Hacker News Reader

Comment by Narann | Hacker News Reader

The real quote would have been:

> Do CPU designers spend area on niche operations such as _binary-field_ multiplication? Sometimes, yes, but not much area. Given how CPUs are actually used, CPU designers see vastly more benefit to spending area on, e.g., vectorized floating-point multipliers.

So, CPUs are not "optimized for video games", they are optimized for "vectorized floating-point multipliers". Something video game (and many others) benefits from.

View on HN

Why are they optimized for vectorized floating-point multipliers? Does the CEO of Intel just tell all the engineers to do this because he likes multiplication?

They are optimized for that because a lot of algorithms can make use of them, from quicksort/mergesort through image rendering and encryption. It is an easy optimization from a hardware perspective -- simple repetitive hardware structure. This is why GPUs are so powerful and games are not the only thing that benefits from this type of optimization. Matrix multiplication is also used in signal processing. The CEO asked, how can we optimize the use of our hardware for the most benefit? And SIMD with wide pipes is at the top of the list. Most of the post is about all the new algorithms that can take advantage of the hardware push. The hardware push is there because it is an easy use of hardware resources.

This is also an optimization that compilers can readily take advantage of on a small scale (similar to pipelining) so the combination of benefit + ability to use + simplicity/low resource use makes it an inevitability.

A sort that uses vectorized floating point multiplication?

From 2008: http://www.vldb.org/pvldb/1/1454171.pdf

Also: http://researcher.watson.ibm.com/researcher/files/jp-INOUEHR... and https://github.com/NumScale/boost.simd

Because CPUs are designed around benchmarks representative of real workloads people are running on their computers. Naturally, multimedia and games are a large part of these workloads, this is what drives SIMD adoption.

No, it's done because it covers two of their largest markets Gaming machines and clusters used for scientific / finance applications.

I'd say modern CPUs are not optimized for anything in particular. They are simply okay at doing general purpose stuff.

Anything more specialized exists in the form of DSPs GPUs, etc..

That's ignoring the history. The definition of 'general purpose stuff' includes things which used to be considered specialized. FPUs used to be physically separate optional add-ons. They were quickly absorbed into the main CPU around the time that the spreadsheet became the killer app. Then SIMD was added when 'multimedia' became a requirement. Now pretty good GPUs are integrated too. Of course all these capabilities are multi-purpose, but they were not core functionality originally.

I agree. Nontheless, modern cpus are nonetheless designed to be acceptably good at different things (versus being extremely efficient at solving a specific task).

Really spreadsheets used floating point? That sounds like terrible idea.

That's a good point. I don't know the answer. Optional PC FPUs and the rise of spreadsheets were contemporaneous in the late 1980s but perhaps not related.

edit: I checked. Microsoft at least used floating point in Excel until much later, until at least 2013. With the expected limitations. https://support.microsoft.com/en-ca/kb/78113

They absolutely were related (as were databases and FPUs). Had these mainstream business applications not driven them, it would have likely been years later before they became standard equipment. The FPU was likely originally added with 'serious' scientific and engineering computing tasks (including CAD) in mind, but it was the more mundane and common spreadsheet and database applications that drove demand.

Here's what I remember from back then: early PCs (8-bit era: Apple // etc) had no socket for a FPU (i.e. not even an expansion option), every IBM-compatible PC had a co-processor socket starting with the 8086 (fun fact: the 8087 wasn't even shipping when the PC was designed, and boy were they expensive for what they did when they shipped) but almost no one bought one until the late-286/early-386 era for general-purpose computing, by the 386 era the FPU was pretty much standard equipment on any 'real' business PC. So the software evolution was: no FPU support, optional FPU support, FPU required. (i.e. by the last generation of DOS applications, several major apps had dropped their floating point emulation libraries and would crap out with an error along the lines of 'coprocessor not detected/installed')

The evolution of the FPU was very similar to how the GPU has played out in terms of becoming a standard, expected component. As with the FPU, CAD and all sorts of other scientific/business applications were the initial drivers of the, but gaming is what caused unit volumes to explode and the reason it is now standard equipment.

> They were quickly absorbed into the main CPU around the time that the spreadsheet became the killer app.

For very loose definitions of "quickly" and/or "around the time". Spreadsheets became the "killer app" with VisiCalc in 1979 -- before the IBM PC was even a thing; but Intel processors through the 386 didn't have an integrated FPU, and the 486 (around 1990) came in both integrated-FPU (486DX) and no-integrated-FPU (486SX) models. It wasn't until the Pentium that integrated-FPU was universal in the Intel line.

(AFAIK, most spreadsheets of the early PC era didn't even support using the FPU, and the main widely-used application category that leveraged FPUs in the optional-FPU era was CAD.)

Yep, my spreadsheet example was not great. Your example of FPU-for-CAD is perhaps better at making the case that the FPU was once a specialized co-processor, but of course is now considered an essential everyday component. The modern CPU evolved to support the set of workloads people found for them over time. They are now heterogeneous parallel systems that are really good at actual workstation and server workloads. They are quite specialized in that sense.

The FPU achieving mass adoption maps pretty well to the onset of "multimedia hype" (CDs, digital audio and video, and of course 3D game rendering). That content doesn't strictly need floating point, but it became an affordance along with 32-bit desktop architectures and their increased memory and storage.

They're optimized for single-precision linear algebra. If you need double-precision all those optimizations go out the window and you're left on your own.

> They're optimized for single-precision linear algebra. If you need double-precision all those optimizations go out the window and you're left on your own.

Not true at all on modern X86. You just get half the FLOPS by using double precision instead of single precision -- 16 double precision FLOPS per core per cycle. Which is very good considering there's also twice as much data to process.

You could rather say X86 CPUs are highly optimized for double precision performance.

That's great. It's been a while since I did a lot of simulations (which used double precision) so I didn't know that had changed.

I still wouldn't say that they're highly optimized, though.

Please see http://agner.org/optimize/optimizing_cpp.pdf chapter 7.3 for detailed technical discussion on this.

"In most cases, double precision calculations take no more time than single precision."

Only for scalar calculations. Vectorized they take about double the amount of time.

Reply on news.ycombinator.com