No buzz about GTX 260?

[taltamir]

Originally posted by: hooflung
The 260 will be an ideal solution when it is 249 before mail in rebates. 2 260's in SLI just aren't that spectacular especially when you could do crossfire on 2 4870's and have much better value.

However, if people start releasing CUDA encoders for media authoring then that is a place where the 260 will shine.

funny you said that. I ordered an eVGA GTX260 SC for 255$, free shipping, and a 30$ MIR (255-30$MIR, not 255 after 30$ MIR).
IF I get the MIR honored then I can expect it to cost 225$ total.


As for "trading blows in min frame rate"... the GTX wins 9 / 11 games, the 4870 wins 2 / 11. Not exactly trading blows, and the GTX wins in min FPS on games that the 4870 wins in average... actually the average FPS is about backwards, with 4870 winning most and the GTX260 winning a few.

 

[Cookie Monster]

Its interesting since the HD4870 can produce a higher max fps i.e producing a higher avg fps compared to the GTX260.

 

[taltamir]

its due to the nature of their SP...

the GTX has single FPU SP.. the AMD cards have 5 FPU SP, so depending on instruction level parallelism then can be using from 1 to 5 of the FPU in each SP... so overall their performance fluctuates more then the nvidia cards. with lower min and higher max per SP. (and with greater per game variation, since each game has different level of scaling within each SP)

 

[Cookie Monster]

Um what is a FPU SP?

An SP or streaming processor is a misleading term in some sense (its more to do with the marketing i guess) but its better to call them scalar ALUs for G200. A G200 has these 8 of these scalar ALUs grouped into a "streaming multiprocessor" that are capable of doing FP32/INT32 operations. (And 64bit FP operations too but at a abysmal 1/8 of the speed!).

Now i dont get what you mean by "instruction level parallelism" but im thinking that you were trying to point out that it takes more work for the scheduler/compiler found in the RV770 GPU to take advantage of its 800 ALUs. (Since these are grouped into 5, i.e sometimes are referred to as vec5 shaders i.e suffer inherent drawbacks associated with such setup). I agree with that, and its one of the drawbacks (when considering gaming) of the architecture currently used by AMD/ATi. But if you consider heavy arithmetic or computational workloads than the RV770 really stretches its legs. When considering FP64 operations, RV770 has 160 FP64 ALUs while G200 has only 30 in layman terms. However, ive always thought that it was also due to the lack of TMUs, probably a combination of both. Its an interesting matter to investigate upon.

 

[taltamir]

ALU = arithmatic logic unit. FPU = floating point unit (like ALU for floating point numbers instead of integers, I think you can call an FPU an ALU since it is a type of ALU AFAIK).
1 FPU SP means that each of the 320 or whatever SP has one FPU unit in it.
5 FPU SP means each one of the 160 has 5 FPU units in it for a total of 800... which are actually 4 single precious FPU and 1 double precious FPU if I am not mistaken.
However, there are also the texture units and other types of dedicated processors.

And the instruction level parallelism is straight from the anandtech's article.
http://www.anandtech.com/video/showdoc.aspx?i=3341&p=3

The thread is first broken up into instructions that can be run in parallel, both AMD and nVidia need to do this. Then sent to each SP. weather that SP is part of an array or not is irrelevant, there is no overhead there.
The actual instruction is given to the SP, nvidia performs them one by one. AMD SP has 5 units in it so it can performs from 1 to 5 calculations per clock per SP.

 

[Cookie Monster]

I think your confusing yourself. A single scalar ALU (for instance found in the G200) can perform either a FP32/INT32 operations (they are supported to do both) without the need to have dedicated FPUs. Calling them FPUs is misleading (since those "FPUs" can perform integer operations!) and i think this is the reason why i was confused in the first place.

How do you think the thread is first broken up into instructions? . Actually what you've said is perfectly fine except thats what i was saying before! hehe. Ive already mentioned about the thread scheduler/compiler overhead (which also determines utilization rates?) in the previous post which is what i think has some sort of influence on the low mid fps compared to its competition. Also as you said, those ALUs that are part of an array are only irrelevant if they are working on the same thread.

 

[taltamir]

http://en.wikipedia.org/wiki/Arithmetic_Logic_Unit

ALUs vs. FPUs

A Floating Point Unit also performs arithmetic operations between two values, but they do so for numbers in floating point representation, which is much more complicated than the two's complement representation used in a typical ALU. In order to do these calculations, a FPU has several complex circuits built-in, including some internal ALUs.

Usually engineers call an ALU the circuit that performs arithmetic operations in integer formats (like two's complement and BCD), while the circuits that calculate on more complex formats like floating point, complex numbers, etc. usually receive a more illustrious name.

Wiki HAS been wrong before though...

 

[Cookie Monster]

Hmm, but arithmetic operations =/= integer operations. So the last statement is what i was trying say. ALUs found both in the G200/RV770 supports both integer and floating point operations which are all arithmetic operations (it supports both formats). Where as saying its an FPU means that it only does arithmetic operations in the FP format.

 

[fierydemise]

Originally posted by: taltamir
As for "trading blows in min frame rate"... the GTX wins 9 / 11 games, the 4870 wins 2 / 11. Not exactly trading blows, and the GTX wins in min FPS on games that the 4870 wins in average... actually the average FPS is about backwards, with 4870 winning most and the GTX260 winning a few.

You keep talking about the GTX wins 9 and the HD4870 wins 2 however what you are ignoring (missing?) is the margin of victory +/- a few FPS (discounting Lost Planet where the HD4870 gets destroyed). Such minor variations are well within the margin of error of the "real world testing" that driverheaven does.