Quote:

If it was POWER6 based it would not really be a POWER6 in any meaningful way. POWER6 added all sorts of on-chip stuff that is completely useless for a console.

I think it's more likely to be based on the A2 used in the wire-speed chip, the A2 appears to be based on the PPE from Cell / Xenon.

POWER6 is better on general purpose stuff than the PPE but that's not very important in games.

The FLOPS numbers are weird, it works out at 12.7 instructions per cycle. Rather more likely is a slightly different clock speed. At 3.66GHz it works out to be 12 instructions per cycle - the same as the PPE.

The double precision FLOPS are the same but why on earth would you want such DP power in a console?

Rest of it sounds fairly plausible.

Quote:


I believe they're quite similar clock for clock. However it'll very much depend on the implementation. A9s tend to have bigger caches and faster memory than the G4s ever had.

Quote:

I meant the lower range of the G5s but remember the G4 was faster per clock than the G5.

Quote:

Sure, what I meant was the Qualcomm chip at 2.5GHz. It's sampling now so should be a year (or less) away. It could get close or beat the G5s.

OTOH It looks like Nvidia's "Project Denver" should blow any G5 clean into next week.

BTW most (or all?) ARMs are mode on low power (LP) or general purpose (GP) silicon processes. Nobody makes them on a High Performance process. If they did, clock rates would soar. With all the talk of servers I wonder will anyone actually do that.

> If it was POWER6 based it would not really be a POWER6 in
> any meaningful way. POWER6 added all sorts of on-chip stuff
> that is completely useless for a console.

The spec listing might refer to a custom-built chip with cores derived from the cores of the POWER6 chip so that POWER6's all sorts of on-chip stuff wouldn't play any role anyway.

> the A2 appears to be based on the PPE from Cell / Xenon.

Btw, does the A2 have AltiVec/VMX?

> why on earth would you want such DP power in a console?

Probably for the same reasons Cell's PPE and Xenon's core also have this DP power per clock and core ;-)

> I believe they're quite similar clock for clock.

Okay, so accordingly G4 above 1.2 GHz should still be faster than current Cortex-A9.

> A9s tend to have bigger caches and faster memory than the G4s ever had.

Faster memory bus yes, but bigger caches than G4s ever had? While the core is specified by ARM for up to 4 MiB L2 cache I couldn't find any actually existing Cortex-A9 with more than 1 MiB L2 cache. MPC7448 has 1 MiB L2 as well. MPC745x has less L2 but can have L3, e.g. MPC7457 has 512 KiB L2 and up to 2 MiB L3.

> I meant the lower range of the G5s

That's what I referred to with "1.6 GHz G5 (the slowest of the bunch)" as that is the very bottom of the G5 range. I still doubt that a 1.8 GHz Cortex-A9 would beat it.

> remember the G4 was faster per clock than the G5.

I don't think that's generally true except for AltiVec/VMX code.

> what I meant was the Qualcomm chip at 2.5GHz.

Ah okay, so that was a misunderstanding from my side, sorry. A 2.5 GHz Krait could possibly beat a 2.5 to 2.7 GHz G5, yes.

> It's sampling now so should be a year (or less) away.

I'm not sure Krait based chips are really sampling at 2.5 GHz already. Some sources say the MSM8960 will only clock up to 1.2 GHz and it will take until the APQ8064 to actually reach 2.5 GHz:

http://www.electronista.com/articles/11/06/02/qualcomm.dual..and.quad.core.processors.inbound/

> It looks like Nvidia's "Project Denver" should blow any G5 clean into next week.

Sure but that's even further away than Cortex-A15.

Quote:

Don't know but they could always add it.

Quote:

IIRC the PPE could do 2 Double Precision instructions per cycle. According to these numbers this core can do 7! The only PPC I know of that can do that is POWER7 (it can do 8 actually).

IIRC some of the PPCs used by Nintendo are standard parts with enhanced vector units.
I wonder could this be something like a set of A2s with a POWER7 style vector units bolted on the side. That makes a lot more sense.

Update:

>>> this started circulating 5 days ago:
>>> http://img12.imageshack.us/img12/1228/prcspec13iu.jpg

>> How accurate do you think that document is? Could it really
>> be a Nintendo internal document or just an elaborate fake?

> We'll know for sure in a few days I guess.

We now know that this sheet must be hoax as it talks about a "32nm process" whereas IBM has to say:

"IBM plans to produce millions of chips for Nintendo featuring IBM Silicon on Insulator (SOI) technology at 45 nanometers (45 billionths of a meter)."
http://www-03.ibm.com/press/us/en/pressrelease/34683.wss

No information on core count (just "multi-core") or frequency, unfortunately. And Nintendo doesn't reveal any better info either:

"CPU: IBM Power®-based multi-core microprocessor."
http://e3.nintendo.com/hw/#about

So let's wait and see if the older triple-core rumour turns out true.

>> does the A2 have AltiVec/VMX?

> Don't know but they could always add it.

Yes, of course they could, but what I would like to know is whether IBM's stock A2 core (as used in the PowerEN and Power BQC chips) has it. I ask because I've read conflicting third party statements on that matter (and no statement by IBM itself).

> IIRC some of the PPCs used by Nintendo are standard parts with
> enhanced vector units.

Yes, Gekko is PPC750CXe plus custom vector unit, and Broadway is PPC750CL (which already includes the same vector unit).

> I wonder could this be something like a set of A2s with a
> POWER7 style vector units bolted on the side.

With "POWER7 style vector units" you mean VSX? Thing is Nintendo claims the Wii U to be backwards compatible to Wii. Assuming there's no real emulation involved this means the vector unit of the new CPU must be either identical to or a superset of Broadway's vector unit. This would rule out VSX (as well as any other standard vector unit like AltiVec/VMX). On the other hand, there could always be more than one, but I doubt it ;-)

@minator

ARM for the future!

@ takemehomegrandma

If these specs are even close to accurate, this new PPC is much more powerful then any announced ARM.

> If these specs are even close to accurate

What specs? I don't think we have any for this Power Architecture chip yet, except that it has more than one core and is manufactured in a rather conservative build process. Or did I miss anything? Or do you refer to the debunked hoax spec sheet here?

Quote:

Only high performance high end stuff tends gets built on the latest nodes. The high volume and especially low cost stuff is usually a generation behind. Everyone does this, even Intel.

It's also easier to build something new on an existing proved process than build it on a newer ramping process. You'll probably find it'll move to 32 or (more likely) 28 nm once they've ramped properly.

Quote:

Some of it is clearly wrong, but some (e.g. eDRAM) is clearly correct.
It's not that adventurous anyway, the important number - single precision floating point - is lower than the PS3's Cell.

> Some of it is clearly wrong, but some (e.g. eDRAM) is clearly correct.

If some of it is wrong the spec sheet as such is wrong and can't be a genuine Nintendo spec sheet I think. Should some details match however I'd regard it as coincidences or educated guesses by the hoaxer at best.

> Key technical features of Freescale's newly announced i.MX6 CPU's include:
> [...]
> - Multistream-capable HD video engine delivering 1080p60 decode,
> 1080p30 encode and 3D video playback in HD
> - Exceptional 3D graphics performance with quad shaders for up to 200 MTPS
> - Separate 2D and vertex acceleration engines for uncompromised user
> interface experiences

Vendor of the GPU cores announced two months ago:

http://www.vivantecorp.com/fsl.html

From Freescale FTF:

Freescale Demonstrates Breakthrough Quad-Core ARM(R) Applications Processor

"Freescale Semiconductor (NYSE: FSL) kicked off the 2011 Freescale Technology Forum Americas in high energy fashion today by demonstrating one of the industry's highest-performance quad-core applications processors. The forum's opening keynote presentation featured a live demonstration presenting the multimedia capabilities of Freescale's i.MX 6Quad applications processor."

http://media.freescale.com/phoenix.zhtml?c=196520&p=irol-newsArticle&ID=1576641&highlight=

Edit:
Video of this 7 days old silicon here: http://www.youtube.com/watch?v=p0lkZlDTq8Q



[ Edited by takemehomegrandma 22.06.2011 - 18:55 ]

> I actually think my Efika MX would kick at least Sam440's butt.

Found some NBench figures of both to compare.

Efika MX: http://www.powerdeveloper.org/forums/viewtopic.php?p=14836#14836
Sam440: http://amigadev.free.fr/powerpc/nbench.html

Result scores (for 800 MHz each):

Efika MX:
- MEMORY: 4.87
- INTEGER: 5.15
- FLOATING-POINT: 1.22

Sam440 (scaled from 553 MHz | scaled from 667 MHz):
- MEMORY: 2.54 | 2.52
- INTEGER: 5.13 | 4.46
- FLOATING-POINT: 3.56 | 3.56

>>> Currently the A9 is in PPC G4 territory performance wise.

>> which clock rate in G4 do you think a 1.2 GHz (which I think is the
>> current max) Cortex-A9 roughly resembles performance-wise?

> I believe they're quite similar clock for clock.

Also here, found some NBench figures of both to compare.

Cortex-A9: http://www.powerdeveloper.org/forums/viewtopic.php?p=14836#14836
G4: http://amigadev.free.fr/powerpc/nbench.html

Result scores (for 1.0 GHz each):

Cortex-A9 (one core of dual-core OMAP4):
- MEMORY: 6.52
- INTEGER: 6.19
- FLOATING-POINT: 6.82
- BYTEMARK INTEGER: 25.4
- BYTEMARK FLOATING-POINT: 12.3

G4 (scaled from 400 MHz | scaled from 1.4 GHz | scaled from 1.5 GHz*):
- MEMORY: 7.49 | 5.76 | 4.49
- INTEGER: 6.69 | 8.27 | 5.76
- FLOATING-POINT: 9.24 | 6.47 | 9.02
- BYTEMARK INTEGER: 20.7 *
- BYTEMARK FLOATING-POINT: 16.3 *


* my machine

Nothing too surprising there. I do expect the G4 will be better at floating point - FP has never been ARM's strong point (the Cortex-A15 should be a lot better).

That said a 15 year old benchmark is probably not a great test - can you run CoreMark?

> can you run CoreMark?

I guess it could be compiled and made to run on Leopard. But that requires registering with EEMBC, which I won't do.

Update:

> I'm not sure Krait based chips are really sampling at 2.5 GHz
> already. Some sources say the MSM8960 will only clock up to
> 1.2 GHz and it will take until the APQ8064 to actually reach 2.5 GHz

New info:
http://www.mobiletechworld.com/2011/07/05/new-qualcomm-2011-2012-roadmap-and-soc-specifications/

Apparently, my doubts were justified and 2.5 GHz Krait will indeed not come until APQ8064 scheduled for 2012 at earliest, but MSM8960 scheduled for Q4/2011 is intended to clock between 1.5 and 1.7 GHz, not just 1.2 GHz which was claimed by the source I quoted.

Update:

>> Probably the fastest ARM around right now CPU and GPU wise

> Yes, it's 20% higher clocked than the SoC on the Snowball board

Apparently, they decreased clock frequency from the annnounced 1.2 GHz to only 1.0 GHz as well:

"Samsung Exynos4210 Cortext-A9 dualcore 1.0GHz"
http://www.origenboard.org/about.php

> http://www.raspberrypi.org/ [...]
> More tech details, performance and such would interesting though.

The SoC has just recently been revealed as the BCM2835:

http://www.raspberrypi.org/?p=106
http://www.raspberrypi.org/?page_id=43&mingleforumaction=viewtopic&t=36.3

In other news, this is from GlobalFoundries:

Quote:

Full story here.

> Video of this 7 days old silicon here: http://www.youtube.com/watch?v=p0lkZlDTq8Q

"Freescale unveiled its i.MX 6 product family in end-June, and revealed the new processors are available in sample and volume quantities."
http://www.digitimes.com/news/a20110831PD212.html

From first silicon to volume production in 11 weeks? Doesn't sound like Freescale at all to me ;-)

Update:

>>> does the A2 have AltiVec/VMX?

>> Don't know but they could always add it.

> Yes, of course they could, but what I would like to know is
> whether IBM's stock A2 core (as used in the PowerEN and
> Power BQC chips) has it. I ask because I've read conflicting
> third party statements on that matter (and no statement by IBM itself).

http://www.theregister.co.uk/2011/08/22/ibm_bluegene_q_chip/

As that detailed article doesn't mention AltiVec/VMX for the A2 core I assume that it's not there.

Update:

> I compiled a small list of DMIPS per MHz and core figures for various
> recent and future cores implementing ARM ISA:
>
> ARM Cortex-A8: 2.0
> Qualcomm Scorpion: 2.1
> Marvell Sheeva PJ4: 2.4
> ARM Cortex-A9: 2.5
> Qualcomm Krait: 3.0...3.5 (estimated)
> ARM Cortex-A15: 3.5 (estimated)

"A new pipeline architecture increases the performance of Krait by over 60% compared to Qualcomm's existing Scorpion CPU micro-architecture."
http://www.qualcomm.com/documents/files/snapdragon-s4-processors-system-on-chip-solutions-for-a-new-mobile-age-white-paper.pdf (page 3)

Going by that, Qualcomm's Krait core could deliver as much as about 3.4 DMIPS per MHz and core.