NVIDIA GF100 & Friends speculation
- Thread starter Arty
- Start date
annihilator
Newcomer
and PhysX is actually related to these new cards....
DavidGraham
Veteran
Well , it is my turn to steer you guys back on topic 
:
in Cypress I noticed that the 4 small ALUs in each 5D SP are only capable of 24-bit INT processing .. while being able to do FP32 in the same time , isn't that a bit odd ?
shouldn't FP32 have the same binary capacity as INT 32 ? binary32 for both?
For example like GF100 ALUs , where each one is able to do FP32 and INT 32.
:in Cypress I noticed that the 4 small ALUs in each 5D SP are only capable of 24-bit INT processing .. while being able to do FP32 in the same time , isn't that a bit odd ?
shouldn't FP32 have the same binary capacity as INT 32 ? binary32 for both?
For example like GF100 ALUs , where each one is able to do FP32 and INT 32.
Well , it is my turn to steer you guys back on topic
in Cypress I noticed that the 4 small ALUs in each 5D SP are only capable of 24-bit INT processing .. while being able to do FP32 in the same time , isn't that a bit odd ?
shouldn't FP32 have the same binary capacity as INT 32 ? binary32 for both?
For example like GF100 ALUs , where each one is able to do FP32 and INT 32.
Actually, FP32 numbers are stored as follows:
23 bits for the mantissa
8 bits for the exponent
1 bit for the sign
The mantissa is guaranteed to have a leading '1' bit at the front for normalized floating-point numbers, which adds a free extra bit of precision - leading to 24 bits of precision for each 32 bit floating point number.
If you're curious about FP, there's a lot to learn - here's a good starting point:
http://en.wikipedia.org/wiki/Floating_point
We could also find useless polls from other sites that ask what color fans most users want Fermi to have. That's related to these new cards! [Full disclosure: I prefer pretty pink fans]
pfft, everyone knows the true names are gt300 and r870, no matter what nvidia and wavey say.
Now back to [strike]gf100[/strike] gt300 discussion!
DavidGraham
Veteran
So because of that , the T-Unit must have a binary capacity of 40-bit right ?
if that was right , then Fermi ALUs must be 40-bit wide too ..
Yup, the T-unit is expanded to 40 bits to handle single-cycle INT32 op's.
No, because the Fermi ALUs can't accept 40-bit wide data - they can accept *either* 32-bit FP or 32-bit int. There's no benefit to supporting a strange FP40 format, since it's not standard and no one could use it, neither under DX11 or OpenCL (or even CUDA).
The Fermi ALUs can accept 32-bit wide datatypes, either 32-bit integers or 32-bit FP. The details of how they've been implemented internally are likely much more complicated than just saying they have a "binary capacity of 40 bits", given that the ALUs are ganged up to operate on larger data types (like FP64).
FUCK THE MAN!pfft, everyone knows the true names are gt300 and r870, no matter what nvidia and wavey say.
Option 5
I only post in the gt300 thread about gt300 related topics
100%
Come on guys three days left, I know you can do it. Yes we can!
Boy George is out of jail?
No Fermi B1 respin?
http://hardforum.com/showthread.php?p=1035493002#post1035493002
If true, Nvidia would be giving up on the high end for this generation because their A3 silicon isn't remotely price/performance competitive with Cypress.
for 5850 prices.
Maybe Charlie was right about Nvidia coming out with their GF108 derivative next. Makes sense as the OEMs hungrily eye and are seduced by all those sexy new ATI features which they can plaster all over their boxes, computers and ads.
http://hardforum.com/showthread.php?p=1035493002#post1035493002
If true, Nvidia would be giving up on the high end for this generation because their A3 silicon isn't remotely price/performance competitive with Cypress.
for 5850 prices. Maybe Charlie was right about Nvidia coming out with their GF108 derivative next. Makes sense as the OEMs hungrily eye and are seduced by all those sexy new ATI features which they can plaster all over their boxes, computers and ads.
Where does it say it has a marginal lead in Dirt 2? For all we know it could be slower.
That said, I don't think you really can extend the "this game prefers AMD over NVIDIA" theme necessarily to GF100. There's probably way more to this, but 2 obvious reasons why a game would run better in theory on radeons compared to geforces relative to other games are 1) the shader code is not too "serially scalar" and 2) the alu:tex ratio is rather high. But GF100 has vastly increased alu:tex ratio hence 2) should no longer be true.
aaronspink
Veteran
FYI, paired double support required the multiplier for each 32b lane to be >2x the size as it would if it only supported SP or DP over 4 lanes. So each fermi lane likely has a 53bx27b multiplier array to support 2 lane ganged DP operation.
aaronspink
Veteran
Do people even read the linked articles? The reason the number looks good in the Nvidia material is because it was running the DX9 path. AKA, it is none comparable but nvidia decided to compare it anyways.
If that graph is accurate it does not bode well for nVidia.
Similar threads
- Deleted member 2197
- Architecture and Products
