Haswell getting AVX2... will Atom too?

[VirtualLarry]

Just curious, if Intel's strategy is to maintain ISA parity between their regular "Core" series of CPUs, and their lower-power Atom CPUs, or if they are intentionally going to limit Atom's ISA extensions, for either power, or more likely, market-segmentation issues.

I was thinking of the server market, since several vendors make chipsets to integrate Atoms into servers, and whether or not the server software market (HPC?) would start to utilize the ISA extensions like SSE2 and AVX2, and if Intel would include them into Atom for this reason.

Edit: Seeing as how Intel doesn't even support 64-bit on Atom on their newest Atom platform/chips, then I guess the answer is likely to be "No AVX2 support" as well.

 

[BenchPress]

Intel has stated to put Atom on a tick-tock development cadence as well. In particular this means a new micro-architecture every two years. And they'll also make it use the latest process technology. So it makes sense to converge toward the desktop chip micro-architecture.

Think about it. The dual-core Haswell chip for ultrabooks will consume 15 Watt, so a single-core (with Hyper-Threading) would consume about 7 Watt, and at half the frequency and lower voltage it would be 3 Watt... All while supporting AVX2.

Perhaps it will take them a couple generations to actually get there, but clearly there's no reason why they can't eventually use the same micro-architecture with the same ISA extensions. Also note that AVX2 is aimed at achieving very high performance per Watt. For a CPU with out-of-order execution the execution units only account for ~20% of the power consumption anyway so no reason not to make them 256-bit wide. Also look at how GPUs achieve high throughput without breaking a sweat.

And AVX-1024 would further lower the power consumption by executing in four 256-bit chunks and thereby lowering the load on the rest of the pipeline. So Intel has good reasons to sooner or later support it all on the Atom line as well.

Not to mention mobile and desktop software development appears to be converging as well so chips with the same capabilities facilitate that...

 

[Edrick]

I doubt Atom will have AVX2 for a long time, if ever. The Atom was built off the P54C core with a few added features like x64. Atom does not even have the instructions that Nehalem had back in 2008, never mind Haswell. And there is good reasons why, mostly it is not needed for the market that Atom is in. Who needs AVX2 on their smartphones today?

Also, adding sets like AVX and AVX2 require specialized cache structures and memory bandwidth to be effective. Atom just doesn't have that. All this will just add size, power comsumption and heat to the Atom, which is not wanted or needed.

 

[Denithor]

Why exactly doesn't Atom support 64-bit OS?

Crap like that is going to prevent us from seeing a complete migration to 64-bit in the next few years.

 

[dealcorn]

This is doable if Intel thinks there is demand at 22 nm (e.g. server farms running AVX2). The way Moore's Law works, it cost roughly half as much to implement at 14 nm as it does at 22 nm. If Intel sincerely believe Atom will be top dog in category at 14 nm, AVX2 might be a nice way to further distinguish Atom from the also rans.

 

[BenchPress]

Who needs AVX2 on their smartphones today?

I'm sorry but you're being too shortsighted. Back when cell phones were merely just wireless phones, most people couldn't imagine what you would need a dual-core 1 GHz chip for. Today it's the norm. So we'll need even higher performance in the not too distant future.

But instead of using a quad-core or a 2 GHz chip for that, it would be far more power efficient to have something like AVX2. It offers an eightfold increase in performance for many of the most time consuming code loops.

Also, adding sets like AVX and AVX2 require specialized cache structures and memory bandwidth to be effective. Atom just doesn't have that. All this will just add size, power comsumption and heat to the Atom, which is not wanted or needed.

No, it doesn't take a lot of size or power consumption at all. Again, just look at GPUs. In relative terms they have very high throughput and bandwidth but low power consumption precisely because of technology quite similar to AVX2!

So there's really no reason why Intel wouldn't want AVX2 support for Atom as soon as possible. And again don't underestimate the value of being able to run the same binary on all devices. Apple has indicated to converge iOS and Mac OS, and Windows 8 clearly bridges the gap between laptops and tablets.

 

[BenchPress]

Why exactly doesn't Atom support 64-bit OS?

Crap like that is going to prevent us from seeing a complete migration to 64-bit in the next few years.

Atom does support x86-64. It's just not enabled on most models, probably because of marketing reasons (giving you an incentive to upgrade when they enable it in next-gen devices).

Yes it stinks, but Atom does support x86-64.

 

[dave_the_nerd]

Atom does support x86-64. It's just not enabled on most models, probably because of marketing reasons (giving you an incentive to upgrade when they enable it in next-gen devices).

Yes it stinks, but Atom does support x86-64.

Chipset support.

 

[Edrick]

No, it doesn't take a lot of size or power consumption at all. Again, just look at GPUs. In relative terms they have very high throughput and bandwidth but low power consumption precisely because of technology quite similar to AVX2!

I am a huge fan and supporter of AVX2. And it would be nice to see it on all Intel CPUs. But it will not happen anytime soon. You can't compare a GPU to an Atom for this arguement, GPU simply has more memory bandwidth and specialized caches that makes it good at what it does. Atom has nothing of the sorts. Remember, Atom was designed from the P54C core from 12 years ago.

The day Intel takes a dual core Haswell, clocks it down to 2Ghz, removes the IGP, and calls it an Atom is the day Atom will support AVX2.

 

[BenchPress]

You can't compare a GPU to an Atom for this arguement, GPU simply has more memory bandwidth and specialized caches that makes it good at what it does.

How is that even remotely relevant? It uses the same laws of semiconductor physics. Anything a GPU can achieve for general purpose programmable computing, can apply to a CPU as well. Nothing is preventing Intel from supporting AVX2 and getting good performance per Watt out of it.

Atom has nothing of the sorts. Remember, Atom was designed from the P54C core from 12 years ago.

Intel has already officially announced that Atom will get a new micro-architecture, designed from the ground up, every two years or less. They already have Silvermont lined up for next year. They've also stated that Atom will be the first to use 14 nm technology, before anything else! So this won't be your father's Atom, it will be an entirely different beast altogether.

Airmont is claimed to be 10x faster than today's Atom. That sounds like AVX2 to me.

 

[Khato]

Remember, Atom was designed from the P54C core from 12 years ago.

While that only applies to the current Bonnell atom microarchitecture, keep in mind that the Larrabee cores are also based on the P54C.

Thus far I don't believe we have any real information as to what all will be included in the 2013 Silvermont based atoms. It's quite possible that they'll include AVX2 right off the bat in order to give Intel that much more of a performance/power edge. I'm not really expecting to hear anything about Silvermont until Fall IDF at the earliest though.

 

[alyarb]

Until the DRAM is on die with a super-wide interposer it's silly to throw in all these additional execution resources without the supporting hardware available to make it worthwhile. Most socs today have 32-bit memory and 1MB of cache. Medfield has widened that to 64-bit, but 128-bit is not realistic and you've got to realize the energy cost keeping the memory bus under constant load. Assuming that you can cut the TDP by 75% by removing a core and clocking downward is also a bit rash.

Mobile devices are really trending toward cloud computing with a GPU-accelerated render/output/display. kilobit vector computing was never in the cards for milliwatt SoCs but who knows what the future holds.

With regard to AVX2-enabled atoms in servers, it would be a highly specialized system, as many workloads would prefer to perform their AVX2 execution on 32 big haswell cores rather than 256 little atoms.

 

[Phynaz]

Why exactly doesn't Atom support 64-bit OS?

Crap like that is going to prevent us from seeing a complete migration to 64-bit in the next few years.

Why exactly would you need > 4GB RAM in an Atom system?

 

[BenchPress]

Why exactly would you need > 4GB RAM in an Atom system?

x86-64 is a lot more than just support for more than 4 GB of RAM.

 

[bronxzv]

Just curious, if Intel's strategy is to maintain ISA parity between their regular "Core" series of CPUs, and their lower-power Atom CPUs

much like today's Atoms provide support for SSEx (processed in 2 chunks AFAIK as desktop CPUs before Conroe) and Intel 64 (the hardware is there, not always enabled) I suppose they will probably provide support for AVX (at 22 nm) and AVX2 at 14nm

 

[bronxzv]

You can't compare a GPU to an Atom for this arguement, GPU simply has more memory bandwidth and specialized caches that makes it good at what it does.

GPUs in embedded SOCs don't have "more memory bandwidth" since the memory bandwidth is shared by all the SOC agents, in fact embedded GPUs all use TBR 3D engines (where most of the work is done in the fast and small on chip memory) to *save bandwidth*, for the same reason an AVX2 CPU will do marvels simply with fast medium size caches, a pure software renderer with fully deferred texturing will save external bandwidth beyond TBR + early Z cull methods and will allow to deal with more complex scenes, it will probably not beat the on chip GPU for all workloads but will be competitive for complex scenes with advanced effects

 

[BenchPress]

Until the DRAM is on die with a super-wide interposer it's silly to throw in all these additional execution resources without the supporting hardware available to make it worthwhile.

You don't need extra RAM bandwidth to make AVX2 worthwhile. Most of the work is done on data within the caches. So it definitely needs additional cache bandwidth, but that's relatively cheap. In fact the power consumption of one full cache line transfer should be lower than that of two half cache line transfers.

It's again comparable to a GPU. Take the GTX 680 for instance. Its 192.2 GB/s RAM bandwidth may seem phenomenal, but its peak arithmetic throughput actually corresponds to 12,360 GB/s! So the internal "bandwidth" is way higher than the external one.

AVX2 does represent a leap in arithmetic throughput for the CPU, but will only require incremental RAM bandwidth.

 

[Phynaz]

x86-64 is a lot more than just support for more than 4 GB of RAM.

That's not what he asked.

 

[BenchPress]

That's not what he asked.

Read the context of his question.

 

[podspi]

Silvermont will be OoO, yes? The 22nm and 14nm Atoms will be very, very interesting. I'm really looking forward to/hoping for mobile phones loaded with Windows 8 (or equivalent) with docking stations.

My Droid 4 can already do this, but two A9's @ 1.2ghz just isn't fast enough for desktop use.

 

[IntelUser2000]

Think about it. The dual-core Haswell chip for ultrabooks will consume 15 Watt, so a single-core (with Hyper-Threading) would consume about 7 Watt, and at half the frequency and lower voltage it would be 3 Watt... All while supporting AVX2.

Actually, being an SoC means CPU core has much smaller impact on power usage. On 17W SKUs, the graphics are(usually) allocated to half of the TDP budget. Assuming its similar for 15W...

8W CPU
7W GPU

Reducing the CPU to a quarter would have the SoC still using 9W. Also at half the frequency and single core, even Haswell would perform pretty poor.

 

[Edrick]

Airmont is claimed to be 10x faster than today's Atom. That sounds like AVX2 to me.

I will believe 10X faster than todays Atoms when I see it.

Again, I am not saying I don't want Atoms to get AVX2, I would love to see it. I think AVX2 is the best thing to hit CPUs in many years. I am just being realistic based on the Atom family tree up to this point. Intel clearly has plans for Atom, but in my opinion, I think AVX2 may be a bit lower on the hit list.

 

[Roland00Address]

Why exactly doesn't Atom support 64-bit OS?

Crap like that is going to prevent us from seeing a complete migration to 64-bit in the next few years.

Most atoms support 64 instructions and 64 bit OS the ones that do not are
Do not support 64bit instructions and OS
Atom Z5xx (mobile internet devices 2008)
Atom N2xx (netbook 2008)

Support 64bit instructions and OS
Atom Z6xx (mobile internet devices 2010)
Atom 2xx and Atom 3xx ( nettop 2008)
Atom N4xx / N5xx (netbook 2010)
Atom D4xx / D5xx (nettop 2010)
Atom D2xxx (nettop late 2011)
Atom N2xxx (netbook late 2011)

Medfield (2012 system on a chip targeting cellphones, mids, tablets) is 32bit if I remember but do not quote me on that.

 

[BenchPress]

I think AVX2 is the best thing to hit CPUs in many years. I am just being realistic based on the Atom family tree up to this point. Intel clearly has plans for Atom, but in my opinion, I think AVX2 may be a bit lower on the hit list.

It's important to realize that AVX2 isn't just about higher peak throughput. It's also about getting the same amount of work done in half the time (or less).

This is particularly critical on a mobile platform where the "sprint to idle" paradigm applies. A powerful CPU doesn't drain your battery any faster when doing the same tasks. It just makes things snappier and might even lower the total power consumption because you can power gate sooner or lower voltage and save on leakage.

So Intel should have AVX2 for Atom pretty high on the list. It would give them an extra advantage over ARM at performance/Watt.