Intel ‘Skylake’ processors for PCs will not support AVX-512 instructions

[patrickjp93]

For all the people complaining, get over it. Intel releases what the market demands, and the reality is most of the consumer market doesn't need AVX 512, and most software barely uses SSE and not in the proportions it should be used. Intel is not going to force the issue. Software has to catch up. Moving to Windows 10 and abandoning any and all support for the Core 2 chips would be a very welcome change and good for the software industry.

 

It's the exact same reason you don't see more cores on the mainstream platform. Intel has profit margins to meet and the demand and usage just isn't there for it to be worth increasing the mainstream chip die sizes. Intel has not been sitting on its tail end since Sandy Bridge despite all the bellyaching and claims to the contrary. If you are a coder, you can prove this to yourself by running the same source code on various machines and normalizing clock rates. Legacy support is currently the name of the software game, and because of that it barely evolves, regardless of Intel's and AMD's efforts.

[vm'N]

For all the people complaining, get over it. Intel releases what the market demands, and the reality is most of the consumer market doesn't need AVX 512, and most software barely uses SSE and not in the proportions it should. Intel is not going to force the issue. Software has to catch up. Moving to Windows 10 and abandoning any and all support for the Core 2 chips would be a very welcome change and good for the software industry.

 

It's the exact same reason you don't see more cores on the mainstream platform. Intel has profit margins to meet and the demand and usage just isn't there for it to be worth increasing the mainstream chip die sizes. Intel has not been sitting on its tail end since Sandy Bridge despite all the bellyaching and claims to the contrary. If you are a coder, you can prove this to yourself by running the same source code on various machines and normalizing clock rates. Legacy support is currently the name of the software game, and because of that it barely evolves, regardless of Intel's and AMD's efforts.

I believe that 2-8 cores is the perfect spot for the consumer market (ranging from low-end to high-end). The fact is, most consumers dont require that much of a computer.

The fact is also, that less and less of the die will be the CPU itself. When we look past 10nm, the integrated GPU will consumer much more die area than the CPU.

If the demand was there, Intel would have been there too.

[Opcode]

AMD are planning for AVX 512 with Zen, supposedly.

 

http://www.kitguru.net/components/cpu/anton-shilov/amds-zen-could-sport-intels-skylake-technologies-new-512-bit-fpu-rumours/

Can confirm Zen will not support AVX-512 either.

[Fulgrim]

Can confirm Zen will not support AVX-512 either.

Give a source then.

[Arct1c0n]

Soo glad I didn't wait for Skylake and just built myself a 4790k rig last month and retired my 2500k. 

 

 

[Misanthrope]

The instruction set would be nice, but can we at least bump the line up for 4 6 and 8 cores for i3 i5 and i7 respectively?

[Victorious Secret]

Can confirm Zen will not support AVX-512 either.

 

What is your source? You don't just say "can confirm" and not give a proper source. Any other site? White paper? Marketing materials from the employee server? Anything? 

[Opcode]

Give a source then.

Check the latest binutils commits. AMD also dropped the XOP instructions from Zen.

[Colonel_Gerdauf]

Check the latest binutils commits. AMD also dropped the XOP instructions from Zen.

 

Can you provide some citations please? URLs are all we ask.

[Victorious Secret]

Can you provide some citations please? URLs are all we ask.

 

Agreed. None of this "go search for it on your own". Thats not how you backup your points. You provide evidence. 

 

Tangent, but I'm just getting tired of all the BS that gets thrown around with no evidence and "go look for it yourself". No. Not how debate works. You provide factual evidence to your positions or you don't have one. 

[Mooshi]

How much does this affect the average PC user who plays vidya, browses the internet and do minor-moderate Photoshop/editing work?

[Faa]

How much does this affect the average PC user who plays vidya, browses the internet and do minor-moderate Photoshop/editing work?

Nothing hence why it's stating consumer grade cpu's won't have avx512. Just like others said, lots of software aren't even using avx at all.

[Misanthrope]

How much does this affect the average PC user who plays vidya, browses the internet and do minor-moderate Photoshop/editing work?

By that logic everybody would still be using pentium 3s to browse geocities websites with animated flames

[Doobeedoo]

Just bring 8-core to price of 4-core, Intel.

[patrickjp93]

Agreed. None of this "go search for it on your own". Thats not how you backup your points. You provide evidence. 

 

Tangent, but I'm just getting tired of all the BS that gets thrown around with no evidence and "go look for it yourself". No. Not how debate works. You provide factual evidence to your positions or you don't have one. 

 

Can you provide some citations please? URLs are all we ask.

He's referring specifically to the GCC commits (compilers are updated almost 2 full years ahead of every instruction release in order to write emulation code for qEMU and such). That said, I've tried reading these things and, like many things in the Linux ecosystem, it's a nightmare if you're not already intimately familiar with a ton of other system terminology.

 

https://gcc.gnu.org/

 

I can't seem to find their original github and subversion repositories. @Opcode, do you know where they're hosting now?

[Faa]

He's referring specifically to the GCC commits (compilers are updated almost 2 full years ahead of every instruction release in order to write emulation code for qEMU and such). That said, I've tried reading these things and, like many things in the Linux ecosystem, it's a nightmare if you're not already intimately familiar with a ton of other system terminology.

 

https://gcc.gnu.org/

 

I can't seem to find their original github and subversion repositories. @Opcode, do you know where they're hosting now?

Still doesn't prove Zen won't have AVX512. Thats what Opcode has been asked to do.

[patrickjp93]

Still doesn't prove Zen won't have AVX512. Thats what Opcode has been asked to do.

There's an update in their compiler repositories well-documented showing only some Zen chips (server only) will have access to the instruction. I just need to find their damn repository which I swear moves every damn month.

[vm'N]

Considering zen would most likely have a 256bit FPU, i always doubted zen would feature AVX512 support (unless it was like bulldozers AVX support)."

[patrickjp93]

Considering zen would most likely have a 256bit FPU, i always doubted zen would feature AVX512 support (unless it was like bulldozers AVX support)."

You could have 2 FPUs per core. Intel's Haswell is arranged that way where there are 2 GP ALUs, a dedicated 256-bit FPU, and a 256-bit AVX ALU.

[vm'N]

You could have 2 FPUs per core. Intel's Haswell is arranged that way where there are 2 GP ALUs, a dedicated 256-bit FPU, and a 256-bit AVX ALU.

It would be dumb having two identical generic 256bit FPUs.

You will waste more ressources than having a single 512bit FPU, and will gain no performance benefits.

 

Dedicated hardware for different operations is something completely different.

[Opcode]

@Fulgrim @Victorious Secret @Colonel_Gerdauf @Faa @patrickjp93

 

I added Excavator as well just for reference (since we don't know a whole lot about it still) which confirms AVX2 support for Carrizo.

   { "CPU_BDVER4_FLAGS",     "Cpu186|Cpu286|Cpu386|Cpu486|Cpu586|Cpu686|CpuSYSCALL|CpuRdtscp|Cpu387|Cpu687|CpuFISTTP|CpuNop|CpuMMX|CpuSSE|CpuSSE2|CpuSSE3|CpuSSE4a|CpuABM|CpuLM|CpuFMA|CpuFMA4|CpuXOP|CpuLWP|CpuBMI|CpuTBM|CpuF16C|CpuCX16|CpuClflush|CpuSSSE3|CpuSVME|CpuSSE4_1|CpuSSE4_2|CpuAES|CpuAVX|CpuPCLMUL|CpuLZCNT|CpuPRFCHW|CpuXsave|CpuXsaveopt|CpuFSGSBase|CpuAVX2|CpuMovbe|CpuBMI2|CpuRdRnd" },+  { "CPU_ZNVER1_FLAGS",+    "Cpu186|Cpu286|Cpu386|Cpu486|Cpu586|Cpu686|CpuSYSCALL|CpuRdtscp|Cpu387|Cpu687|CpuFISTTP|CpuNop|CpuMMX|CpuSSE|CpuSSE2|CpuSSE3|CpuSSE4a|CpuABM|CpuLM|CpuFMA|CpuFMA4|CpuBMI|CpuF16C|CpuCX16|CpuClflush|CpuSSSE3|CpuSVME|CpuSSE4_1|CpuSSE4_2|CpuAES|CpuAVX|CpuPCLMUL|CpuLZCNT|CpuPRFCHW|CpuXsave|CpuXsaveopt|CpuFSGSBase|CpuAVX2|CpuMovbe|CpuBMI2|CpuRdRnd|CpuADX|CpuRdSeed|CpuSMAP|CpuSHA|CpuXSAVEC|CpuXSAVES|CpuClflushOpt|CpuCLZERO" },

Strangely enough BTVER2 (Jaguar) is listed with AVX2 and AVX-512 (I'm guessing through FMA) support.

   { "CPU_BTVER2_FLAGS",@@ -245,6 +247,8 @@ static initializer cpu_flag_init[] =     "CpuMMX|CpuSSE|CpuSSE2|CpuSSE3|CpuSSSE3|CpuSSE4_1|CpuSSE4_2|CpuAVX|CpuAVX2|CpuAVX512F|CpuAVX512IFMA" },

[LAwLz]

Wow... the old "we don't need more" argument. People have been using it since the days of 500Mhz single core CPUs, and we were very lucky that nobody actually believed that horse shit back then.

This thread is pretty funny. The people upset would probably not even have noticed it unless someone told them, and the people defending Intel would probably have been praising them for being forward thinking if they did release with it.

 

Bottom line is, Skylake without AVX512 is objectively inferior to what Skylake would have been if it supported it. How much worse it is varies, but it is worse nonetheless.

[Mooshi]

By that logic everybody would still be using pentium 3s to browse geocities websites with animated flames

 

Except the typical user isn't browsing geocities to stare at animated flames, not on this forum. Come on, that wasn't an accurate comparison, we're not talking mothers who can barely use the internet. We're talking those who use Steam and do other things like Blender, Maya, Photoshop ect. Not quite doing workloads that require a Xeon, but typical things many were already doing on their PC right now. I was asking how much of a real impact lacking this instruction set will have on essentially everyone today that's not using a crazy workstation powered by a Xeon.

 

Edit: Even if it's not really utilized, I'm not excusing AMD or Intel for dropping it on the consumer line. Especially for Intel, their gear has enough power on the mainstream platform to be used by prosumers and even professionals on a budget.

[Misanthrope]

Except the typical user isn't browsing geocities to stare at animated flames, not on this forum. Come on, that wasn't an accurate comparison, we're not talking mothers who can barely use the internet. We're talking those who use Steam and do other things like Blender, Maya, Photoshop ect. Not quite doing workloads that require a Xeon, but typical things many were already doing on their PC right now. I was asking how much of a real impact lacking this instruction set will have on essentially everyone today that's not using a crazy workstation powered by a Xeon.

 

Edit: Even if it's not really utilized, I'm not excusing AMD or Intel for dropping it on the consumer line. Especially for Intel, their gear has enough power on the mainstream platform to be used by prosumers and even professionals on a budget.

 

No it was an exaggeration to illustrate a point: without the increased hardware capabilities we wouldn't have the increased resource usage we need for modern websites nowadays. The way it is right now a lot of programs run the risk of getting stagnated since its difficult to move developers to multithreaded programming. If we have intel not introducing instruction sets and limiting the number of threads on their processors adoption of this technologies will take even longer.

 

Yes only very specific power users would get any advantage out of it right now but if the tech was out there more devs would be encouraged to take advantage of it.

[patrickjp93]

It would be dumb having two identical generic 256bit FPUs.

You will waste more ressources than having a single 512bit FPU, and will gain no performance benefits.

Dedicated hardware for different operations is something completely different.

Not true. The larger the FPU the more cycles you add to its operations. Two smaller 256-bit is actually more efficient. That's Power 8 have 4 dedicated 256-bit FPUs per core in addition to 4 256-bit integer ALUs.