I personally find it hard to believe Hawaii is going to stick around in a new lineup when Fiji comes out. A ~500mm2 die, a ~435mm2 die, and a ~365mm2 die? Wow that is quite a bit of reel estate for 3 dies.
I think Hawaii is going to be retired and Fiji will have 3 desktop skus, followed by the "full" Tonga sku. If Hawaii is coming back, then AMD is going to keep losing market share. Plain and simple.
agreed. Hawaii in its present form is not competitive with GM204 in terms of perf, perf/watt, perf/sq mm and DX12 feature set support (Tier 3).
Fundamentally I think the reason for AMD to take so much effort in flushing the channel and take roughly 2-3 quarters to clear existing channel inventory is that the R9 3xx GPU stack is an all new GPU stack with significant architectural improvements over Hawaii and even further improvements over Tonga.
The time is also being utilized to build decent product volume as they expect strong demand. That is why I believe we are seeing a late H1 2015 launch. I expect the R9 3xx products to be competitive with Maxwell in terms of perf, perf/sq mm and perf/watt. That is not possible with existing chips.
Fully agreed.
Large Fiji die can definitely be harvested multiple times to fit a performance segment.
Actually I think there is a dedicated cost optimized chip for R9 380 and R9 380X with HBM. GF 2.5D presentation clearly talked of Cu pillar bump termination being sufficient for die sizes upto 400 sq mm. CRTM bump termination is required for die sizes above 500 sq mm.
http://www.youtube.com/watch?v=po29B53bpic
There are cost and yield benefits in using Cu Pillar as CRTM is a custom solution for one specific customer. No surprise that its AMD and the product that needs it is R9 390 series. So it would have higher costs. GF also said large die (>500 sq mm) large interposer had more yield challenges than die sizes <= 400 sq mm. Since the R9 380 series will be higher volume and most probably in the USD 300-500 price range I expect AMD to have a lower cost higher yielding chip with die size < 400 sq mm and interposer size roughly as big as Kalahari their 28nm CPI test vehicle.
http://www.3dincites.com/2014/11/globalfoundries-3d-ducks-row/
"Kalahari is the technology baseline qualification test vehicle for chip-package interface (CPI) qualification and final inputs to defect management (DM) and the process design kit (PDK).
“
Kalahari mimics a real system that’s going to be built,” explained Alapati. “We have high confidence that we can take any 28nm technology in 2.5D very fast.”
www.amkor.com/go/Kalahari-Brochure
My guess is the chip powering R9 380 series has a < 400 sq mm die size with a interposer size as big as the one used in Kalahari (32 x 26 = 832 sq mm). This chip will be area and cost optimized and have 1/16 fp64 performance like Tonga. This is the higher volume higher yielding chip and can work with CuPillar.
The flagship R9 390 series will use CRTM which is higher cost and will support large die / large inerposer. My guess is die size could be even higher than 550 sq mm as this chip will have full fp64 performance at 1/2 fp32 rate. For Radeon though 1/8 fp64 rate will be supported like R9 290X and for Firepro it will support 1/2 fp64 rate. You have to remember Nvidia gutted fp64 to remain around 600 sq mm. AMD due to HBM which reduces memory controller size and GF 28SHP can easily design a 550 - 580 sq mm chip with fp64 performance.
If you remember the dual link interposer tech allowed 8 GB with 4 HBM stacks on R9 390X. So 4 GB is then possible with just 2 HBM stacks which is how I think we will see this play out. With a 2048 bit memory controller running at 1.25 Ghz the bandwidth is half of what the R9 390X will have. But the 320 GB/s combined with Tonga color compression which provides 40% better memory bandwidth efficiency should be enough to efficiently feed a 2560-2816 sp chip which has 15-20% better perf/sp compared to Hawaii.
Anyway I am excited to see what AMD has designed below R9 390 series because thats what can turn their market share around. There is no way AMD is going to gain back market share with Hawaii and Tonga. The effort they are taking to flush the channel inventory makes me certain that the newer chips are completely new designs with significant improvements to micro-architecture, efficiency and performance.
Facebook
Twitter