More PowerMac G5 Details...

[sinisterdesign]

so Apple FINALLY gets all the great press that they deserve, all the wonderful writeups in numerous magazine's & endless tech articles praising OSX, iPod and steve's turnaround of the company aaaaand now...we have nothing new to show. 😕

it's not so much the GHz war that really chaffs me (although i DID have high hopes for IBM and was so excited that we partnered w/ them), it's just the thought of paying top dollar for more mediocre technology. stale grx cards, gorgeous LCDs w/ so-so resolution, stupid 1-button mice, etc, etc.

i have a few switchers here around the office, most of them just grew tired of windows probs and decided to try the now "cool" Mac, but on more than one occasion i hear, "why is _____ so slow on my new Mac? it seemed much faster on my old PC laptop."

i feel like i'm having to defend the average performance of the machines that i love and have worked on for years. i'm tired of that. why can't i say, "yeah, see how F%*#&ING FAST it is on your new Mac??!? smokes the S#!+ out of that Dell you used to use, huh?"

when can i say that, steve?? huh?

 

[ACED]

If you deliberately misinform staff who are bound by an NDA to not go outside the company then I think you would find it very hard to prosecute the company, especially if the intention is to uncover persons known to be leaking secrets to the market.

The moment "you deliberately misinform staff who are bound by an NDA to not go outside the company" you are releasing information into the public domain ... particularly if there is established wide-spread suspicion and a realistic chance that the misinformation WILL enter the market, AND consequently, influence share prices. If they break their NDA agreement, thats entirely another matter, which shouldn't justify the means, by flagantly abusing the public arena, with reckless misinformation.

Besides anyone adversely affected by such a release, could sue Apple for deliberately mis-using the public forum to expose errant employees, IMHO.

 

[Evangelion]

The moment "you deliberately misinform staff who are bound by an NDA to not go outside the company" you are releasing information into the public domain

No, that would be confidential company-information being told to someone inside the company. If you told the same info in a keynote to a wider public, then you would be releasing it in to "public domain".

If they break their NDA agreement, thats entirely another matter, which shouldn't justify the means, by flagantly abusing the public arena, with reckless misinformation.

THEY are not "abusing the public arena". The ones who leak the information outside the company might be. The third party that reports the info might be. But company telling something confidential to their employees (who are under an NDA) are NOT "abusing public arena". I think you have pretty weird ideas as to what is "public" and what is not.

Besides anyone adversely affected by such a release, could sue Apple for deliberately mis-using the public forum to expose errant employees, IMHO.

Uh, no. Apple didn't release any info to the public, I fail to see how they could be held accountable.

 

[nexusfx]

Awake again.....

You talk about x86 (instruction-set), and then talk about how eliminating northbridge was a great idea. What does northbridge have to do with the CPU's instruction-set? And then you talk about specific CPU-architectures (pipeline-stages). Again, that has nothing to do with the instruction-set. And, IIRC G5 has pretty long pipeline as well. Number of pipelines is more a feature of particular CPU-families and not the instruction-set. Athlons have alot less pipeline-stages than P4 does, yet they are both x86 (A64 is x86-64 as well)

Of course AMD 64 is x86, every Wintel Processor is. Eliminating the Northbridge allowed for smaller pipeline stages on the Athlon 64's x86 based chipset.

And then you talk about specific CPU-architectures (pipeline-stages)

Pipeline Stages are "part" of a systems architecture not the whole thing. The G5's specifically go through 9 stages, the G4's actually went through 7, that doesn't make the G4 faster, it's the rest of the whole that makes the Power4/G5 such a signifigant improvement in speed.

I don't care how "crappy" x86 is. I care about the performance. why is it that Pentium/Athlon64 (with their "crappy" instruction-set), compete just fine with G5 (with it's "superior" instruction set)?

I explained this in another post, the software does not fully take advantage of the System Architecture. First off, none of it is fully 64 bit, secondly some software doesnt even take advantage of Dual processing, that's two KEY components of the G5's speed right there. There are plenty more examples but you get the idea.

You talk about bandwidth. Funny, Athlon64 is alot smarter when it comes to bandwidth that G5 is. In G5, when the CPU wants to talk with the RAM, it first talks with the northbridge (sharing the FSB-bandwidth with other devices), northbridge talks with the RAM, RAM then talks with the northbridge which then talks with the CPU. That wastes bandwidth and it takes a long time. Contrast that to Athlon64, where the CPU talks directly with the RAM using a dedicated channel, leaving the FSB to other devices.

First, the FSB on any system shares bandwidth with other devices, unlike the Athlon 64 however, the Power4 System controller prevents that data from fighting for bandwidth across that FSB by giving each subsystem a dedicated throughput to the CPU and Main Memeory. NOTHING FIGHTS FOR BANDWIDTH. Second,

northbridge talks with the RAM, RAM then talks with the northbridge which then talks with the CPU

replace northbridge with Power4 and you get

CPU talks directly with the RAM using a dedicated channel, leaving the FSB to other devices.

except on the G5 side, the dedicated channel is running the full speed of the FSB and is running seamlessly within the system, unlike the Athlon 64's dedicated channel which runs at a slower speed and adds another pipeline stage to the overall proccessing of data.

If you have any information conrtadicting what I said please share it, if I'm mistaken about something I'd like to know. Wow this is alot of post's for me in one night 🙂

 

[aswitcher]

The moment "you deliberately misinform staff who are bound by an NDA to not go outside the company" you are releasing information into the public domain ...

No your not...your releasing information to a select few internal company personnel. Its a loyalty test. These things happen in government and in other forms in sciety all the time. Your argument followed to the logical extreme would suggest that no "testing" of staff loyalty can take place, and that even more drastic trust related actions like police stings are illegal.

Apple have a right to protect their IP and manage their own release schedule to make the best of their business. A loyalty test / internal "sting" operation on their part would (if it is) be being done for the best of the companies and shareholders long term interests. If they believe the risk is worth it then its not different from other busienss and marketing decisions. Breaking an NDA is illegal. No one forces anyone in the company to do it but if they do and are caught then they have broken a fair agreement that they entered into and hence wear the consequences. That Apple try to protect their corporate interests by providing disinformation is no crime - its being doen for the companies good.

 

[AidenShaw]

not talking about "consumer" machines here, right?

Look on the bright side, by the time consumer windows software actually takes advantage of the second core, the 970MP should be out.

LOL - what a case of "the Emperor has no clothes"....

You realize that 87% of the Macs sold are single processors, and most of those are G4 systems!

For the pro apps, those have been running on multi-processor Xeons for quite some time - they'll take advantage of dual-core or dual-cpu without any issues.

Don't forget that HyperThreading has made "multi-processor" aware programs useful even on the typical higher-end single CPU Intel box - many of the more power-hungry "consumer" apps have been taking advantage of HT.

And, BTW, one of the most power hungry "consumer" apps on Windows is the free Windows Media Encoder (also the encoder piece of Windows Media Player). This app supports at least 4 CPUs currently.

When I rip a CD, it often sits above 100% for CPU utilization - and ripping is CDROM limited. When re-encoding (for example, converting the CDs from lossless format to VBR WMA for the MuVo), it keeps all 4 CPUs on my HyperThreaded Xeon busy.

Also, don't forget that the Pentium D will soon be out - a much cheaper dual core (cheaper than the current Extreme Edition dual-core) so that the price premium for dual-core will be very small.

By the end of the year, dual-core will be through out the Intel and AMD product lines. There have even been sightings of early samples of the dual-core "Centrino" laptop chip - so get ready to see the adverts for dual CPU laptops before too long.

 

[Hattig]

And if you look at Opteron.... It started at 1.4Ghz in April 2003 and it's today at 2.6Ghz 🙂. And it too has received mem-controller improvements and the like (although no SSE3 yet IIRC). So it has had increase of 85% in two years.

The launch Opterons were 1.4 GHz, 1.6GHz and 1.8GHz, so 1.8 to 2.6 is 44% faster. That's a good record though comparitively, and given that we are talking about PowerMacs here, a reasonable comparison point.

The G5 has had FSB improvements too of course, and we don't know if the IBM made northbridge has had any memory controller improvements (were the first PowerMacs PC2700 or PC3200?). The 2.6GHz Opteron will have SSE3 as it will be based upon the new 90nm Rev. E core. It will be released around the same time as these new PowerMacs give or take a few weeks.

Oh, if you want to feel better, go and price up a dual processor Dell Precision 670 Workstation 😛

 

[duffman9000]

so Apple FINALLY gets all the great press that they deserve, all the wonderful writeups in numerous magazine's & endless tech articles praising OSX, iPod and steve's turnaround of the company aaaaand now...we have nothing new to show. 😕

it's not so much the GHz war that really chaffs me (although i DID have high hopes for IBM and was so excited that we partnered w/ them), it's just the thought of paying top dollar for more mediocre technology. stale grx cards, gorgeous LCDs w/ so-so resolution, stupid 1-button mice, etc, etc.

i have a few switchers here around the office, most of them just grew tired of windows probs and decided to try the now "cool" Mac, but on more than one occasion i hear, "why is _____ so slow on my new Mac? it seemed much faster on my old PC laptop."

i feel like i'm having to defend the average performance of the machines that i love and have worked on for years. i'm tired of that. why can't i say, "yeah, see how F%*#&ING FAST it is on your new Mac??!? smokes the S#!+ out of that Dell you used to use, huh?"

when can i say that, steve?? huh?

Yeah, its frustrating. People ask me all the time should they switch. I ask them why switch if they're happy with the speed of their computer since they would most likely be switching to a slower mac. Also, if they're satisfied with windows why should they switch? Not everyone has the cash to buy a G5 and for office work almost any cheapo pc will smoke a mac in price and speed.

 

[ACED]

It has to do with intent...

If Apple released it with the belief that the misinformation would end up on TS, that is a deliberate release, based on its proir knowledge of a definite leaks in the past...prior knowledge. Its an 'established' channel.

It would be reckless of Apple, and I don't think they would risk it. It would also be a PR disaster, if this was revealed as its intentions.

AN attitude of "to hell what the public thinks of the misinformation, we're prepared to do this to get a rat" would not be acceptable, and readily contestible, if consequential loses occurred

 

[Hattig]

Of course AMD 64 is x86, every Wintel Processor is. Eliminating the Northbridge allowed for smaller pipeline stages on the Athlon 64's x86 based chipset.

Please stop commenting, it makes you look silly and what you are saying is nonsense.

The internal make-up of an A64 is this:

[core 1]---[SRQ]---[Crossbar]---[Dual Channel Memory Controller]
[core 2]-----+ . . . . . +---[Hypertransport]

The core is what does all the processing, the pipeline length is not affected by having the northbridge on-die at all, that's just stupid. What incorporating the northbridge does allow for is extremely low latency memory access, and the Hypertransport link to the chipset will have much lower contention because it doesn't carry memory access traffic except for chipset generated memory traffic (which is the same case on all chipsets, so a moot point).

 

[mad jew]

Sorry for keeping the thread off topic but I don't see how the law has to be involved at all. Apple isn't after the financial benefits ( 🙄 ) of suing the culprit, they just want to relocate him/her so that information isn't leaked anymore. This may mean retrenchment, but I think that's pretty unlikely considering the points everybody has raised about the legalities. If Apple just feeds misinformation to the leak, and that gets published by Thinksecret, it just means they can internally rearrange things so it won't happen again.

Back on topic:

If they don't bring out new iMacs soon I'm just gonna grab an old eMac as a stop gap. I desperately need a desktop that can be kept on 24/7 and to be honest, speed isn't a huge factor, but it'd be nice to have a 128Mb video carded G5. 🙂

 

[le_coc]

Can ANYONE tell me what time normally the updates are?

Cause I think it won't happen today anymore?

 

[Evangelion]

Of course AMD 64 is x86, every Wintel Processor is. Eliminating the Northbridge allowed for smaller pipeline stages on the Athlon 64's x86 based chipset.

Uh, those things are not related to each other in any way at all! Eliminating the northbridge (or rather: having integrated mem-controller) allowed for better mem-bandwidth and lower-mem-latencies. It had no effect on the number of pipeline-stages. it doesn't really matter to the core how it talks with the RAM. Only difference the core will see, is that with integrated mem-controller, accessing the RAM is alot faster.

Pipeline Stages are "part" of a systems architecture not the whole thing. The G5's specifically go through 9 stages, the G4's actually went through 7, that doesn't make the G4 faster, it's the rest of the whole that makes the Power4/G5 such a signifigant improvement in speed.

Actually, G5's pipeline is up to 23 stages long. And that's longer than on the A64 for example. If G4 had faster FSB, it would rival G5 in performance on many instances.

I explained this in another post, the software does not fully take advantage of the System Architecture. First off, none of it is fully 64 bit, secondly some software doesnt even take advantage of Dual processing, that's two KEY components of the G5's speed right there. There are plenty more examples but you get the idea.[/quote9

And very little of Mac-software is 64bit either. Hell, even the OS isn't 64bit! And besides, just about all software on my Athlon64-machine IS 64bit, including the OS.

And the systems are dual-processor aware. If you have more than one process running, you will benefit from dual-processing.

[quote9First, the FSB on any system shares bandwidth with other devices, unlike the Athlon 64 however, the Power4 System controller prevents that data from fighting for bandwidth across that FSB by giving each subsystem a dedicated throughput to the CPU and Main Memeory. NOTHING FIGHTS FOR BANDWIDTH.

If you have X amount of bandwidth, and your need for bandwidth is X + 20%, you can be damn sure that there will be a "fight for bandwidth". And you are talking about Power4. Does it relate to G5?

And on the A64/Opteron, the devices (IDE, PCI, AGP, USB etc. etc.) have a dedicated 1GHz HyperTransport link to the CPU. Yes, they share that bus, but it has so much bandwidth that it doesn't really matter. On Intel-systems and on Macs, you have RAM using that same bus as well, and that eats ALOT of bandwidth. Not so in the A64/Opteron.

Second, replace northbridge with Power4 and you get except on the G5 side, the dedicated channel is running the full speed of the FSB and is running seamlessly within the system, unlike the Athlon 64's dedicated channel which runs at a slower speed and adds another pipeline stage to the overall proccessing of data.

What on Earth are you talking about? Replacing the northbridge with a Power4-CPU? Are you talking about buses on SMP-machines? On Opteron-machines, each CPU has a dedicated 1GHz HyperTransport-bus (unlike on Xeons', where they share one bus) and each CPU has a dedicated mem-channel as well (and each CPU can access other CPU's RAM as well, so that means that memory-bandwidth goes up as you increase the number of CPU's. 2x Opterons have 2x mem-bandwidth when compared to 1x Opterons. 4x Opterons have 4x the mem-bandwidth and so forth).

AS to your talks about A64's "dedicated channels running slower".... I really don't know what you are saying in there. Mem-channel on the G5 doesn't run at CPU-speed either, so I don't see your point. And their FSB's are slower than the CPU, so again, I don't see your point.

 

[nexusfx]

Oye.......

Please stop commenting, it makes you look silly and what you are saying is nonsense.

Proove me wrong, look throught he technical documentation for Power4/G5 send me a link that "prooves" me wrong, then I will willing acknowledge I didn't have all the info I needed. Don't just tell me to stopposting because of what you understand to be correct.

The internal make-up of an A64 is this:

[core 1]---[SRQ]---[Crossbar]---[Dual Channel Memory Controller]
[core 2]-----+ . . . . . +---[Hypertransport]

The core is what does all the processing, the pipeline length is not affected by having the northbridge on-die at all, that's just stupid. What incorporating the northbridge does allow for is extremely low latency memory access, and the Hypertransport link to the chipset will have much lower contention because it doesn't carry memory access traffic except for chipset generated memory traffic (which is the same case on all chipsets, so a moot point).

It doesn't work the same on the G5, the Power4 System controller makes for NO CONTENTION from data going to and from the CPU, EVERYTHING has a dedicated throughput. K I'm cutting this short, I'm sure you'll havbe something to say when I come back online.

 

[javiercr]

blue ray

why do you want blue ray now? even dual layer blank media is hard to find (and expensive)! i wouldn't know where to buy a blank blue ray disc.

 

[aethier]

i just hope that IBM isn't sending all the fast cpus to microsoft for the release of xbox 360. it does have two or three 3.5ghz g5's. and to ibm, i am sure the xbox will bring them much more cpu money then apple selling them in the g5 powermac, and imac


aethier

 

[Hattig]

Proove me wrong, look throught he technical documentation for Power4/G5 send me a link that "prooves" me wrong, then I will willing acknowledge I didn't have all the info I needed. Don't just tell me to stopposting because of what you understand to be correct.

It doesn't work the same on the G5, the Power4 System controller makes for NO CONTENTION from data going to and from the CPU, EVERYTHING has a dedicated throughput. K I'm cutting this short, I'm sure you'll havbe something to say when I come back online.

Let's do this for a 2.5GHz G5.

[G5]---5GB/s each way FSB---[Northbridge]---[Dual Channel PC3200]

Northbridge also connects to 2GB/s AGP8x (with PCIe that would be 4GB/s each way), and a 3.2GB/s HyperTransport bus IIRC.

The FSB also has to transfer cache coherency data in a dual-processor system. This isn't minor at all.

As you can see the diagram is actually VERY SIMILAR to the Athlon 64, except that it is split into two chips instead of one. Of course, the A64 has an internal 'FSB' per core that runs at core speed and is 64-bits wide. On a 2.4GHz A64 that is 20GB/s. More than the G5.

I still don't get what this "power 4 system controller" is, the G5 connects to a standard IBM northbridge designed for the G5, IBM PPC910 or something, I forget the name. It is functionally the same as your typical Intel northbridge.

 

[vouder17]

Can ANYONE tell me what time normally the updates are?

Cause I think it won't happen today anymore?

At 7:00PM in Netherlands...It is the same as 9:00AM in Cupertino.

 

[AidenShaw]

you've made so many mistakes, one doesn't know where to start

It doesn't work the same on the G5, the Power4 System controller makes for NO CONTENTION from data going to and from the CPU, EVERYTHING has a dedicated throughput.

For example, there's no POWER4 system controller in a G5, the POWER4 chip is a big, hot, fast dual-core IBM processor. The G5 uses a PPC970, which is a cost-reduced lightweight single core chip based on the POWER4 - but with lots of changes.

The G5 uses the U3 or U3H Northbridge, with a 16-bit link to the I/O devices.

There are lots of places for contention - FSB <-> RAM, to AGP, down the narrow HT bus to the PCI tunnels.

http://developer.apple.com/document...rMacG5/2Architecture/chapter_3_section_2.html\

______________________________________

But I agree with the earlier posts - you've made so many mistakes in your arguments that it's apparent that you don't understand much of what you're reading. You don't even get the words right (using G5 and POWER4 interchangeably, for example).

When we try to correct you, you don't listen. It's not worth the effort to do a complete point-by-point rebuttal.

 

[Hattig]

i just hope that IBM isn't sending all the fast cpus to microsoft for the release of xbox 360. it does have two or three 3.5ghz g5's. and to ibm, i am sure the xbox will bring them much more cpu money then apple selling them in the g5 powermac, and imac

Well, the XBox360 will be using a different processor to the G5, as previously discussed. You can also be certain that the margin on each processor will be a lot less. Microsoft might be able to take a loss on the consoles in the first year or two, but with a launch price of $399, and a $100 loss that would be $499 to spread between the processor, graphics system, I/O system, controllers, case, marketing, design, motherboard, DVD drive and so on. I certainly don't think that IBM is going to get more than $100 a processor, whereas they can probably get $200 for even a low end G5, and $500 for a high end one.

20m * $20 profit = $400m
2m * $200 av. profit = $400m

Probably worth the same over a period of a few years, possibly the G5 would be more lucrative even, depending on the per-processor profit, and IBM's own G5 systems, etc.

(I actually don't think that IBM will get more than $50 a processor for XBox360 in the long term)

 

[nexusfx]

And again

For example, there's no POWER4 system controller in a G5, the POWER4 chip is a big, hot, fast dual-core IBM processor. The G5 uses a PPC970, which is a cost-reduced lightweight single core chip based on the POWER4 - but with lots of changes.

The G5 uses the U3 or U3H Northbridge, with a 16-bit link to the I/O devices.

There are lots of places for contention - FSB <-> RAM, to AGP, down the narrow HT bus to the PCI tunnels.

http://developer.apple.com/document...rMacG5/2Architecture/chapter_3_section_2.html
______________________________________

But I agree with the earlier posts - you've made so many mistakes in your arguments that it's apparent that you don't understand much of what you're reading. You don't even get the words right (using G5 and POWER4 interchangeably, for example).

When we try to correct you, you don't listen. It's not worth the effort to do a complete point-by-point rebuttal.

That's the point, you're "trying" to correct me but you have yet to do so, I'm listening "very carefully" like I said earlier if you can proove me wrong I have no problem with that, however you haven't, though your actually bringing up a good point and actually included a diagram from the developers pages, which I do understand.

http://www.apple.com/powermac/architecture.html read point 4 then dowload the tech pdf's from http://www.apple.com/g5processor/ if I'm wrong I'm wrong, if I'm misunderstanding, I'm misunderstanding. Just explain it to me if I'm wrong, I'm not trying to be a pain.

 

[quta]

this board is filled with crap

There is so much crap here, I can't believe it. So many of you are debating using technobabble you barely understand, about machines you'll probably never actually use to their full potential. What really matters, I think, is how well you can actually use your machine to do what you want, effectively. For some people, that means games; so go with a machine that supports lots of games, and has a good graphics card (i.e., get a console and stop worrying!) For others, it means the ability to get work done efficiently and effectively without spending a lot of time dealing with problems. Even with a slower overall computational speed or memory throughput, an OS X machine is a much more productive environment for me than a Windows or even Linux box (I do scientific software development, technical writing, web stuff, digital photography, home movies, interface with my iPod, manage finances, etc., all on the same machine). I could probably compile my code a little quicker on a fast Opteron, and run it a little faster too, but I can assure you I spend less time getting it written in the first place, and my time is more valuable than the computer's.

 

[nexusfx]

Good Call.

There is so much crap here, I can't believe it. So many of you are debating using technobabble you barely understand, about machines you'll probably never actually use to their full potential. What really matters, I think, is how well you can actually use your machine to do what you want, effectively. For some people, that means games; so go with a machine that supports lots of games, and has a good graphics card (i.e., get a console and stop worrying!) For others, it means the ability to get work done efficiently and effectively without spending a lot of time dealing with problems. Even with a slower overall computational speed or memory throughput, an OS X machine is a much more productive environment for me than a Windows or even Linux box (I do scientific software development, technical writing, web stuff, digital photography, home movies, interface with my iPod, manage finances, etc., all on the same machine). I could probably compile my code a little quicker on a fast Opteron, and run it a little faster too, but I can assure you I spend less time getting it written in the first place, and my time is more valuable than the computer's.

AMEN! I'm tired of arguing this point. So to hell with it all, I'm gonna play my Xbox, then watch some Viva La BAM, render some 3d on my G5, and then continue to be lazy the rest of the day cause I'm on friggin paid vacation woot. 🙂

 

[dicklacara]

Microsoft might be able to take a loss on the consoles in the first year or two, but with a launch price of $399, and a $100 loss that would be $499 to spread between the processor, graphics system, I/O system, controllers, case, marketing, design, motherboard, DVD drive and so on.

Mmmm.... the Justice Department would certainly be interested in that!

 

[Hattig]

That's the point, you're "trying" to correct me but you have yet to do so, I'm listening "very carefully" like I said earlier if you can proove me wrong I have no problem with that, however you haven't, though your actually bringing up a good point and actually included a diagram from the developers pages, which I do understand.

http://www.apple.com/powermac/architecture.html read point 4 then dowload the tech pdf's from http://www.apple.com/g5processor/ if I'm wrong I'm wrong, if I'm misunderstanding, I'm misunderstanding. Just explain it to me if I'm wrong, I'm not trying to be a pain.

Going back to your original discussion that involved x86 and pipelines in some wierd way ...

Basically the chipset does not affect the core in any way. The core has its own execution pipeline, the chipset doesn't do any "x86 execution" on an x86 system. The more pipeline stages, the faster you can run a processor in general, with the downside being that branches are heavily affected in longer pipelines if they are guessed incorrectly, hence better and better branch prediction units. Memory latency delays a processor too, hence larger caches, and moving the memory controller closer to the processor.

The G5's system layout is very similar to an Intel x86 system, except that instead of a shared FSB, each processor has its own FSB. In that respect, it is similar to the old AMD K7 x86 760MPX platform of 3 or 4 years ago. In fact, I'd say it was almost pretty much exactly the same as this, except with more up to date AGP and memory bus options as you'd expect.

Your information is coming from the marketing website of Apple. It isn't very revolutionary, it isn't new. It's slightly more advanced than Intel's current solution, the same as AMD's old solution, and behind AMD's current solution.

In fact, the dual-core Freescale G4 chip will be close to AMD's current solution, as that has an entire northbridge on chip, even including PCIe!