ARM Cortex-A15 the successor of ARM Cortex-A9

[Grall]

I wonder how fast a consumer-level ethernet router could be with a chip like this driving it... Considering current top-end routers from D-Link, Netgear and so on employ 200-400MHz-ish ARM chips of generally unknown generation and capability (ARM 6-7 probably) and route 100MB/s and up, perhaps we could see full gigabit line speed from a multicore 2.5GHz CPU.

 

[cal_guy]

I wonder how fast a consumer-level ethernet router could be with a chip like this driving it... Considering current top-end routers from D-Link, Netgear and so on employ 200-400MHz-ish ARM chips of generally unknown generation and capability (ARM 6-7 probably) and route 100MB/s and up, perhaps we could see full gigabit line speed from a multicore 2.5GHz CPU.

Aren't most routers powered by MIPS based chips?

 

[rpg.314]

One feature that I hope Cortex-A15 will have is the ability to share one NEON core among multiple CPU cores. The designs are fairly decoupled as it is so I hope that this will be a possibility.

AKA Bulldozer

 

[Exophase]

AKA Bulldozer

Exactly

 

[iwod]

I wonder what is the use 4GB + memory in server market. Most of the Hosting,Grid Hosting, Cloud Hosting are / can be done in a VM environment where you dont get more then 4GB Memory in one instance.

So that is at least part of the server market that can be addressed with ARM. It may not be a large % in server market shares, but it is a good fit for ARM nevertheless.

Software Stack are mainly open sourced software. So i am not at all worried about running windows like all internet forums have been crying about. It just means we have 2 - 3 years to polish LAMP and LLVM for ARM.

 

[Alexko]

Is any of that problematic if the chip skipped ahead to 64 bits?
PAE may fit that usage case, but why target that one situation at the expense of being marketable for all the other ones?

How does ARM's business model work exactly? I mean, do they get significantly more if TI/Samsung/Qualcomm/Whoever sells an ARM server chip than if they sell a smartphone chip?

The whole point of the server market for the x86 guys (and indeed IBM et al.) is that while volume is low, margins are very high. But does that apply to ARM as well, since they don't sell chips but only licenses?

If not, perhaps it explains their reluctance to invest too much silicon in features that wouldn't be very useful outside the server market.

 

[Arun]

How does ARM's business model work exactly? I mean, do they get significantly more if TI/Samsung/Qualcomm/Whoever sells an ARM server chip than if they sell a smartphone chip?

No, but iirc they do get royalties relative to the number of cores, so at least they'd get more than on an handset. Still not much given the massive lower volumes though.

On the other hand, ARM seems to be investing directly in an ARM-based server startup so they do stand to benefit more directly: http://www.eetimes.com/electronics-news/4206171/ARM--ATIC--TI-invest--48M-in-server-chip-startup-

 

[Exophase]

I read somewhere that ARM gets 1% in licensing, so if true it's proportional to chip cost and they won't lose out on server sales being lower volume.

IMO, even supporting 16 cores isn't going to be very useful outside of the server market (or at the very least, probably won't be seen in mobile devices in 2013, or even high end embedded where multi-core is common now)

 

[digitalwanderer]

Ok, someone make this simple for me. Does this mean that we'll be seeing uber cool/cheap/portable netbooks soon? :|

 

[3dilettante]

I read somewhere that ARM gets 1% in licensing, so if true it's proportional to chip cost and they won't lose out on server sales being lower volume.

The fixed cost for validating the extensions and designing/validating the IP would have to be amortized over that small volume. These chips will never have the comparatively massive prices of high-end x86, and their best prices are likely far below the cheapest chips AMD or Intel would be willing to sell.
That probably is figured into the calculus that went into how they extended the address space. It's a low-cost evolution, which probably means the expectations are that even if the prices are higher than is customary for an ARM, those combined with the small volume and competition are not going to justify a full transition to 64-bit at this point.

 

[Exophase]

The fixed cost for validating the extensions and designing/validating the IP would have to be amortized over that small volume. These chips will never have the comparatively massive prices of high-end x86, and their best prices are likely far below the cheapest chips AMD or Intel would be willing to sell.
That probably is figured into the calculus that went into how they extended the address space. It's a low-cost evolution, which probably means the expectations are that even if the prices are higher than is customary for an ARM, those combined with the small volume and competition are not going to justify a full transition to 64-bit at this point.

I agree. They're probably taking their time on extensions. PAE style extension is a lot simpler to approach. I still think something that extended AGU addressing a bit would be simpler/cheaper to implement than full-on 64-bit, but it does seem pretty hackish. I'm not sure this approach has been used anywhere. It would have fit pretty well with 68k's programming model.

Ok, someone make this simple for me. Does this mean that we'll be seeing uber cool/cheap/portable netbooks soon? :neutral:

Netbooks are already really cheap (can be found for ~$250 or less on the lower end) and portability is dictated by screen size and overall weight. As for cool, I'll leave that to you. For netbooks other components like large displays and hard drives tend to use up the power budget a lot more than in mobiles, so shaving a little bit off of CPU consumption doesn't get you as much.

If you're particularly interested in the netbook space you should look out for AMD's Bobcat CPUs, which may end up giving Atom netbooks a serious run for their money.

 

[kinshadow]

I read somewhere that ARM gets 1% in licensing, so if true it's proportional to chip cost and they won't lose out on server sales being lower volume.

The royalty rates are all vendor specific and the guys mentioned here all paid a premium to be lead partners.

How does ARM's business model work exactly? I mean, do they get significantly more if TI/Samsung/Qualcomm/Whoever sells an ARM server chip than if they sell a smartphone chip?

ARM doesn't really sell chips. ARM sells the processor core IP (RTL and hard macros ..... like a software script that defines hardware logic). Many other companies integrate the processor IP (along with a lot of other stuff, GPUs, video processors, DSPs, display controllers, etc.) into SoC chips. Any physical 'chip' that ARM makes is for development, etc. and isn't really mass market.

 

[aaronspink]

I read somewhere that ARM gets 1% in licensing, so if true it's proportional to chip cost and they won't lose out on server sales being lower volume.

In general the licensing costs are quite varied. Usually there is some base min price, a max price, and in between a % price based on the selling price of the chip. The min price may be as little as 10 cents while the max price may be as much as $10 while the percentage can be anything from .1% to 5%. A lot of this also varies based on min/prepaid fees as well. If you are willing to contract to a min royalty of 10mil per year, you are likely to get better min/max/variable cost than if you go with zero or 1mil upfront. Things also vary greatly depending on if you are licensing IP vs just the ISA (architectural license) and hard vs soft macros (hard macros you can end up paying the foundry vs arm directly depending on the various arrangements, sometimes the foundry will subsidize the macro cost in order to get business, it doesn't matter to arm because the foundry ends up paying their negotiated to ARM per part regardless).

 

[aaronspink]

ARM doesn't really sell chips. ARM sells the processor core IP (RTL and hard macros ..... like a software script that defines hardware logic). Many other companies integrate the processor IP (along with a lot of other stuff, GPUs, video processors, DSPs, display controllers, etc.) into SoC chips. Any physical 'chip' that ARM makes is for development, etc. and isn't really mass market.

I'm not even sure that ARM makes any chips per se at this point. Their designs are simple enough that they can synthesis to FPGAs and its likely that the foundries/partners do the hard macro work and testing.

 

[Squilliam]

How well would something like this compare to say a full blooded desktop processor such as an Athlon X2? I find it difficult to fathom the performance of handheld processors when all I can think about is relative desktop terms. For instance I guessed that the Ipad CPU was like a Pentium 3 800Mhz or so when my friend asked me, but I pulled that one out of my butt.

 

[darkblu]

How well would something like this compare to say a full blooded desktop processor such as an Athlon X2? I find it difficult to fathom the performance of handheld processors when all I can think about is relative desktop terms. For instance I guessed that the Ipad CPU was like a Pentium 3 800Mhz or so when my friend asked me, but I pulled that one out of my butt.

The currently-popular A8 design is more akin to P5 (original pentium) with MMX.. Actually more so to AMD's K6 with 3DNow. Of course, most A8's run at clocks which those x86 never achieved (incl. memory buses, etc), neither do they "feature" x86 legacy frontends. And by 'legacy' I mean something that should have been dead and buried decades ago.

 

[Exophase]

The currently-popular A8 design is more akin to P5 (original pentium) with MMX.. Actually more so to AMD's K6 with 3DNow. Of course, most A8's run at clocks which those x86 never achieved (incl. memory buses, etc), neither do they "feature" x86 legacy frontends. And by 'legacy' I mean something that should have been dead and buried decades ago.

How more akin to K6? That was an out of order design, more like P6. Is it because of the 2-way SIMD FPU?

Anyway, he asked how A15 would compare, not A8. A15 is two designs ahead of A8, and is 3-issue out of order instead of 2-issue in order, and who knows what other improvements. At +50% IPC over Cortex-A9 it might start looking around K8 level performance like asked (where clocks and core counts match, anyway, and minus 64-bit support)

 

[Squilliam]

Thanks people, that makes things a little more comprehendable.

 

[Flux]

In general the licensing costs are quite varied. Usually there is some base min price, a max price, and in between a % price based on the selling price of the chip. The min price may be as little as 10 cents while the max price may be as much as $10 while the percentage can be anything from .1% to 5%. A lot of this also varies based on min/prepaid fees as well. If you are willing to contract to a min royalty of 10mil per year, you are likely to get better min/max/variable cost than if you go with zero or 1mil upfront. Things also vary greatly depending on if you are licensing IP vs just the ISA (architectural license) and hard vs soft macros (hard macros you can end up paying the foundry vs arm directly depending on the various arrangements, sometimes the foundry will subsidize the macro cost in order to get business, it doesn't matter to arm because the foundry ends up paying their negotiated to ARM per part regardless).

in what bulk volume orders do you have to have if you want to get the lower price quotes? 100k-250k+?

 

[metafor]

How more akin to K6? That was an out of order design, more like P6. Is it because of the 2-way SIMD FPU?

Anyway, he asked how A15 would compare, not A8. A15 is two designs ahead of A8, and is 3-issue out of order instead of 2-issue in order, and who knows what other improvements. At +50% IPC over Cortex-A9 it might start looking around K8 level performance like asked (where clocks and core counts match, anyway, and minus 64-bit support)

3-issue OoOE is probably closer to a Pentium 3 after the cache changes. I seriously doubt they put a 9-way issue back-end in this thing like the K8 has.