Do GPUs have instruction sets and caches?

[Firestrider]

Simple question.

Caches meaning like L1 and L2 on a CPU and instruction sets like SSE2.

I know Larrabee will have the x86 architecture but I'm not sure exactly what that entails... will it just have the SIMD instructions onboard?

Also if a GPU has its own ISA will that mean there is no more communication between the CPU and GPU? I think the GPU will still have to store information in main memory.

 

[dogbert_2001]

No. GPUs just have registers. When the CPU writes to the register, the GPU goes off and does something. The GPU does not read instructions from a memory location.

A DMA engine in the GPU may read descriptors from memory, but it's nothing like a real instruction code.

 

[Firestrider]

It seems from here that graphics cards do have some L1/L2 cache:

http://www.amd.com/us-en/assets/con...aphicsAMD2008AnalystDayBreakout11-18-2008.pdf

 

[MadMan007]

They do in a broad sense but I wouldn't compare them directly to a CPU, it also depends upon the exact GPU about which you're talking. They definitely have caches now, you can find that in reviews where they go over the architecture in detail. ATi cards since the HD2000's have been VLIW (very long instruction word) processors where there is a software layer that packages instructions for efficient execution, in that sense there is an instruction set for them it's just buried underneath that VLIW layer. It's not a standard or open instruction set though and is interpreted through drivers or programming language interfaces like CUDA or ATi Stream.

 

[electronicsguy]

I know it's an old thread, but I just came across is today. The answers here are all wrong. Simple answer: Yes, GPUs have cache! Size, bandwidth and coherency may be different. Here's a simple article explaining it: http://supercomputingblog.com/cuda/cuda-memory-and-cache-architecture/ or http://meseec.ce.rit.edu/551-projects/spring2015/3-2.pdf

 

[wingman99]

Very interesting.