PCGH_Carsten
Ex-Redakteur
Hallo,
Thanks for taking the time to answer our questions!
PCGH:
Why did you re-integrate NVIO-functionality into the main die and how much die space (or transistors) did that roughly cost?
Nvidia:
Going to 65nm allows us to integrate more transistors into the chip and reduce routing complexity.
PCGH:
Does the G92/GF8800 GT chip support double-precision to be used in the Tesla-line of cards? If so, what about the number of cycles compared to single-precision?
Nvidia:
The Geforce 8800 GT does not support double precision.
PCGH:
Did you improve on the number of registers per Thread? If i remember correctly from your CUDA-documents, there were about 10,6 for every thread.
Nvidia:
Registers per thread are the same.
PCGH:
How many Quad-TMUs does GF8800 GT use? Are the TMUs still coupled to the 7 TPCs or is their number freely scalable now?
Nvidia:
Up to 8 Texture address ops and texture filtering ops can be done per clock, per TPC. 7 * 8 = 56 Units for the chip. The texture units are coupled to TPCs.
PCGH:
Did any of the filtering characteristics change compared to G80? Does G92 filter some formats now single-cycle which G80 couldn't?
Nvidia:
G92 can do up to 2x more bilinear filtering ops per clock, per TPC.
PCGH:
What is the minimal thread size? Did it change compared to G80?
Nvidia:
It's still the same.
PCGH:
Are the TPC still organized as 16x SIMDs or did you improve on granularity?
Nvidia:
They are the same.
PCGH:
Any important changes to the ROPs? Still single-cycle 4x MSAA, 4 pixels each / alt. 8 Z-Values per clock?
Nvidia:
ROP Compression system is improved for high resolutions with AA. Compression now covers scenes up to 2560 x 1600 with 4xAA (thought it may be shader bound at this point)
PCGH:
Did you improve the geometry shader amplification performance, esp. under heavy load?
Nvidia:
No, it's the same as G80.
PCGH:
Did you improve on triangle-setup? How many tri/s can you setup before and now?
Nvidia:
Up to one triangle per clock, like G80.
PCGH:
Is the GF8800 GT's VP identical to the one used in G84/G86?
Nvidia:
Yes.
PCGH:
Is the mentioned TDP of 100 watts the absolute maximum on stock clocks? Can you give a rough estimate as to what percentage of the TDP one would typically experience in a 3DMark-run?
Nvidia:
110W is the max board power.
PCGH:
Will the G80 chip continue to be used only in the highest-end cards, e.g. GTX and Ultra?
Nvidia:
Yes, absolutely. GTX is more than 20% faster than the 8800 GT at 25x16 4xAA.
PCGH:
What about AGP? Is there a new brigde-chip coming to support GF8800GT chips?
Nvidia:
Sorry, no comments on unannounced products.
PCGH:
SLI: Did you improve on inter card bandwidth? What kind of protocol do you use for SLI communication? Something like PCI-Express?
Nvidia:
We use a combination of PCI Express and the SLI bridge. This remains the same on GeForce 8800 GT.
PCGH:
Thanks for the answers
Das Interview führte Carsten Spille.
Thanks for taking the time to answer our questions!
PCGH:
Why did you re-integrate NVIO-functionality into the main die and how much die space (or transistors) did that roughly cost?
Nvidia:
Going to 65nm allows us to integrate more transistors into the chip and reduce routing complexity.
PCGH:
Does the G92/GF8800 GT chip support double-precision to be used in the Tesla-line of cards? If so, what about the number of cycles compared to single-precision?
Nvidia:
The Geforce 8800 GT does not support double precision.
PCGH:
Did you improve on the number of registers per Thread? If i remember correctly from your CUDA-documents, there were about 10,6 for every thread.
Nvidia:
Registers per thread are the same.
PCGH:
How many Quad-TMUs does GF8800 GT use? Are the TMUs still coupled to the 7 TPCs or is their number freely scalable now?
Nvidia:
Up to 8 Texture address ops and texture filtering ops can be done per clock, per TPC. 7 * 8 = 56 Units for the chip. The texture units are coupled to TPCs.
PCGH:
Did any of the filtering characteristics change compared to G80? Does G92 filter some formats now single-cycle which G80 couldn't?
Nvidia:
G92 can do up to 2x more bilinear filtering ops per clock, per TPC.
PCGH:
What is the minimal thread size? Did it change compared to G80?
Nvidia:
It's still the same.
PCGH:
Are the TPC still organized as 16x SIMDs or did you improve on granularity?
Nvidia:
They are the same.
PCGH:
Any important changes to the ROPs? Still single-cycle 4x MSAA, 4 pixels each / alt. 8 Z-Values per clock?
Nvidia:
ROP Compression system is improved for high resolutions with AA. Compression now covers scenes up to 2560 x 1600 with 4xAA (thought it may be shader bound at this point)
PCGH:
Did you improve the geometry shader amplification performance, esp. under heavy load?
Nvidia:
No, it's the same as G80.
PCGH:
Did you improve on triangle-setup? How many tri/s can you setup before and now?
Nvidia:
Up to one triangle per clock, like G80.
PCGH:
Is the GF8800 GT's VP identical to the one used in G84/G86?
Nvidia:
Yes.
PCGH:
Is the mentioned TDP of 100 watts the absolute maximum on stock clocks? Can you give a rough estimate as to what percentage of the TDP one would typically experience in a 3DMark-run?
Nvidia:
110W is the max board power.
PCGH:
Will the G80 chip continue to be used only in the highest-end cards, e.g. GTX and Ultra?
Nvidia:
Yes, absolutely. GTX is more than 20% faster than the 8800 GT at 25x16 4xAA.
PCGH:
What about AGP? Is there a new brigde-chip coming to support GF8800GT chips?
Nvidia:
Sorry, no comments on unannounced products.
PCGH:
SLI: Did you improve on inter card bandwidth? What kind of protocol do you use for SLI communication? Something like PCI-Express?
Nvidia:
We use a combination of PCI Express and the SLI bridge. This remains the same on GeForce 8800 GT.
PCGH:
Thanks for the answers
Das Interview führte Carsten Spille.
Zuletzt bearbeitet: