Does anyone know what a half gflop is?

A forum for discussing FAH-related hardware choices and info on actual products (not speculation).

Moderator: Site Moderators

Forum rules
Please read the forum rules before posting.
Post Reply
Green
Posts: 41
Joined: Wed Oct 22, 2014 6:46 pm
Location: Lexington, KY 40503
Contact:

Does anyone know what a half gflop is?

Post by Green »

A half gflop if you wiki "Amd gpu" ya digg, how you find all the speeds of amd and nvidia cards, on their wiki pages under gflops. Right, so now that were all on the right track, does a RX Vega go any faster because it carries half GLFOP points as well as single GFLOP points?
Very Respectfully
Ben D.C. Bruner
toTOW
Site Moderator
Posts: 6296
Joined: Sun Dec 02, 2007 10:38 am
Location: Bordeaux, France
Contact:

Re: Does anyone know what a half gflop is?

Post by toTOW »

I think you mean half precision ? This is work on 16 bit floating point numbers (this is also known as FP16). The single precision which we mostly use on FAH works on 32 bits floating point numbers (FP32). Some scientific computations use double precision (64 bits - called FP64) to avoid numeric instabilities introduced by rounding. FAH sometimes use mixed precision, which means that most of the simulation is using single precision, but some parts require double precision.

You can learn more about these formats here :
https://en.wikipedia.org/wiki/Half-prec ... int_format
https://en.wikipedia.org/wiki/Single-pr ... int_format
https://en.wikipedia.org/wiki/Double-pr ... int_format
Image

Folding@Home beta tester since 2002. Folding Forum moderator since July 2008.
bruce
Posts: 20910
Joined: Thu Nov 29, 2007 10:13 pm
Location: So. Cal.

Re: Does anyone know what a half gflop is?

Post by bruce »

There is a long history of single-precision floating point operations which you can read elsewhere. In the early days of home computing, say around 1960s, the 8086 chip did not include the capabilities of performing floating point operations (flops) but if you needed to do scientific computations, you had to buy a "math coprocessor" (also identified as a 8087) which could be plugged into a second socket and it could do FP32 operations, though at a much slower clock rate than the basic integer operations of the 8086. After several generations of chip design, the two were integrated on to the same silicon and a single socket was provided.

FP32 is accurate enough for MOST scientific calculations although some require Double-precision, hence FAH's use of Mixed precision. Half precision has been having a heyday because it can be used in things like 3D video games where the slight errors introduced aren't important. Since FP16 uses less hardware, those operations can run faster, improving the experience of your video game.
JimboPalmer
Posts: 2573
Joined: Mon Feb 16, 2009 4:12 am
Location: Greenwood MS USA

Re: Does anyone know what a half gflop is?

Post by JimboPalmer »

Since Half Precision is 'easier', the GPU can do more of them in a fixed period of time.

It is possible you saw stats showing that Full Precision did 1/2 the FLOPS of Half Precision.

(Some GPUs take a huge drop going to Double Precision, as much as 1/24th the throughput. It is widely believed this is deliberate, to prevent scientific use of gaming cards)

https://en.wikipedia.org/wiki/List_of_A ... ega_Series
shows some Single/Full/Double throughput in GFLOPS for recent AMD GPUs.
Tsar of all the Rushers
I tried to remain childlike, all I achieved was childish.
A friend to those who want no friends
Post Reply