Take the 2-minute tour ×
Super User is a question and answer site for computer enthusiasts and power users. It's 100% free, no registration required.

Top GeForce video cards have 480 GUDA cores (60 multicores by 8 stream each). They are like general purpose CPUs, but with specialized instructions, registers and memory working modes, and are oriented to float-point processing.

Top Radeon video cards usually have less speed than GeForce cards (at two times less price!). However, the official specs state that they have 1408 stream processors and 22-24 SIMD engines (by 16 streams each).

What are stream processors, and what does SIMD engine mean? Is a SIMD engine like CUDA cores? Are stream processors exactly the same as ALUs?

share|improve this question
    
SIMD stands for "Single Instruction, Multiple Data". –  ChrisF Sep 3 '11 at 21:07
    
@ChrisF OK, I already know this )) –  gavenkoa Sep 4 '11 at 7:14
    
I didn't :) I thought it might help others too. –  ChrisF Sep 4 '11 at 12:42

1 Answer 1

up vote 5 down vote accepted

From some AMD presentation:

Each SIMD unit contains 16 stream cores, and each stream core houses five processing elements. Thus, each compute unit in the ATI Radeon HD 5870 has 80 (16 × 5) processing elements.

Analogue for CUDA core is Stream core, and SIMD unit is analogue to nVidia's Streaming Multiprocessor.

The difference between CUDA core and Stream core is that CUDA core consists of one processor, though is capable of performing only one operation at a time, and Stream core has 5 processors and can simultaneously perform up to 4 'simple' operations (like addition/multiplication), but only for vector types (like int4, float4). For scalar types 4 of this 5 processors wait while one does all the work.

The main difference in performance is not due to number of cores, but rather due to how this cores are made. nVidia's GPUs are better for floating point operations (so they are used in most scientific applications), when ATI's GPUs are unsurpassed in field of integer and bitwise operations, and are used for cryptography.

share|improve this answer
    
Great answer! Welcome to Super User :D –  Kyle Sep 6 '11 at 19:57
    
@aland thanks for answer. In specs written that CUDA perform all operation (float/integer) in parallel on single multicore (as 8 subcores you have 8 operation). So CUDA more effectively use integer operations (as it have more integer ALU then ATI as ATI have **CORE_CNT**/5)? –  gavenkoa Sep 7 '11 at 16:05
    
@gavenkoa Multicore is the same as streaming multiprocessor (== SIMD unit). SM has 8 cores in older devices and 32 cores in geforce 4xx and newer. The number of cores that are usually advertised is actually amount of processors (480 for GTX480) and not amount of SM (15 for GTX480). AMD advertisements also just tell number of processors (1600 for HD5870), though for scalar operations max. 20% of them are used (in CUDA there is no such tricks) –  aland Sep 7 '11 at 17:20

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Not the answer you're looking for? Browse other questions tagged or ask your own question.