Intel "Tiger Lake" Based Pentium and Celeron to Feature AVX2, an Instruction the Entry-Level Brands were Deprived Of

[efikkan]

I am honestly not sure if I use the benefits of AVX2 -- I always thought that AVX always referred to 1.0 but your post made me do a bit of research, and now I can't really tell if it automatically uses 2. When I had Skylake-X and had AVX-512 it was NEVER used (which was a shame) so I had lumped AVX 2.0 into that category as well.

Most of the mentioned software probably uses AVX2.

"AVX" can be confusing, since it can refer to AVX(1) or the whole family (AVX(1), AVX2, AVX-512).
In a way you can view AVX2 as the first full 256-bit SIMD instruction set, while AVX(1) was mostly a partial 256-bit extension of SSE4, but with a new syntax. AVX2 added more operations and flexibility over AVX(1), and added full 256-bit integer support. From a software development perspective, there isn't much use in using AVX(1), since AVX2 is more flexible and complete, only at the expense of dropping support for Sandy Bridge and Ivy Bridge (SB didn't have any good AVX performance anyway). By using AVX2 you usually get the bonus feature of FMA too, which really can accelerate some algorithms.

AVX-512 is unfortunately (to my knowledge) not yet used in consumer software. This is the chicken and the egg problem, as usual. But it's important to get these features widespread so software can start to utilize it. AVX-512 will be massively powerful when we see real applications use it, there will be no going back.

 

[biffzinker]

AVX-512 will be massively powerful when we see real applications use it, there will be no going back.

Except there's a hit to clockspeed, and thermals when AVX-512 is in use. Everybody loves their 5GHz overclock on the Intel side which ends up impossible with AVX-512.

 

[GoldenX]

Except there's a hit to clockspeed, and thermals when AVX-512 is in use. Everybody loves their 5GHz overclock on the Intel side which ends up impossible with AVX-512.

Even with that, the performance benefit will offset the lower clocks.
FMA alone can produce some very nice 40% uplifts, for example.

 

[windwhirl]

Except there's a hit to clockspeed, and thermals when AVX-512 is in use. Everybody loves their 5GHz overclock on the Intel side which ends up impossible with AVX-512.

I think Intel' engineers will find some way to solve that problem without losing any or very little of the speedups that AVX-512 may provide. Though it may take a while.

 

[GoldenX]

Fun that this took a decade. Looks like Intel is angry at the R3 3100 and any incoming Zen3 replacement.

 

[ratirt]

How come?
AMD have in some ways been better at adopting AVX2, featuring it even in their entry level products, and have supported it since Excavator(2015).
If anything, AVX-512 should be Intel's stronghold, if only they featured that across the entire lineup.

What I mean is Intel doesn't have much to play with to counter AMD so AVX is the one to go. In that Area the performance is not bad so Intel is adding it to every processor available or at least most.

 

[R-T-B]

Just a question - why PFSense (built on BSD) in this day and age when there is IP Fire built on Linux (SMP support) with comparable features? Granted, there's no ARM64 version but that is more or less the only drawback in my book.

There are certainly ARM64 linux versions.

 

[efikkan]

I think Intel' engineers will find some way to solve that problem without losing any or very little of the speedups that AVX-512 may provide. Though it may take a while.

They already have, in Ice Lake-SP.

 

[windwhirl]

They already have, in Ice Lake-SP.

Yeah, I forgot about it...

 

[GoldenX]

Reminder that AVX-512 is a complete mess:

 

[ExcuseMeWtf]

okay, unless there will be a consistent set implemented across the board in several generations at minimum, it won't be adopted in any remotely widespread manner.

 

[windwhirl]

Reminder that AVX-512 is a complete mess:

View attachment 172581

Yeah, but it seems like Ice Lake and Tiger Lake are gonna bring a bit of order... That aside, I think some of AVX-512 instructions are in reality "512" versions of instructions that were previously implemented in SSE/AVX/AVX2 instructions sets, so if you really don't need it, you could probably use the older versions.

okay, unless there will be a consistent set implemented across the board in several generations at minimum, it won't be adopted in any remotely widespread manner.

Again, IC and TL seem to bring a little order here, implementing a rather respectable amount of AVX-512 sets.

And just for the record, your username made this very hilarious

 

[efikkan]

I think some of AVX-512 instructions are in reality "512" versions of instructions that were previously implemented in SSE/AVX/AVX2 instructions sets, so if you really don't need it, you could probably use the older versions.

AVX-512 certainly extends the feature set of AVX2 to 512-bit, and existing code can be ported very easily.
But on the instruction level it also changes the opcodes to allow much more advanced operations on data sets, which is where the true power of AVX-512 is, beyond just being a "double AVX2". AVX-512 is getting close to being a "sub instruction set" of x86.

The challenge of all versions of AVX is the difficulty of using it, it requires expert level programmers to gain substantial performance gains. But the good news is that just enabling automatic optimizations usually gives ~10-30% performance gains "for free" (probably >50% with some minor effort), since the compiler can auto-vectorize and unroll some things, but in order to get that >10x performance gain, it still requires handcrafted low-level code. I believe compilers have some potential to improve here, but ultimately they can only deal with the code written by the programmer.

Again, IC and TL seem to bring a little order here, implementing a rather respectable amount of AVX-512 sets.

Some of these feature sets are mostly relevant to enterprise users, like those "AI" features.
The good thing about having feature sets is that it makes it easier for e.g. AMD to implement the relevant features for consumers.

 

[dragontamer5788]

AVX-512 certainly extends the feature set of AVX2 to 512-bit, and existing code can be ported very easily.
But on the instruction level it also changes the opcodes to allow much more advanced operations on data sets, which is where the true power of AVX-512 is, beyond just being a "double AVX2". AVX-512 is getting close to being a "sub instruction set" of x86.

The challenge of all versions of AVX is the difficulty of using it, it requires expert level programmers to gain substantial performance gains. But the good news is that just enabling automatic optimizations usually gives ~10-30% performance gains "for free" (probably >50% with some minor effort), since the compiler can auto-vectorize and unroll some things, but in order to get that >10x performance gain, it still requires handcrafted low-level code. I believe compilers have some potential to improve here, but ultimately they can only deal with the code written by the programmer.

ISPC and OpenCL target AVX512, and are probably what I'd recommend to anyone who is writing any serious amount of CPU-SIMD code these days. That is, if you need SIMD, but for some reason GPU-SIMD is too high latency or something, so you need to fall back to CPU-SIMD.

Intrinsics are still useful for a few applications, but its far harder to use intrinsics than to use a dedicated language like ISPC: https://ispc.github.io/

If that's still too much to ask for, then "#pragma omp simd" is the next recommendation. Works in C, C++, and Fortran on a variety of compilers (like GCC and LLVM). A shame about Microsoft Visual Studio... you can't win them all.

 

[MrMilli]

They needed AVX2? I thought they only needed plain AVX? (Missing on the Phenom II's pre-bulldozer, lots of people didn't upgrade to Bulldozer because it was worse from those, until a few titles in their DRM used AVX...)

That's a bit irrelevant isn't it, considering even the newest Pentium's and Celeron's don't support AVX1.

Reminder that AVX-512 is a complete mess:

View attachment 172581

That only looks like a mess because it's full of obsolete things. You can remove Knights Landing, Knights Mill and Cannon Lake.
Skylake-SP, Cascade Lake & Cooper Lake are all niche workstation and 8-way multiprocessing server products. Not really relevant for desktop software.
That leaves Skylake-X, Ice Lake and Tiger Lake. Again Skylake-X was a niche product.
Whatever's that left for actual desktop usage, is not a mess.

Edit:
I forgot to add that in any case all CPU's support AVX-512F (AVX-512 Foundation). If you program for AVX-512, you can always rely on AVX-512F instructions and check for more (required anyway because of fall-back code).