and graphics alone can draw more.
I think though it will only have 8 ROPs but given the bandwidth that should make no difference (so it might be more like upscaled Mars with 8 CUs rather than downscaled Cape Verde with 8 CUs).
DDR3 will kill any chance this thing has of being a truly capable gaming chip. They need either GDDR5 or high speed embedded memory like Iris Pro.
With the PS4 consuming so much GDDR5, perhaps they could purchase the leftovers that don't make the 5.5Gbps cut? Even 5Gpbs GDDR5 would massively benefit over DDR3.
The OEM's who would have to build that have no interest in such a cost adder with negligable sales effect. Especially for an AMD platform that won't generate any significant revenue in the first place.
Compare Richland HD8670D /48TMU ~40.5 GTexel/s
Kaveri almost half of the texturing performance.
Compare Richland HD8670D /48TMU ~40.5 GTexel/s
Kaveri almost half of the texturing performance.
Richland has 24 TMU's ( 6 SIMD's x 4 TMU's per SIMD).
Well I'm not sure they have many options with GDDR5 to be honest. If you make a side-band interface for just the GPU you sort of undercut the whole APU unified memory story. Conversely, if you try and replace the main memory interface with GDDR5 you hurt the CPU performance (pretty significantly @ 4Ghz) due to increased latency and remove socketing/dimms and other conventional desktop stuff.
Even that review you quoted has an average of 64% which is very very close to 2/3 and not really 1/2. I was going by ht4u review (http://ht4u.net/reviews/2012/msi_radeon_hd_7750_2gb_ddr3_test/index36.php) which had 65% at the end.
But yes this is _very_ dependent on the actual title as you can easily see as the results ranged from -15% to -54%. But your quoting seems a bit biased, if you wanted a somewhat accurate between range you'd have to say between 4/5 and 1/2 or so... I'd certainly agree though that even 2/3 is quite a poor showing.
Performance loss due to DDR3 should be even smaller for Kaveri as it has DDR3 2133 official support and DDR3 2400+ unofficial support. Tested HD7750 DDR3 had 1600MHz memory. Besides AMD probably tweaked few things to get better resource utilization out of slow memory available on the desktop chip.
Yes faster ram will help no doubt - as long as you can get it cheap, otherwise it doesn't make much sense.
I'm not sure how much it is tweaked - well I guess dropping half the ROPs (though that's just my guess) would be such a tweak
. Plus the lower shader clock will also make this a bit more balanced (but obviously not help absolute performance).Yes faster ram will help no doubt - as long as you can get it cheap, otherwise it doesn't make much sense.
I'm not sure how much it is tweaked - well I guess dropping half the ROPs (though that's just my guess) would be such a tweak
Kaveri's GPU is based on the same IP level as Bonaire and Hawaii, so the front-end is different from Cape Verde's. But I'm not sure that will do anything for bandwidth requirements. There's always the relatively straightforward option of making the L2 bigger.