Thracks :
I am not a developer, but I will answer as much as I can.
Thx. I know my questions are always not so easy to answer.
1) 4 blocks with 11 CUs each is the configuration of the R9 290X.
Yeah i know, but what is with the real physical implementation on the DIE?
When i think about crossbars, codings and so on, 11 is a REALLY ugly number. As a student of computerachitekturs it is really hard to belief, that there are really only 11 CUs per block. It is really ugly, and i would never design something ugly, when there are no really good reasons, and die size is no really good reaseon. There is enough space for 4 more CUs.
2) GDS via GWS should still be in the architecture.
Yeah, but there is no documentation, and the bit of documentation is gone...
The ISA documentation is no big help, when you don't know the limitations and so on. And nearly nobody would write Assembler code... For SIMD it is possible, but it is notheing you want to do without a really good documentation.
3) Yes, the R9 290 and 290X are 1:8 DP. As to "why" I do not know, except to say that these are primarily gaming graphics cards, so it would make sense that the architectural implementations are optimized for these types of workloads.
Yeah, but the Radeon card were really good development cards and also for students for casestudys. I think there are a lot of people who are really sad, that AMD begin also to cripple there Consumercards. AMD was there a real model student in the past.
I hope really, this is a feature the GPGPU folks wait for years, and it is a REAL gamechanger. It make complet new algorithms on a GPU possible. To bad, that i am waiting since over a year on this feature ;(
5) I do not know what the FirePro team plans.
6) Not sure!
Is there a possibility to get more informations about this? To poor that i am not able to go to the SC....
RDMA shouldn't be such a big thing. More software and a bit of MMU work than everything else. I know what i say, i am porting a linux kerneldriver for a HPC networkcard.
8) 290 and 290X support system unified addressing.
Yeah i know, but the important part was the "
coherent"
Oh i have also some more questions.
9. nVidia made some marketing stunts with there FCAT. Why did AMD no real deep analysis of FCAT, and the problem of divergency between Ingametime and output-/real-time? Framepasing is not the golden eag. It is just a patch, but it do not eliminate the reason for µstuttering. To poor, that AMD not blame nVidia for this.
Is there a chance, that Mantle give the Game-developer enough knowledge about the rendering and the output, so he is able to sync the gametime with the output-/real-time?
10. Is Hawaii 28nm HPM or 28nm HP process from TSMC?
11. I belong to the GPGPU/HPC folk. So FASTRA was a really interesting thing for me since years. How is the situation with FirePro and Radeons, when i want to put 6+ GPUs on a single Motherboard. Will it boot?
I know somebody who let 8 GPUs run under linux, but under Windows the machine don't boot. Is there a driver fix yet, or will it there in there future? GPGPU Development with VirtualStudio and your reall! good performanceanalyser plugin for it makes you want to develop under windows.
EDIT:
Oh and there is one more thing. I try to get some connections since over 2 years to AMD for informations how to read the GPU registers directly.
Why i want this? I want to write my own "MSI Afterburner" for windows and for linux! Unther linux i know no program that allows you to change the voltage, clocks and fan during you monitore the temperatures. linux becomes more important, and so i think there a more than enough people out there who would be happy to have such a program.