Just discovered my posts about Titan were mentioned in this thread, and nobody knew what I was on about. I didn't give the whole story so I'll explain:

I attended a talk by Intrinsity at work where they talked about their technology and their involvement in Titan.

The idea behind Titan was to use the Intrinsity Fast14 technology on a cheap PPC. By using Fast14 they could make the chip fast and low power but they could also save money by using a 90nm process.

But, Titan wasn't an existing core with Fast14 added to it. It was a completely new core designed by Intrinsity under contract from AMCC.

When Apple bought Intrinsity I suspect Apple wanted the engineers to work on their projects and pretty much paid AMCC to drop Titan.

However since AMCC owned the design they could still port it to a different process, presumably without the Fast14 tech.

Titan was a specific implementation of the core, it got cancelled, and AMCC started work on a new implementation on IIRC 45nm.

So that why Titan was the core that was cancelled, but wasn't.

--

BTW The Fast14 tech is very impressive, it can pretty much double performance without increasing power consumption. Before they were bought, there were rumours of Intrinsity working on a 2Ghz Cortex-A9.

I suspect the iPad 2 might be rather more potent than anyone suspects

It's a pity Apple scarfed up the technology.
I had some suspicion that Apple's purchase of Intrinsity was behind the shift away from Titan.

Have you got any idea how close the new 45nm processor is to the Titan design?

[ Edited by Jim on 2011/1/19 22:44 ]

> Titan wasn't an existing core with Fast14 added to it. It was a completely
> new core designed by Intrinsity under contract from AMCC.

But why is it then that Titan is often connected to the PPC4xx family of cores?

https://morph.zone/modules/newbb_plus/viewtopic.php?topic_id=6268&forum=11&post_id=76580#76580
https://morph.zone/modules/newbb_plus/viewtopic.php?topic_id=6268&forum=11&post_id=79905#79905

Is it maybe just that Titan was designed to be compatible with the existing PPC4xx cores in supervisor mode, rendering it a new member of the PPC4xx core family technically? And then there's the clue that Titan was announced to deliver 2.0 DMIPS/MHz, just like PPC440 and PPC460 (and presumably PPC450) do.

> When Apple bought Intrinsity I suspect Apple [...] pretty much paid AMCC to drop Titan.

Yes, Applied Micro said that in April 2010 they had got their 5.4 million USD investment refunded by Apple. See last quote there:

https://morph.zone/modules/newbb_plus/viewtopic.php?topic_id=7001&forum=3&post_id=77002#77002

> since AMCC owned the design

As they got their investment refunded I would conclude that they don't own the design anymore. Any objections?

> they could still port it to a different process, presumably without the Fast14 tech.
> Titan was a specific implementation of the core, it got cancelled, and AMCC
> started work on a new implementation on IIRC 45nm. So that why Titan was
> the core that was cancelled, but wasn't.

So you think the 40nm PPC465 core based PacketPro/Mamba/APM86xxx is just a quick interim solution on the way to a processor based on a 40nm (not 45) implementation of the Intrinsity developed core? Admittedly, Applied Micro claimed in July 2009 that the Titan core could easily be migrated to smaller process nodes like 40nm. See last quote there:

https://morph.zone/modules/newbb_plus/viewtopic.php?topic_id=7001&forum=3&post_id=77005#77005

But then, this was before Apple had refunded Applied Micro's Intrinsity investment, and even before Apple had acquired Intrinsity.

> Before they were bought, there were rumours of Intrinsity
> working on a 2Ghz Cortex-A9.

Samsung's Hummingbird core used for instance in Apple's A4 chip is an Intrinsity (i.e. Fast14) enhanced Cortex-A8 core. And while there're already 2 GHz Cortex-A9 based chips in existence (e.g. by Nufront) I guess that a Fast14 enhanced 2 GHz Cortex-A9 would draw even less power.

Quote:

Indeed! Sparked by what you've said, I immediately found a document about it, and it's a fascinating read.

Andreas_Wolf wrote:

Quote:

Maybe it's designed to be compatible.

OTOH it could just be marketing. AMCC have the 460 label on different chips with 440 cores and others with 460 cores...

Quote:

I don't know the details so I wouldn't conclude anything.

I would imagine they own the core but not the 90nm version. Certainly $5.6 million isn't nearly enough to design a core.

The deal would have been much bigger and probably very complex so $5.6 million might have been a part payment for something they didn't get.


Quote:

Indeed.

> Maybe it's designed to be compatible.

Okay, that would be what I told I suspected regarding supervisor mode compatibility.

> OTOH it could just be marketing.

I doubt that, else debugging tools developers Abatron and Lauterbach wouldn't list it under PPC4xx I think.

> AMCC have the 460 label on different chips with 440 cores
> and others with 460 cores...

I'm aware of that. But still, subsuming any core/chip under 'PPC4xx' is yet another quality than just calling a PPC440 core based chip a 'PPC460'.

> Certainly $5.6 million isn't nearly enough to design a core.

So I guess 5.4 million wouldn't suffice all the more ;-) Thanks for the insight as I'm really not aware of how much money is usually involved in such undertakings.

Quote:

I'm surprised at this. I thought designing a CPU core, based on existing ARM technology, wasn't that complex. Due all respect, of course.

Or perhaps it's my misunderstanding, as it's not like ARM gives you some schematics, and then you tweak and build around them. Perhaps what ARM gives you is the instruction set and nothing else, and that's not at all like having the plans for a building.

Also, ARM might ask you for several million already in order to get a license for their ISA.

So, what does it take to make a CPU? Being it ARM makes it any easier? x86 maybe is? As I understand it would be a very lengthy explanation, pehaps there's something worth reading there that you could point me at.

By the way, congratulations Andreas: I've never seen a member here with such a "Yokemate of Keyboards" award! Are you the most posting user here? It wouldn't surprise me - and I'd say 90% of your posts are useful, that would count double than this post from mine.

Andreas and I have been discussing Freescale's reference designs for the P2020. Prompted by my complaint that you couldn't get he first board (it was already discontinued) and it's replacement wasn't available, our incredibly resourceful Wolfbot dug up this.

http://www.freescale.com/files/32bit/doc/brochure/PWRARCHQIQSG.pdf

The line you want to look at is under Reference Platforms where it mentions
"P2020/P2010 P2020RDPC $595 QorIQ P2020/P2010 Reference Design Board"

If you look this up on Freescake's website you'll find the P2020DS-PA Development System.

http://cache.freescale.com/files/netcomm/doc/fact_sheet/P2020DS.pdf?fsrch=1&sr=1

It comes complete with a hard drive with Linux loaded on it and appears to be a larger, better expanded version of the older boards.

http://www.freescale.com/webapp/sps/site/prod_summary.jsp?code=P2010RDB&fsrch=1&sr=1

The processor (a P2020) is a dual core E5500 derivative running at 800 to 1200Mhz. The boars has two PCIe slots, another slot referred to as an SGMI riser card slot (apparently related to networking), and what appears to be a PCI slot (which isn't mentioned in the documentation).
The board has the usual sound and USB ports and three 10/100/1000 Ethernet connectors as well as a connector for NAND Flash memory.
The board has one slot for DDR3 memory.

So $595, listed as for purchase with expected delivery in mid-April. This looks more powerful than the SAM460EX. While the P2020 lacks AltiVec instructions its E5500 cores should be more powerful than the SAM460EX (and its cheaper than any board from Acube).

I should have let Andreas post this for you guys, but after thinking about it overnight, I decided to go ahead (since he's sure to post corrections to any errors I've made and his observations).

So, what do you think?


[ Edited by Jim on 2011/1/21 12:15 ]

[ Edited by Jim on 2011/1/21 13:13 ]

> I thought designing a CPU core, based on existing ARM technology,
> wasn't that complex.

Titan uses Power ISA, not ARM ISA. As for the scale of costs, I think it depends on whether it's about a new core from scratch or about the enhancement of a pre-existing core. In case of Titan minator says it's a new core from scratch, thus more costly to develop, while in case of the Hummingbird core Intrinsity enhanced the pre-existing ARM Cortex-A8 core with their Fast14 technology, thus less costly to develop.

> Or perhaps it's my misunderstanding, as it's not like ARM gives you some
> schematics, and then you tweak and build around them. Perhaps what ARM
> gives you is the instruction set and nothing else, and that's not at all like having
> the plans for a building.

I already tried to explain the difference between an ISA license (here: Titan according to minator, albeit Power instead of ARM) and a core license (here: Hummingbird) to you, using the example of Marvell (seems I didn't succeed, unfortunately):

https://morph.zone/modules/newbb_plus/viewtopic.php?topic_id=6726&forum=11&post_id=78281#78281

So for developing Titan Intrinsity allegedly had only the Power "instruction set and nothing else", while for Hummingbird they had access to "the plans for a building".

> ARM might ask you for several million already in order to get
> a license for their ISA.

While that might be true (not only regarding ARM Ltd., but IBM as well) it doesn't apply here. Intrinsity developed Titan for Applied Micro, who have already had a Power ISA license (as well as a PPC4xx core license) from IBM, and they developed Hummingbird for Samsung, who have already had a Cortex-A8 core license from ARM Ltd. So no additional 3rd party licensing costs were involved in either case.

> congratulations Andreas: I've never seen a member here with such a
> "Yokemate of Keyboards" award!

Thanks. I think I got it by passing 2000 postings.

> Are you the most posting user here?

Second most actually, with Targhan being first and magnetic being third. But that's really not important to me as I don't post with the intention to raise my posting count. You're #9 btw ;-) You can see current stats there on the left:

https://morph.zone/modules/lastposts/

To hell with the number of posts. It his accuracy, recall, and the information content of his posts that both impress and scare me a little.

BTW - Its early in the morning on the East coast of the US and I've got a slight headache while I'm listening to Rob Halford belt out an old Judas Priest song "Dissident Aggressor"

A live version of this song on Judas Priest's A Touch of Evil album won the award for Best Metal Performance at the 2010 Grammys. It was the first time the band had won one of the coveted trophies.

Chorus

Stab! Bawl! Punch! Crawl!
Hooks to my brain are well in
Stab! Bawl! Punch! Crawl!
I know what I am, I'm Berlin


Appears to relate to WWII - intense song

[ Edited by Jim on 2011/1/21 13:25 ]

> The line you want to look at is under Reference Platforms where
> it mentions "P2020/P2010 P2020RDPC $595 QorIQ P2020/P2010
> Reference Design Board"
> If you look this up on Freescake's website you'll find the
> P2020DS-PA Development System.
> http://cache.freescale.com/files/netcomm/doc/fact_sheet/P2020DS.pdf?fsrch=1&sr=1

...which is not only larger (ATX instead of mITX) than the P2020RDB but also almost 6 times as expensive (P2020DS-PC just being a more recent revision of the P2020DS-PA).

> appears to be a larger, better expanded version of the older boards.
> http://www.freescale.com/webapp/sps/site/prod_summary.jsp?code=P2010RDB&fsrch=1&sr=1

It has nothing to do with "older" or "newer". Freescale basically offers two lines of boards for the P2020/P2010, one line being the reference boards at 595 USD called P2020RDB and the other line being the development boards at 3500 USD called P2020DS. Both lines exist in parallel, each having undergone several minor revisions so far ('PA' to 'PB' to 'PC').

> The processor (a P2020) is a dual core E5500 derivative

No, the P2020's cores are e500v2, not e5500. Where did you get that misinformation from?

> The boars has two PCIw slots, another slot referred to as an SGMI riser card slot
> (apparently related to networking), and what appears to be a PCI slot (which isn't
> mentioned in the documentation).

From the P2020DS block diagram I conclude the PCIe x2 slot and the SGMII slot to work mutually exclusive, as well as the PCIe x1/x2 slot and the M1575 (including Audio, (e)SATA and the PCI slot) to work mutually exclusive (in a sense that you could either operate the slot at x2 speed and have the M1575 turned off, or operate the M1575 and that slot at x1 speed each at the same time).

> So $595 [...] its cheaper than any board from Acube

No, 3500, which is multiple times the cost of the Sam460ex.

> This looks more powerful than the SAM460EX. [...] its E5500 cores
> should be more powerful than the SAM460EX

The P2020's e500v2 (not e5500) cores don't have a proper FPU, which the PPC460EX has:

https://morph.zone/modules/newbb_plus/viewtopic.php?topic_id=6196&forum=11&post_id=62074#62074

> I should have let Andreas post this for you guys

I wouldn't have posted this since on 'our market' I don't see the point in a CPU lacking proper FPU at all. Furthermore, I don't see the point in a 3500 USD board for our purposes.

> he's sure to post corrections to any errors I've made

You bet :-)

> what do you think?

I think you got everything confused that can possibly be confused ;-)

Almost everything. Yes, definitely screwed that up. I should have paid more attention to your responses.

Hey at least I got the form factor difference right. And as you pointed out to me several hours ago (if I'd bothered reading it carefully enough) the e500 cored products aren't worth consideration.

Pretty foolish mistaking P2020 and P2010 for P5020 and P5010.

Maybe I shouldn't mixed migraine medication, heavy metal and sleeplessness with bad posts.

Anyway, you've got more patience with my stupidity than I do today.

BTW - Where are the e5500 cored processors?

[ Edited by Jim on 2011/1/21 13:33 ]

[ Edited by Jim on 2011/1/21 13:46 ]

> the e500 cored products aren't worth consideration.

True for e500 through e500v2, yes. The e500mc has a proper FPU (albeit half-clocked), though.

> Where are the e5500 cored processors?

According to Freescale's June 2010 announcement they should be sampling already and go into volume production in second half of 2011.

> Unfortunately, no announcement of specific AltiVec enabled QorIQ chips in 2010 :-/

That didn't prevent Power.org from retrospectively telling the following about the year 2010:

"Freescale [...] incorporated AltiVec technology into its QorIQ family of multicore products"
http://www.power.org/news/pr/view?item_key=2c310e9c78c16f9b85b23f0c08ec4e4acf1778ef

Quote:

It's incredibly complex designing any core but it'll be a lot simpler than trying to design an x86.

It is a frighteningly expensive business:

You need:
1) For an ARM core you'll need an architecture license, I don't know the price but I once heard a rumour of $20 million.
2) A group of very clever and experienced people. They're hard to find and not cheap.
3) A very large computer to run the EDA software on. So large it'll be in the Top500 supercomputer list.
4) EDA software. All of it fear inducingly expensive (it's typically priced in the hundreds of thousands.) ...and you need lots and lots of it.
5) Time - it takes years to design a core. We might see what the PA-Semi team at Apple have been up to next year.
6) Cell libraries - you're not going to design you're own transistors so you buy designs for them.
7) Silicon. A new core will require several revisions at a millions each (I suspect the $5.4 million was for a new revision). It normally takes 3 months to make a chip so getting it right takes quite some time - at least a year.

It depends on the complexity of course but don't expect much (if any) change out of $100 million. Intel and AMD and the likes of Nvidia each spend around $500 million to design a new core.

- All that before they go into production.


While there are a number of companies that design ARM cores, ARM makes most of it's money from selling pre-designed cores. They're not exactly cheap but it's a fraction of the price of designing your own core. You also pay a royalty per chip you make. It's not much per chip but at over a billion per month it adds up.

ARM also do most of the stuff that goes around the core - bus systems, memory controllers and other cores like GPUs etc, etc. This gets used on all sorts of chips, not just ARM based.

> We might see what the PA-Semi team at Apple have been up to next year.

AFAIK it has been established that the P.A.Semi team designed Apple's A4 SoC around Samsung's/Intrinsity's Hummingbird core.

Update:

>> Unfortunately, no announcement of specific AltiVec enabled QorIQ chips in 2010 :-/

> That didn't prevent Power.org from retrospectively telling the following about the year 2010:
> "Freescale [...] incorporated AltiVec technology into its QorIQ family of multicore products"
> http://www.power.org/news/pr/view?item_key=2c310e9c78c16f9b85b23f0c08ec4e4acf1778ef

Jim pointed me to a Freescale PDF file from November 2010 containing some more information about AltiVec enabled QorIQ. This made me search for more and the search came up with these results (all files from November 2010):

http://www.freescale.com/files/training/doc/dwf/AMF_NET_T0425.pdf (pages 14/15)
http://www.freescale.com/files/training/doc/dwf/AMF_NET_T0610.pdf (page 71)
http://www.freescale.com/files/training/doc/dwf/AMF_ENT_T1117.pdf (pages 15 and 17)
http://www.freescale.com/files/training/doc/MULTICORE_MORE.pdf (page 32)
http://www.freescale.com/files/training/doc/POWERQUICC_TO_QORIQ.pdf (pages 50/51)

So it seems that both cores e5500 (QorIQ P5) and e500mc (QorIQ P4, P3 and some P2) will have variants with AltiVec, i.e. e5500 + AltiVec comes first and e500mc + AltiVec after that.

Addendum:

> I've yet to find numeric floating point performance figures for the e5500. All I
> know is that compared to the e500mc the e5500 has twice the single precision
> performance per clock and up to quadruple the double precision performance
> per clock. But I don't have any real numbers for the e500mc either.

Now I have:

http://www.freescale.com/files/training/doc/MULTICORE_MORE.pdf (page 28)
http://www.freescale.com/files/training/doc/POWERQUICC_TO_QORIQ.pdf (page 47)

e500mc: 1 MFLOPS/MHz in SP; 0.5 MFLOPS/MHz in DP
e5500: 2 MFLOPS/MHz in SP; 2 MFLOPS/MHz in DP


On the matter of DMIPS/MHz figures I find it odd that the two documents have contradicting figures: table in 1st linked file has the well known ones of 2.5 for e500mc and 3.0 for e5500, whereas table in 2nd linked file has figures of 2.3 for e500mc and 2.4/2.9 for e5500 (and lacks information why for e5500 there're two figures instead of just one).

Update:

> the future QorIQ T series

As of November 2010 there's a fresh roadmap on QorIQ T series:

http://www.freescale.com/files/training/doc/dwf/AMF_NET_T0425.pdf (page 7)

T5: T5040, T5020
T4: T4160, T4120
T3: T3080, T3060
T2: T2040

According to this roadmap depiction, QorIQ T series will be manufactured in a 28 nm process node instead of the previously announced 32 nm.
Moreover, I suspect they'll all have AltiVec.

(Besides, new on the roadmap from QorIQ P series: P4060)

It's now available to order:-

http://amigakit.leamancomputing.com/catalog/EUR.php?url=product_info.php?products_id=1014

|||

[ Edited by Velcro_SP 19.04.2011 - 15:08 ]

> MorphOS on some old PPC Mac would probably be just as -- I'd like to
> sees some comparative benchmarks sometime.

https://morph.zone/modules/newbb_plus/viewtopic.php?topic_id=7001&forum=3&post_id=78955#78955
https://morph.zone/modules/newbb_plus/viewtopic.php?topic_id=7001&forum=3&post_id=80157#80157

Update:

> The PSP2 will most likely be nVidia Tegra 2xx (ARM Cortex-A9 core) driven.
> http://www.semiaccurate.com/2010/07/14/sonys-psp2-powered-nvidias-tegra-line/

ARM Cortex-A9 in the PSP successor ("NGP") has been confirmed by Sony:

http://www.scei.co.jp/corporate/release/110127a_e.html

As the GPU is announced to be the SGX543MP4+ (i.e. non-nVidia) the CPU can't be Tegra, though.

Addendum:

>> what is the impact of the missing Altivec instructions
>> (in particular in regard to MorphOS).

> That question is best answered by a MorphOS Team member.

...who now did on another site:

http://www.amiga.org/forums/showpost.php?p=610058

...and another one on yet another site:

http://amigaworld.net/modules/newbb/viewtopic.php?topic_id=33358&forum=28&start=40#605540

[ Edited by Andreas_Wolf on 2011/3/15 16:58 ]

Quote:

Pre-order. Not available yet.