Question on AVX2 and FMA (Haswell)

[phillyman36]

Ok I tried my best to find the answer myself thru wiki, youtube, google searches and still dont know. Can someone tell me/explain
1) What is FMA is and what it does?(examples)
2) What AVX2 is and does? (examples)

I seem to find technical answers that I don't understand. In simple terms what do they help ie video encoding, gaming etc etc. How do they effect everyday computing? What software would benefit from these in Haswell (ie handbrake, photoshop)
Building a pc is simple. Understanding the the thing you all talk about and debate is a whole other level
Thanks for any help.

 

[Azuma Hazuki]

FMA - Fused Multiply-Add. A single-instruction multiply-add, which saves much heat and processor time.

AVX2- Advanced Vector eXtensions, ver. 2. Something like SIMD (SSE) for integer operations; massively speeds up integer computation where supported, as the first AVX did for floating-point operations.

 

[john3850]

AVX2 gets released with Haswell and extends the avx
http://software.intel.com/en-us/art...rformance-improvements-and-energy-efficiency/

Both prime and ibt use the avx ext.
The software will take a few years to catch up.

 

[cytg111]

1. this op A = A * B + C, can be done much faster

2. look at the code here and think integer instead of floating
http://software.intel.com/en-us/articles/using-avx-without-writing-avx-code

 

[NTMBK]

And to make that a little clearer...

SIMD - single instruction, multiple data. This lets the processor do sums on multiple numbers at once, instead of just one. SSE lets it operate on 4 numbers at once, and AVX lets it operate on 8 numbers at once. (AVX2 adds support for integers, whereas AVX1 was just floating point [decimal numbers]).

FMA - this lets the processor work out A = (B*C)+D all in one calculation, instead of doing the multiply and the addition separately.

 

[Idontcare]

This is a great question, and a great thread!

Thanks OP for posting the question, and thanks contributors for taking time to respond Good stuff all around :thumbsup:

 

[Pilum]

SIMD - single instruction, multiple data. This lets the processor do sums on multiple numbers at once, instead of just one. SSE lets it operate on 4 numbers at once, and AVX lets it operate on 8 numbers at once. (AVX2 adds support for integers, whereas AVX1 was just floating point [decimal numbers]).

To enhance on AVX2: when you can get away with 16-bit or 8-bit integer calculations (image&video processing), AVX2 can work on 16 or 32 elements at once.

E.g. for image scaling you can usually use 16-bit integers with fixed-point math instead of floating point. With the two AVX pipelines Haswell can do 32 16-bit operations per core per cycle, that is 128 ops/cycle for a 4-core HSW. In contrast, GPUs can't split their FP execution units like that, if you work with 16-bit integers, you basically waste half of the processing power. This will make Haswell rather strong compared to GPGPU for appropriate workloads, once the software has been adopted. And of course porting from SSE2+ to AVX2 is rather easy, as you basically only need to adopt the code to the wider instructions and change the pointer adjustments; this is much easier than designing and writing whole new OpenCL/CUDA routines.

 

[Phynaz]

To make it a little more clear:

AVX2 is the (marketing) name given to a set of software instructions added to Haswell processors that do what other posters have pointed out.

 

[cytg111]

To make that even a little more clear :

 

[guskline]

Excellent question and thread. I actually learned something today!

 

[Greenlepricon]

Thanks everyone! Maybe we should just make a thread for all of these crazy acronyms that I don't understand.

 

[BTRY B 529th FA BN]

Gettin my armchair education here! Thanks!

 

[SocketF]

AVX2 gets released with Haswell and extends the avx with to 256 bit

No, AVX can handle 256bit already, what AVX2 does is delivering more INT instructions (and some other weired ones (gather/scatter). AVX so far had only FP-instructions.

In other words: AVX2 is a bit more than AMD's XOP-instructions (which were limited to 128bit).

AMD planned SSE5 long ago with INT&FP SIMD 3operand-instructions plus FMA instructions from the start. Intel than added 256bit support, invented the VEX-prefix used only FP and named it AVX.

Because AVX is only a SSE5 subset but with 256bit, AMD used it, too, and the leftovers of SSE5 (the INT instructions) were called XOP.

Now with AVX2 you get INT 3op-operations, too. Haswell also has FMA, but that extension has its own CPUID-flag, i.e. it shouldnt be counted as AVX2.

For example, one of the next VIA CPUs (yes they are still alive) will support AVX2, but not FMA.