Will AVX2 ISA allow Intel to jettison X86 legacy/baggage?

Our "pub" where you can post about things completely Off Topic or about non-silent PC issues.

Moderators: NeilBlanchard, Ralf Hutter, sthayashi, Lawrence Lee

Post Reply
dan
Posts: 1243
Joined: Sun Jan 25, 2004 2:01 am
Contact:

Will AVX2 ISA allow Intel to jettison X86 legacy/baggage?

Post by dan » Mon Jan 02, 2012 5:36 pm

Haswell will come with a new instruction expansion, AVX2 for Interger (AVX was for Floating)

AVX2 will introduced 3-operand encoding which is RISC

Could Intel introduce future chips that support AVX2 and x86 in emulation?

Can software be written in clean AVX2 code?

It is reported

What’s new with Intel Haswell architecture

DDR3-1600 memory
8MB of L3
FMA (Fused Multiply Add) feature
Coding similar to AVX2 RISC code
Like AMD, Intel is planning to AVX as the basis of a future “running on an x86 emulation layer AVX2″ and dismissing the “pure x86″.
http://lenzfire.com/2011/11/intel-haswe ... ors-35797/

higher-performing "RISC" lower power, and backwards compatible to x86

NeilBlanchard
Moderator
Posts: 7681
Joined: Mon Dec 09, 2002 7:11 pm
Location: Maynard, MA, Eaarth
Contact:

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by NeilBlanchard » Tue Jan 03, 2012 3:10 am

Dan,

Thanks for posting. What is your take on this -- is it evolution or revolution? Will it affect the end user, or is this just changes "under the hood"?

Ingeneer
Posts: 2
Joined: Fri Jan 06, 2012 2:27 am

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by Ingeneer » Fri Jan 06, 2012 3:08 am

AVX2 is both evolution and revolution. It's an evolution in the sense that it's just another x86 instruction set extension. It's the successor of SSE1, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2, and AVX1.

It is revolutionary for the addition of 'gather' support and other instructions which make it feasible to run up to 8 times more instructions in parallel. Gather instructions are the vector equivalent of load instructions, and replace 18 legacy instructions! It also brings fused multiply-add (FMA) support to Intel processors (and unlike AMD it will double the peak performance per core).

Note also that both gather and FMA are GPU features. So the difference in their throughput computing capabilities is fading, while the CPU retains its high efficiency qualities for complex code...

dan
Posts: 1243
Joined: Sun Jan 25, 2004 2:01 am
Contact:

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by dan » Sat Jan 14, 2012 8:20 am

NeilBlanchard wrote:Dan,

Thanks for posting. What is your take on this -- is it evolution or revolution? Will it affect the end user, or is this just changes "under the hood"?
You know how many mobile devices use ARM Risc, b/c x86 is too hot and power hungry?

AVX2 3-operand would allow a Risc like design and potentially lower power consumption. So revolutionary.

dan
Posts: 1243
Joined: Sun Jan 25, 2004 2:01 am
Contact:

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by dan » Sat Jan 14, 2012 8:22 am

Ingeneer wrote:AVX2 is both evolution and revolution. It's an evolution in the sense that it's just another x86 instruction set extension. It's the successor of SSE1, SSE2, SSE3, SSSE3, SSE4.1, SSE4.2, and AVX1.

It is revolutionary for the addition of 'gather' support and other instructions which make it feasible to run up to 8 times more instructions in parallel. Gather instructions are the vector equivalent of load instructions, and replace 18 legacy instructions! It also brings fused multiply-add (FMA) support to Intel processors (and unlike AMD it will double the peak performance per core).

Note also that both gather and FMA are GPU features. So the difference in their throughput computing capabilities is fading, while the CPU retains its high efficiency qualities for complex code...
thanks for replying. How much of a performance increase would this result in? Is this similar in spirit to VLIW Itanium EPIC, only compatible with x86

I know that it will be 3-operand, so does this mean AVX2 will essentially be RISC

nutball
*Lifetime Patron*
Posts: 1304
Joined: Thu Apr 10, 2003 7:16 am
Location: en.gb.uk

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by nutball » Sat Jan 14, 2012 8:59 am

dan wrote:Is this similar in spirit to VLIW Itanium EPIC, only compatible with x86
Not really, no. VLIW is more about instruction encoding than anything else.
I know that it will be 3-operand, so does this mean AVX2 will essentially be RISC
Not really, no. RISC means a whole pile of things, and it's not obvious how a bunch of extra instructions added to a CISC instruction set can in any way be called reduced instruction set. RISC doesn't imply 3-operand, and 3-operand doesn't imply RISC. It's not that simple.

By the way when that article talks about emulating x86, I think what it means is emulating the really old bits of x86, the pre-x86_64 stuff. The x86_64 stuff will remain unemulated.

Lastly, the view that RISC is somehow magically clean, simple and efficient, and that x86 is dirty and inefficient in a way that really matters is a bit naive... a bit 90's to be honest. It might matter in mobile phones, but pretty much doesn't anywhere else. If you remove ARM from the equation and look at the next best-selling pure RISC chip, it can hardly be called simple, nor low-power. If you're interested there have some good ding-dongs between Linus Torvalds and advocates of the RISC=good/x86=bad mantra over on RWT, they make fun reading.

dan
Posts: 1243
Joined: Sun Jan 25, 2004 2:01 am
Contact:

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by dan » Sat Jan 14, 2012 12:44 pm

nutball wrote:
dan wrote:Is this similar in spirit to VLIW Itanium EPIC, only compatible with x86
Not really, no. VLIW is more about instruction encoding than anything else.
I know that it will be 3-operand, so does this mean AVX2 will essentially be RISC
Not really, no. RISC means a whole pile of things, and it's not obvious how a bunch of extra instructions added to a CISC instruction set can in any way be called reduced instruction set. RISC doesn't imply 3-operand, and 3-operand doesn't imply RISC. It's not that simple.

By the way when that article talks about emulating x86, I think what it means is emulating the really old bits of x86, the pre-x86_64 stuff. The x86_64 stuff will remain unemulated.

Lastly, the view that RISC is somehow magically clean, simple and efficient, and that x86 is dirty and inefficient in a way that really matters is a bit naive... a bit 90's to be honest. It might matter in mobile phones, but pretty much doesn't anywhere else. If you remove ARM from the equation and look at the next best-selling pure RISC chip, it can hardly be called simple, nor low-power. If you're interested there have some good ding-dongs between Linus Torvalds and advocates of the RISC=good/x86=bad mantra over on RWT, they make fun reading.

I'm kinda curious as to what that RISC chip is -- Powerpc? MIPS?
There's a lot of talk about Apple ditching x86 for ARM, esp laptops
How come Nintendo Sony Xbox use powerpc?

Personally I would like to see Haswell in PS4 or Xbox 720 so as to make porting easier.

washu
Posts: 571
Joined: Thu Nov 19, 2009 10:20 am
Location: Ottawa

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by washu » Sat Jan 14, 2012 3:52 pm

dan wrote: I'm kinda curious as to what that RISC chip is -- Powerpc? MIPS?
There are lots of definitions as to what makes RISC, but the one I usually go by is that the chip implements a load-store architecture. IE, the CPU must use separate instructions to move data to and from the registers and cannot operate on data in memory directly.

A reduced number of instructions doesn't really apply any more, Powerpc for example has more instructions than x86.
There's a lot of talk about Apple ditching x86 for ARM, esp laptops
I don't think that is very likely unless Apple also ditches OSX on those laptops. The 68k->Powerpc and Powerpc->x86 switches were increases in CPU power. X86->ARM would be a major decrease in CPU power. There is no way an ARM could emulate x86 programs at a reasonable speed.

Apple might bring out a "low end" ARM laptop running iOS, but I don't think they will be getting rid of their x86 computers anytime soon.
How come Nintendo Sony Xbox use powerpc?
Simple, Sony and Microsoft needed two things for game console CPUs:
1 - Good FP performance.
2 - Cheap.

At the time x86 could kinda do 1 but it could not do it at the same time as 2. The only reason 2 was met by the powerpc chips was that their general purpose performance was cut down immensely. The powerpc chips in game consoles are not the same as "normal" powerpcs. They would make terrible CPUs for general purpose use, but they are great for gaming.

Imagine you had a cheap VIA C7 CPU but with the FP performance of a Sandy Bridge. It would be great for some games and some scientific applications, but as a general purpose computer it would suck.

dan
Posts: 1243
Joined: Sun Jan 25, 2004 2:01 am
Contact:

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by dan » Sat Jan 14, 2012 5:09 pm

washu wrote:
dan wrote: I'm kinda curious as to what that RISC chip is -- Powerpc? MIPS?
There are lots of definitions as to what makes RISC, but the one I usually go by is that the chip implements a load-store architecture. IE, the CPU must use separate instructions to move data to and from the registers and cannot operate on data in memory directly.

A reduced number of instructions doesn't really apply any more, Powerpc for example has more instructions than x86.
There's a lot of talk about Apple ditching x86 for ARM, esp laptops
I don't think that is very likely unless Apple also ditches OSX on those laptops. The 68k->Powerpc and Powerpc->x86 switches were increases in CPU power. X86->ARM would be a major decrease in CPU power. There is no way an ARM could emulate x86 programs at a reasonable speed.

Apple might bring out a "low end" ARM laptop running iOS, but I don't think they will be getting rid of their x86 computers anytime soon.
How come Nintendo Sony Xbox use powerpc?
Simple, Sony and Microsoft needed two things for game console CPUs:
1 - Good FP performance.
2 - Cheap.

At the time x86 could kinda do 1 but it could not do it at the same time as 2. The only reason 2 was met by the powerpc chips was that their general purpose performance was cut down immensely. The powerpc chips in game consoles are not the same as "normal" powerpcs. They would make terrible CPUs for general purpose use, but they are great for gaming.

Imagine you had a cheap VIA C7 CPU but with the FP performance of a Sandy Bridge. It would be great for some games and some scientific applications, but as a general purpose computer it would suck.
so what is the definition of AVX2? not VLIW and not RISC. CISC still?

this is very interesting ... I remember the old days of powerpc vs intel. I'm sure Intel (or AMD) could design such a part solely for console ? Brazos for example?

Would a clean-room "pure" AVX2-x86-64 with all legacy removed and with backwards compatibility provided through software emulation, be both faster at same power consumption and/or same speed and lower power consumption due to simpler decoding logic? I recall the original pentium pro had problems with 16-bit code which was corrected with the pentium 2.

Ingeneer
Posts: 2
Joined: Fri Jan 06, 2012 2:27 am

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by Ingeneer » Sat Jan 14, 2012 6:20 pm

dan wrote:How much of a performance increase would this result in?
It highly depends on the application. Some could run two times faster or even more (if they didn't use SIMD instructions before). Some will see practically no benefit from AVX2. But keep in mind that Intel focusses on improving performance for applications for which it actually matters... Multimedia and games are likely to benefit greatly from AVX2.
Is this similar in spirit to VLIW Itanium EPIC, only compatible with x86
No. VLIW bundles multiple operations together into one instruction, so they can execute in parallel. AVX2 on the other hand is an SIMD extension which performs the same operation on multiple data elements. For instance it can perform eight 32-bit additions in parallel.
I know that it will be 3-operand, so does this mean AVX2 will essentially be RISC
No. One major characteristic of RISC is that it has separate instructions for memory accesses. AX2, like all previous x86 instructions, allows to access memory with practically every arithmetic instruction. It would be a waste to have made it RISC, since then the complex decoders would not be used.

washu
Posts: 571
Joined: Thu Nov 19, 2009 10:20 am
Location: Ottawa

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by washu » Sat Jan 14, 2012 6:52 pm

dan wrote:
so what is the definition of AVX2? not VLIW and not RISC. CISC still?
I'd say CISC because they are doing multiple steps in one instruction. Still, there are "RISC" cpus that have similar instructions so the line is quite blurry.
this is very interesting ... I remember the old days of powerpc vs intel. I'm sure Intel (or AMD) could design such a part solely for console ? Brazos for example?
Sure they probably could, but I don't see them spending the resources to do so. AMD might, but they have their GPUs that handle graphics just fine. Intel makes money hand over fist without making something custom.

Brazos isn't nearly one sided enough towards FP, even if you count the built in GPU.

I remember the old days of powerpc vs intel as well. The nice thing now is that you can run essentially the same OS (Linux, BSD) on both a power pc and x86 chips of the same eras and see which one was really faster. Here's the spoiler: Apple was lying.
Would a clean-room "pure" AVX2-x86-64 with all legacy removed and with backwards compatibility provided through software emulation, be both faster at same power consumption and/or same speed and lower power consumption due to simpler decoding logic? I recall the original pentium pro had problems with 16-bit code which was corrected with the pentium 2.
Removing 16/32 bit compatibility might save a tiny bit of power (emphasis on might). The instruction decoders on modern x86 CPUS are a such a minuscule part of the overall package even on an Atom that they really don't hurt power much. In addition, modern x86 CPUs are very good at shutting down unused parts. When you are running a 64 bit OS the 16/32 bit decoders are using practically no power.

Speed wise that wouldn't help at all. It's not like the 16/32 bit instructions somehow slow down the 64 bit ones. X86 is currently the 2nd fastest architecture on the market, it's not like they are really lagging. Intel has beaten all the RISC/VLIW processors (including their own VLIW CPU) in performance except one, and that one (POWER) isn't particularly energy efficient or cost effective.

Now compatibility wise this would be a nightmare for no benefit. Either the whole boot process of a normal PC would need to be changed or the BIOS would have to run an emulator just to boot. One of the main reasons for using x86 is compatibility. No other architecture has even close to the amount of legacy code. As a professional system administrator I still regularly get people asking why their 16 bit apps don't run on 64 bit Windows. If you can't run the code you need to then it doesn't matter how fast or energy efficient your CPU is, it's useless. Emulation isn't perfect and doesn't work in all situations. I've seen PCs as new as Core 2s running DOS just for legacy apps.

For many things legacy support and backwards compatibility is everything. Intel themselves have come up with four different non-x86 architectures meant to replace x86. They all failed or ended up in niche markets. IA64 was just the most recent one.

dan
Posts: 1243
Joined: Sun Jan 25, 2004 2:01 am
Contact:

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by dan » Sat Jan 14, 2012 7:34 pm

washu wrote:
dan wrote:
so what is the definition of AVX2? not VLIW and not RISC. CISC still?
I'd say CISC because they are doing multiple steps in one instruction. Still, there are "RISC" cpus that have similar instructions so the line is quite blurry.
this is very interesting ... I remember the old days of powerpc vs intel. I'm sure Intel (or AMD) could design such a part solely for console ? Brazos for example?
Sure they probably could, but I don't see them spending the resources to do so. AMD might, but they have their GPUs that handle graphics just fine. Intel makes money hand over fist without making something custom.

Brazos isn't nearly one sided enough towards FP, even if you count the built in GPU.

I remember the old days of powerpc vs intel as well. The nice thing now is that you can run essentially the same OS (Linux, BSD) on both a power pc and x86 chips of the same eras and see which one was really faster. Here's the spoiler: Apple was lying.
Would a clean-room "pure" AVX2-x86-64 with all legacy removed and with backwards compatibility provided through software emulation, be both faster at same power consumption and/or same speed and lower power consumption due to simpler decoding logic? I recall the original pentium pro had problems with 16-bit code which was corrected with the pentium 2.
Removing 16/32 bit compatibility might save a tiny bit of power (emphasis on might). The instruction decoders on modern x86 CPUS are a such a minuscule part of the overall package even on an Atom that they really don't hurt power much. In addition, modern x86 CPUs are very good at shutting down unused parts. When you are running a 64 bit OS the 16/32 bit decoders are using practically no power.

Speed wise that wouldn't help at all. It's not like the 16/32 bit instructions somehow slow down the 64 bit ones. X86 is currently the 2nd fastest architecture on the market, it's not like they are really lagging. Intel has beaten all the RISC/VLIW processors (including their own VLIW CPU) in performance except one, and that one (POWER) isn't particularly energy efficient or cost effective.

Now compatibility wise this would be a nightmare for no benefit. Either the whole boot process of a normal PC would need to be changed or the BIOS would have to run an emulator just to boot. One of the main reasons for using x86 is compatibility. No other architecture has even close to the amount of legacy code. As a professional system administrator I still regularly get people asking why their 16 bit apps don't run on 64 bit Windows. If you can't run the code you need to then it doesn't matter how fast or energy efficient your CPU is, it's useless. Emulation isn't perfect and doesn't work in all situations. I've seen PCs as new as Core 2s running DOS just for legacy apps.

For many things legacy support and backwards compatibility is everything. Intel themselves have come up with four different non-x86 architectures meant to replace x86. They all failed or ended up in niche markets. IA64 was just the most recent one.
Powers vs intel, Apple claimed that running say photoshop on mac os on powerpc doing certain benchmarks is faster than photoshop on windows on intel, which is relevant to apple's core market. Apple didn't claim that playing a 3D FPS was faster on powerpc than on intel with 3DFX back in the day -- whether these benchmarks were fair is another issue. apple didn't claimed running linux on powerpc was faster than linux on intel.

why is Atom having such a hard time competing with ARM then? sure the decorder logic has something to do with it.

I would think Intel and Microsoft and all publishers who also cross-publish on PC, like Call of Duty, would have some interest in making console x86 custom chips since then porting them to windows would be straight forward or more straight forward than from xbox xenon to ps3 cell to intel x86.

running c2duo for dos for legacy seems a bit overkill since Atom could do it just the same.

I know Itanium has RAS that is superior to Xeon, I wonder if they could build a Xeon from the ground up with comparable RAS as Itanium.
Do you think if Dec Alpha had continued development it would also outperform Xeon?

Doesn't the movement towards UEFI to replace BIOS eliminate some of that boot problems? and UEFI is 64-bit, and windows 7 and 8 also come in 64-bit, so in effect, esp when dealing with larger than 2tb, you would be in pure x64 from boot to apps.

Also don't forget that Apple just loves to jettison legacy at every chance. Apple's has a very large priority on reducing power consumption for sleek industrial designs, and their OS Lion and future OS require 64-bit.

washu
Posts: 571
Joined: Thu Nov 19, 2009 10:20 am
Location: Ottawa

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by washu » Sat Jan 14, 2012 8:20 pm

dan wrote: Powers vs intel, Apple claimed that running say photoshop on mac os on powerpc doing certain benchmarks is faster than photoshop on windows on intel, which is relevant to apple's core market. Apple didn't claim that playing a 3D FPS was faster on powerpc than on intel with 3DFX back in the day -- whether these benchmarks were fair is another issue. apple didn't claimed running linux on powerpc was faster than linux on intel.
While Photoshop was often a focus point, Apple most certainly did claim that PowerPC (not POWER, different but related CPU) was faster than x86 in general. The rigged benchmarks using GCC on G5 vs the P4 were a big example. Thing is, GCC has improved over the years for both CPUs. While the G5 has a slight IPC advantage over the P4, the P4 had such a huge clock speed advantage that it didn't matter, the P4 was the faster CPU. Apple also conveniently ignored AMD during those years when it was beating both PowerPC and Intel's x86 offerings. Apple lied outright and it wasn't just with Photoshop.

Look at the reviews of the Apple "Developer Transition Kit", which was pretty much a standard high end P4. The reviews raved about how much faster it was than the G5s out at the time.
why is Atom having such a hard time competing with ARM then? sure the decorder logic has something to do with it.
ARM and the Atom aren't really competitors (yet). For what the Atoms are used for (netbooks) their power consumption was fine so Intel didn't go lower.

Now Intel has come out with the Medfield CPUs which might actually challenge ARM seriously in the performance/watt area. While we will have to wait and see for some truly third party benchmarks to come out, Medfield seems to be doing fine despite having the "burden" of x86 decoder logic.

Don't forget that modern ARM CPUs have decoder logic too. They don't execute everything directly either.
I would think Intel and Microsoft and all publishers who also cross-publish on PC, like Call of Duty, would have some interest in making console x86 custom chips since then porting them to windows would be straight forward or more straight forward than from xbox xenon to ps3 cell to intel x86.
Most games aren't really written direct to the hardware like they were in older console generations. Porting is much easier, most of the difficulty is differing APIs not the CPU architecture. As long as the underlying CPU does the job it doesn't matter as consoles don't have the legacy issues PCs do.

running c2duo for dos for legacy seems a bit overkill since Atom could do it just the same.
The PC was from before Atoms were available.

My workplace uses Atom's extensively in "legacy" applications. While not going back to DOS, we do run some kiosks that have custom 32 bit drivers. Emulation or 64 bit only CPUs would be useless.

I know Itanium has RAS that is superior to Xeon, I wonder if they could build a Xeon from the ground up with comparable RAS as Itanium.
Do you think if Dec Alpha had continued development it would also outperform Xeon?
Alpha was an awesome architecture, but DEC didn't have the money to out engineer Intel. It may have taken longer, but x86 would have likely won in the end.

Doesn't the movement towards UEFI to replace BIOS eliminate some of that boot problems? and UEFI is 64-bit, and windows 7 and 8 also come in 64-bit, so in effect, esp when dealing with larger than 2tb, you would be in pure x64 from boot to apps.
Yes and no. If you are running a pure 64 bit OS then UEFI would be fine. However, that is very unlikely right now and it throws away backward compatibility. I'm running Win 7 64 bit on a UEFI machine and over half my processes are 32 bit. I would rather not take the performance hit from having to emulate those 32 bit instructions.

That also again ignores the huge advantage of x86, backwards compatibility. If I need to (and I have) I can run DOS on my i7-2600K. It's rare, but it is still useful. There's no point in throwing all that away unless I could get the performance of a POWER 7 at similar prices and power usage. Every other non-x86 CPU would be a step down in performance, so there is no point.

Also don't forget that Apple just loves to jettison legacy at every chance. Apple's has a very large priority on reducing power consumption for sleek industrial designs, and their OS Lion and future OS require 64-bit.
Apple isn't going to piss off all their graphic designers and other professionals who really need the performance that high end x86 offers. Running Photoshop, even if it wasn't emulated and even a quad core ARM would be brutal compared to a fast x86.

Besides, who would make all those ARM iOS apps? You can't make them on anything other than an x86 Mac.

dan
Posts: 1243
Joined: Sun Jan 25, 2004 2:01 am
Contact:

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by dan » Sun Jan 15, 2012 6:18 am

washu wrote:
dan wrote: Powers vs intel, Apple claimed that running say photoshop on mac os on powerpc doing certain benchmarks is faster than photoshop on windows on intel, which is relevant to apple's core market. Apple didn't claim that playing a 3D FPS was faster on powerpc than on intel with 3DFX back in the day -- whether these benchmarks were fair is another issue. apple didn't claimed running linux on powerpc was faster than linux on intel.
While Photoshop was often a focus point, Apple most certainly did claim that PowerPC (not POWER, different but related CPU) was faster than x86 in general. The rigged benchmarks using GCC on G5 vs the P4 were a big example. Thing is, GCC has improved over the years for both CPUs. While the G5 has a slight IPC advantage over the P4, the P4 had such a huge clock speed advantage that it didn't matter, the P4 was the faster CPU. Apple also conveniently ignored AMD during those years when it was beating both PowerPC and Intel's x86 offerings. Apple lied outright and it wasn't just with Photoshop.

Look at the reviews of the Apple "Developer Transition Kit", which was pretty much a standard high end P4. The reviews raved about how much faster it was than the G5s out at the time.
why is Atom having such a hard time competing with ARM then? sure the decorder logic has something to do with it.
ARM and the Atom aren't really competitors (yet). For what the Atoms are used for (netbooks) their power consumption was fine so Intel didn't go lower.

Now Intel has come out with the Medfield CPUs which might actually challenge ARM seriously in the performance/watt area. While we will have to wait and see for some truly third party benchmarks to come out, Medfield seems to be doing fine despite having the "burden" of x86 decoder logic.

Don't forget that modern ARM CPUs have decoder logic too. They don't execute everything directly either.
I would think Intel and Microsoft and all publishers who also cross-publish on PC, like Call of Duty, would have some interest in making console x86 custom chips since then porting them to windows would be straight forward or more straight forward than from xbox xenon to ps3 cell to intel x86.
Most games aren't really written direct to the hardware like they were in older console generations. Porting is much easier, most of the difficulty is differing APIs not the CPU architecture. As long as the underlying CPU does the job it doesn't matter as consoles don't have the legacy issues PCs do.

running c2duo for dos for legacy seems a bit overkill since Atom could do it just the same.
The PC was from before Atoms were available.

My workplace uses Atom's extensively in "legacy" applications. While not going back to DOS, we do run some kiosks that have custom 32 bit drivers. Emulation or 64 bit only CPUs would be useless.

I know Itanium has RAS that is superior to Xeon, I wonder if they could build a Xeon from the ground up with comparable RAS as Itanium.
Do you think if Dec Alpha had continued development it would also outperform Xeon?
Alpha was an awesome architecture, but DEC didn't have the money to out engineer Intel. It may have taken longer, but x86 would have likely won in the end.

Doesn't the movement towards UEFI to replace BIOS eliminate some of that boot problems? and UEFI is 64-bit, and windows 7 and 8 also come in 64-bit, so in effect, esp when dealing with larger than 2tb, you would be in pure x64 from boot to apps.
Yes and no. If you are running a pure 64 bit OS then UEFI would be fine. However, that is very unlikely right now and it throws away backward compatibility. I'm running Win 7 64 bit on a UEFI machine and over half my processes are 32 bit. I would rather not take the performance hit from having to emulate those 32 bit instructions.

That also again ignores the huge advantage of x86, backwards compatibility. If I need to (and I have) I can run DOS on my i7-2600K. It's rare, but it is still useful. There's no point in throwing all that away unless I could get the performance of a POWER 7 at similar prices and power usage. Every other non-x86 CPU would be a step down in performance, so there is no point.

Also don't forget that Apple just loves to jettison legacy at every chance. Apple's has a very large priority on reducing power consumption for sleek industrial designs, and their OS Lion and future OS require 64-bit.
Apple isn't going to piss off all their graphic designers and other professionals who really need the performance that high end x86 offers. Running Photoshop, even if it wasn't emulated and even a quad core ARM would be brutal compared to a fast x86.

Besides, who would make all those ARM iOS apps? You can't make them on anything other than an x86 Mac.

ARM has outlined a 64-bit extension. There are credible rumors that AMD might abandon x86 and feature ARM. AMD or even Apple could engineer high performance ARM processors. Apple has pissed off prior when they switched from PPC to Intel.

I have wondered if a legacy-free Skylake or Skymount could achieve Power 7 performance.
I don't know how much of a performance improvement could be had with AVX2 only but if it is substantial, it could still be backwards compatible AND faster with software emulation, as many x86-64 can run most x86 at native speeds.

I wonder where we would be if Apple went with Alpha rather than Powerpc. Since Alpha, PowerPC, MIPS are all Risc what made Alpha outperform PowerPC and MIPS?


I thought that the reason games look so good even though on much inferior hardware is that they are written close to the metal, unlike PC's. If I were Intel and software publishers, I'd urge Sony and MS to standardize the hardware and API.


http://lowendmac.com/hodges/06/0817.html

Ted Hodges - 2006.08.17

Was it all a lie?

The Mac Pro is the newest, coolest, fastest, and best Mac ever - or so Apple says.

Apparently, it's up to twice as fast as it's immediate predecessor, the Power Mac G5 Quad.

Sounds great, doesn't it?

A new system with a new architecture that's up to twice as fast is sure to bring in big bucks for Apple and make plenty of customers happy.
Wasn't PowerPC Faster?

But wait, there's something not quite right about all of this, something that I'm sure a lot of people are thinking: "I thought Intel chips were half as fast as the PowerPCs, not two to four times faster."

I seem to remember a time, about nine years ago, when Apple was bragging about how it's PowerPC G3 CPU was up to twice as fast as its Pentium II counterpart.

Well you can say that Apple has issued a retraction, it's now saying that Intel is about to 2-4 times faster than POWERPC. Donate your powerpc mac to charity and BUY the New Intel Macs!!!

washu
Posts: 571
Joined: Thu Nov 19, 2009 10:20 am
Location: Ottawa

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by washu » Sun Jan 15, 2012 8:16 am

dan wrote: ARM has outlined a 64-bit extension. There are credible rumors that AMD might abandon x86 and feature ARM. AMD or even Apple could engineer high performance ARM processors. Apple has pissed off prior when they switched from PPC to Intel.
I don't think the rumours of AMD switching to ARM have much credibility. What would be the benefit? Right now AMD has one real competitor, if they went to ARM they would have many.

I don't think you appreciate the difference in performance between even a "fast" ARM and a fast x86. It's not like the difference between an i3 and an i7, it is like difference between a slow Atom and an i7. Simply put ARM CPUs are slow, very slow. They have great performance/watt because that is what they are designed for.

Now, could AMD/Apple make an ARM that is performance competitive with x86? Sure, but it would take years and cost billions. That is also ignoring the fact that Intel is going to take years and spend billions improving x86 because that is what they do. In the end you might have an ARM that performs like a Core 2, where as Intel will have Skylake or beyond.

Apple pissed of a few people when they switched processors, but again it is about performance. What is worse: Having to get new apps that perform way better, or having to get new apps that perform way worse? x86->ARM would be the latter.
I have wondered if a legacy-free Skylake or Skymount could achieve Power 7 performance.
I don't know how much of a performance improvement could be had with AVX2 only but if it is substantial, it could still be backwards compatible AND faster with software emulation, as many x86-64 can run most x86 at native speeds.
I already went over this and the answer is no. The legacy bits of x86 are not slowing it down. Removing them would have no benefit and lots of downsides. X86-64 CPUs are not emulating 32 bit instructions (or even 16 bit) they execute them natively and at full speed. X86 CPUs will continue to operate this way as it makes sense.
I wonder where we would be if Apple went with Alpha rather than Powerpc. Since Alpha, PowerPC, MIPS are all Risc what made Alpha outperform PowerPC and MIPS?
We'd probably be in the same place. Choosing a different CPU after the 68k transition would not have stopped Apple from being completely boneheaded in the 90s. Apple made some PowerPC Macs that make the worst Cyrix emachines look high end. They would have done the same with Alphas. The slight increase in Alpha sales would not have made much difference to DEC's bottom line and Intel would have still beaten them in the end.

Alpha outperformed everything at the start for the same reason Intel and POWER outperform everything now, money. While DEC was going down when the Alpha came out, they still made massive investments in it. They also had lots of experience in microprocessors; Their CISC VAX CPUs were quite fast for the time, just expensive.

dan
Posts: 1243
Joined: Sun Jan 25, 2004 2:01 am
Contact:

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by dan » Sun Jan 15, 2012 4:49 pm

washu wrote:
dan wrote: ARM has outlined a 64-bit extension. There are credible rumors that AMD might abandon x86 and feature ARM. AMD or even Apple could engineer high performance ARM processors. Apple has pissed off prior when they switched from PPC to Intel.
I don't think the rumours of AMD switching to ARM have much credibility. What would be the benefit? Right now AMD has one real competitor, if they went to ARM they would have many.

I don't think you appreciate the difference in performance between even a "fast" ARM and a fast x86. It's not like the difference between an i3 and an i7, it is like difference between a slow Atom and an i7. Simply put ARM CPUs are slow, very slow. They have great performance/watt because that is what they are designed for.

Now, could AMD/Apple make an ARM that is performance competitive with x86? Sure, but it would take years and cost billions. That is also ignoring the fact that Intel is going to take years and spend billions improving x86 because that is what they do. In the end you might have an ARM that performs like a Core 2, where as Intel will have Skylake or beyond.

Apple pissed of a few people when they switched processors, but again it is about performance. What is worse: Having to get new apps that perform way better, or having to get new apps that perform way worse? x86->ARM would be the latter.
I have wondered if a legacy-free Skylake or Skymount could achieve Power 7 performance.
I don't know how much of a performance improvement could be had with AVX2 only but if it is substantial, it could still be backwards compatible AND faster with software emulation, as many x86-64 can run most x86 at native speeds.
I already went over this and the answer is no. The legacy bits of x86 are not slowing it down. Removing them would have no benefit and lots of downsides. X86-64 CPUs are not emulating 32 bit instructions (or even 16 bit) they execute them natively and at full speed. X86 CPUs will continue to operate this way as it makes sense.
I wonder where we would be if Apple went with Alpha rather than Powerpc. Since Alpha, PowerPC, MIPS are all Risc what made Alpha outperform PowerPC and MIPS?
We'd probably be in the same place. Choosing a different CPU after the 68k transition would not have stopped Apple from being completely boneheaded in the 90s. Apple made some PowerPC Macs that make the worst Cyrix emachines look high end. They would have done the same with Alphas. The slight increase in Alpha sales would not have made much difference to DEC's bottom line and Intel would have still beaten them in the end.

Alpha outperformed everything at the start for the same reason Intel and POWER outperform everything now, money. While DEC was going down when the Alpha came out, they still made massive investments in it. They also had lots of experience in microprocessors; Their CISC VAX CPUs were quite fast for the time, just expensive.
I enjoy this trip down memory lane.

So how come intel has to date been unable to design an Atom SOC that Apple would use for their IPAD, or Samsung for their Galaxy, or any tablet or smart phone? I am aware of medfield but no apparent takers yet.

nutball
*Lifetime Patron*
Posts: 1304
Joined: Thu Apr 10, 2003 7:16 am
Location: en.gb.uk

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by nutball » Mon Jan 16, 2012 1:39 am

dan wrote:why is Atom having such a hard time competing with ARM then? sure the decorder logic has something to do with it.
Atom isn't really competing with ARM - Atoms job is to run Windows, ARM is useless for this as has been pointed out already. If you want to take the cynical view, Intel carefully bounded the performance of Atom to avoid having it compete with their own low-end but higher-priced x86 products (Celeron, and Pentium as it's now known).
Do you think if Dec Alpha had continued development it would also outperform Xeon?
Maybe, maybe not, but even if Alpha were around today I'd wager it'd be a niche product just as Itanium and POWER are. They are not attractive on the desktop due to lack of legacy support; they are not attractive in the low- and mid-range server market because they're too expensive and power-hungry. So they've ended up where they are - being sold to people with the workloads to justify them and who don't really care about the price.

Arguably Itanium and POWER can only exist today because of the massive financial and technical resources that their respective parents, Intel and IBM, have behind them. I'm not sure there are many or even any other companies who could sensibly give Alpha the support it would require to compete today whilst keeping a straight face when trying to justify such an effort to its shareholders.

Despite DECs grandiose plans for Alpha to take over the desktop, it failed partly because Intel delivered a blinder with P6, but just as much because it lacked in the legacy support department. Itanium has suffered the same fate on the desktop, out-competed at the low-end initially by AMD then by Intel itself, whilst also lacking in the legacy support department.

HFat
Posts: 1753
Joined: Thu Jul 03, 2008 4:27 am
Location: Switzerland

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by HFat » Mon Jan 16, 2012 2:40 am

nutball wrote:If you want to take the cynical view, Intel carefully bounded the performance of Atom to avoid having it compete with their own low-end but higher-priced x86 products (Celeron, and Pentium as it's now known).
I'm not sure Atoms with significantly higher performance would be efficient compared to mainstream CPUs for most users (for some niche uses, I guess a high x86 core count in a small and affordable package could be nice).
Regardless, most people are OK with the CPU performance of the better Atoms (the kind of people who post on tech forums aren't representative). That's not what's keeping Atoms from competing with other low-end Intel products. Two-thirds of my recent x86 purchases have been Atoms.
The issue with Atoms (and the reason 90% of my purchases haven't been Atoms) are the limitations of the chipsets and the graphics. Instead of working with Nvidia or server/NAS players, Intel is trying to lock everyone out. Meanwhile they still aren't shipping chipsets with decent I/O or functional graphics drivers for last year's Atoms.

On topic, I have to agree: Intel and AMD are obviously not going to jettison x86 any time soon.

dan
Posts: 1243
Joined: Sun Jan 25, 2004 2:01 am
Contact:

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by dan » Mon Jan 16, 2012 10:47 am

nutball wrote:
dan wrote:why is Atom having such a hard time competing with ARM then? sure the decorder logic has something to do with it.
Atom isn't really competing with ARM - Atoms job is to run Windows, ARM is useless for this as has been pointed out already. If you want to take the cynical view, Intel carefully bounded the performance of Atom to avoid having it compete with their own low-end but higher-priced x86 products (Celeron, and Pentium as it's now known).
Do you think if Dec Alpha had continued development it would also outperform Xeon?
Maybe, maybe not, but even if Alpha were around today I'd wager it'd be a niche product just as Itanium and POWER are. They are not attractive on the desktop due to lack of legacy support; they are not attractive in the low- and mid-range server market because they're too expensive and power-hungry. So they've ended up where they are - being sold to people with the workloads to justify them and who don't really care about the price.

Arguably Itanium and POWER can only exist today because of the massive financial and technical resources that their respective parents, Intel and IBM, have behind them. I'm not sure there are many or even any other companies who could sensibly give Alpha the support it would require to compete today whilst keeping a straight face when trying to justify such an effort to its shareholders.

Despite DECs grandiose plans for Alpha to take over the desktop, it failed partly because Intel delivered a blinder with P6, but just as much because it lacked in the legacy support department. Itanium has suffered the same fate on the desktop, out-competed at the low-end initially by AMD then by Intel itself, whilst also lacking in the legacy support department.
I'm surprised given margins and process lead that Intel or AMD doesn't create a x64 to compete with Power 7

Windows 7 is useless but windows 8 and windows phone do run on ARM. Legacy isn't much of an issue yet for smartphone/tablet. If Intel doesn't get marketshare, of course, legacy on ARM could be an issue of the future.


it's interesting we are having this conversation since
ie
http://www.dailytech.com/ARM+and+Intel+ ... e23772.htm

The war between x86 and ARM is almost upon us
One of the most compelling storylines to emerge from the 2012 Consumer Electronics Show was the struggle between ARM and x86, two architectures vying for CPU global domination. ARM’s proponents argue that the reduced instruction set computer (RISC) architecture used in ARM is inherently more power efficient than x86. x86 proponents disagree, often arguing that ARM cores are too weak for serious computing. x86's proponents also brag about their lead in semiconductor manufacturing processes, which they feel will yield more efficient features, putting x86 ahead of ARM in power efficiency.


ARM currenty dominates the majority of embedded markets (think dishwashers, microwaves, etc.). It also owns a previously unchallenged monopoly on smartphone chip architecture, and a near-complete monopoly on tablet SoCs.


http://www.ibtimes.com/articles/270778/ ... .htm[quote]

ARM vs X86: Could Intel Offer a Cheaper, More Powerful Smartphone Processor?

As of today, hardware firm ARM has a dominant share of the market. Its technology forms the basis of several smartphone chips. Meanwhile, U.S.-based chip giants, Intel, are trying to make inroads into the mobile computing race, through its x86 technology and are expected to reap benefits in the coming days.
[/quote]

the reason this matters is
"While our cycle framework tells us it is too early to Buy INTC, this analysis makes us more constructive on INTC's long term potential in mobile applications," added Lipacis who has a "hold" rating on Intel stock.

nutball
*Lifetime Patron*
Posts: 1304
Joined: Thu Apr 10, 2003 7:16 am
Location: en.gb.uk

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by nutball » Tue Jan 17, 2012 12:39 am

dan wrote:I'm surprised given margins and process lead that Intel or AMD doesn't create a x64 to compete with Power 7
Intel will do this eventually once Itanium runs out of steam. AMD don't have the capability (technical or financial) to do it at all. But it's not just something you do at a drop of a hat, the differences between POWER7/Itanium and top-end x86 are very significant, and are as much to do with the architecture of the whole system as with the CPU itself. IBM has the advantage of having control of pretty much the whole system, whereas Intel doesn't work like that (at the moment).
Windows 7 is useless but windows 8 and windows phone do run on ARM.
Windows 8 will run on ARM, legacy Windows binaries won't. As has already been pointed out, ARM cannot compete with low-end x86 in a non-power-constrained environment.
Legacy isn't much of an issue yet for smartphone/tablet. If Intel doesn't get marketshare, of course, legacy on ARM could be an issue of the future.
I don't think the two scenarios are really comparable. Phone apps tend to be written in languages which are readily portable, or which don't directly target the instruction set of the underlying CPU, rather they use JITs and so on. This is different from the situation with older Windows apps (maybe even modern Windows apps, those not written in C# or similar). Users of phone apps don't expect to be able to run 10-year-old apps on their latest phones, and even if they did it might well work if the relevant JIT were available.

dan
Posts: 1243
Joined: Sun Jan 25, 2004 2:01 am
Contact:

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by dan » Tue Jan 17, 2012 9:20 am

nutball wrote:
dan wrote:I'm surprised given margins and process lead that Intel or AMD doesn't create a x64 to compete with Power 7
Intel will do this eventually once Itanium runs out of steam. AMD don't have the capability (technical or financial) to do it at all. But it's not just something you do at a drop of a hat, the differences between POWER7/Itanium and top-end x86 are very significant, and are as much to do with the architecture of the whole system as with the CPU itself. IBM has the advantage of having control of pretty much the whole system, whereas Intel doesn't work like that (at the moment).
Windows 7 is useless but windows 8 and windows phone do run on ARM.
Windows 8 will run on ARM, legacy Windows binaries won't. As has already been pointed out, ARM cannot compete with low-end x86 in a non-power-constrained environment.
Legacy isn't much of an issue yet for smartphone/tablet. If Intel doesn't get marketshare, of course, legacy on ARM could be an issue of the future.
I don't think the two scenarios are really comparable. Phone apps tend to be written in languages which are readily portable, or which don't directly target the instruction set of the underlying CPU, rather they use JITs and so on. This is different from the situation with older Windows apps (maybe even modern Windows apps, those not written in C# or similar). Users of phone apps don't expect to be able to run 10-year-old apps on their latest phones, and even if they did it might well work if the relevant JIT were available.

I have a question then -- if IBM had chosen Motorola 68000 rather than 8088 for its first PC, resulting in R&D behind 68k comparable to x86, would it be possible to develop the 68k to as fast if not faster than what Intel has today?

I thought x86 was register starved.

nutball
*Lifetime Patron*
Posts: 1304
Joined: Thu Apr 10, 2003 7:16 am
Location: en.gb.uk

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by nutball » Tue Jan 17, 2012 9:52 am

dan wrote:I have a question then -- if IBM had chosen Motorola 68000 rather than 8088 for its first PC, resulting in R&D behind 68k comparable to x86, would it be possible to develop the 68k to as fast if not faster than what Intel has today?
Maybe as fast. Possibly not noticeably faster. Who knows. Honestly though that's like asking whether cars would be more or less fuel efficient today if Henry Ford had made some different design choices. Comparing a modern x86 CPU to 8088 is like comparing a modern turbo-diesel Ford Mondeo to a Model-T.
I thought x86 was register starved.
Was. Once upon a time. Before the invention of super-scalar, out-of-order execution, register renaming and all that jazz, and x86_64 of course.

Pop quiz: how many registers are there in a modern x86 CPU?

dan
Posts: 1243
Joined: Sun Jan 25, 2004 2:01 am
Contact:

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by dan » Tue Jan 17, 2012 10:31 am

nutball wrote:
dan wrote:I have a question then -- if IBM had chosen Motorola 68000 rather than 8088 for its first PC, resulting in R&D behind 68k comparable to x86, would it be possible to develop the 68k to as fast if not faster than what Intel has today?
Maybe as fast. Possibly not noticeably faster. Who knows. Honestly though that's like asking whether cars would be more or less fuel efficient today if Henry Ford had made some different design choices. Comparing a modern x86 CPU to 8088 is like comparing a modern turbo-diesel Ford Mondeo to a Model-T.
I thought x86 was register starved.
Was. Once upon a time. Before the invention of super-scalar, out-of-order execution, register renaming and all that jazz, and x86_64 of course.

Pop quiz: how many registers are there in a modern x86 CPU?
I know that there are 8-16 that are directly accessible to the compiler but due to register renaming there are nearly 100 in the CPU.

I know that the reason that AMD expanded from 8 to 16 GP registers is due to x86 encoding even though 32 or 64 registers would provide more performance in some tasks (at the expense of swollen pointer) but I'm kinda disappointed that AMD/intel did not tinker with encoding so as to have 32-64 GP registers.

Do you think removing the archaic x87 FPU entirely and replacing it with a "pure" highly performance optimized AVX/SSE unit would be a smart engineering choice for target market?

nutball
*Lifetime Patron*
Posts: 1304
Joined: Thu Apr 10, 2003 7:16 am
Location: en.gb.uk

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by nutball » Tue Jan 17, 2012 11:33 am

dan wrote:I know that there are 8-16 that are directly accessible to the compiler but due to register renaming there are nearly 100 in the CPU.
So not miles off from any of the relevant competition then, at either end of the performance spectrum.
I'm kinda disappointed that AMD/intel did not tinker with encoding so as to have 32-64 GP registers.
Can you give any concrete examples of where the difference in available GPs has made a measurable difference to you personally, in a task that scales to an economically relevant number of customers?
Do you think removing the archaic x87 FPU entirely and replacing it with a "pure" highly performance optimized AVX/SSE unit would be a smart engineering choice for target market?
Erm... which target market? We seem to be talking about at least three (we are talking ARM, x86 and POWER7, they are demonstratively distinct target markets if you care to draw a Venn diagram).

Anyway, this has happened already has it not?!

dan
Posts: 1243
Joined: Sun Jan 25, 2004 2:01 am
Contact:

Re: Will AVX2 ISA allow Intel to jettison X86 legacy/baggage

Post by dan » Wed Jan 18, 2012 8:52 am

nutball wrote:
dan wrote:I know that there are 8-16 that are directly accessible to the compiler but due to register renaming there are nearly 100 in the CPU.
So not miles off from any of the relevant competition then, at either end of the performance spectrum.
I'm kinda disappointed that AMD/intel did not tinker with encoding so as to have 32-64 GP registers.
Can you give any concrete examples of where the difference in available GPs has made a measurable difference to you personally, in a task that scales to an economically relevant number of customers?
Do you think removing the archaic x87 FPU entirely and replacing it with a "pure" highly performance optimized AVX/SSE unit would be a smart engineering choice for target market?
Erm... which target market? We seem to be talking about at least three (we are talking ARM, x86 and POWER7, they are demonstratively distinct target markets if you care to draw a Venn diagram).

Anyway, this has happened already has it not?!

What was the reason that Itanium EPIC failed to deliver superior price/performance/energy given it was a clean-room design, no legacy?
Intel's original plan was to make the pentium 4 the last and 64-bt by the Itanium
yeah, i wonder what cpu ps4 and xbox720 will have.

Post Reply