Where's a good place to learn more about NUMA and how it might migrate down to the desktop?
I don't know. I expect it'll look somewhat different than NUMA in high end machines/systems. As you probably know, it means "Non-Uniform Memory Access" and all that means is that different pieces of memory are accessed in a different ways by the processor's hardware, although generally it looks uniform to the application -- i.e. the hardware or the OS. In a consumer system this will probably be done using some kind of a packet based communications bus, like RapidIO. The RapidIO.org website has a couple of interesting papers which talk about how RIO might be used to implement a multi-processor system with seperate memory pools. RIO operates at a fairly small burst level -- ~32 byte cacheline transactions are probably going to be the most common. Another alternative would be to have a non-data addressed bus (like Ethernet) and use the hardware's DMA engine and the OS' VMM system to page memory across it in medium-sized chunks (i.e. pages of 4K+).
I understand the feeling. However, competition is good. Multiple suppliers is good. Keeping options open is good. Keep 'em coming, Motorola. Let's see what you've got.
BTW, three cheers to Outsider for a truly superb thread. I also would like to see more links to some of this info.
Sorry if I led people to believe that the above roadmap is written in stone somewhere at IBM but it's mostly info from the IBM roadmap, some guess work and yes, some wishful thinking. The 900 series part should be pretty accurate up to the 990, which we don't know much about. Only that it is most likely based on the POWER6. The 850 series is based off the IBM roadmaps for the G3 (750 series) with VMX added in.
Programmer brings up some good point on embedded RAM. Eventually the difference between processor core and memory will become so blurry that we will have to rethink our attitudes on expansion. An idea is to intricately link each core with a sizable memory chunk, like 256MB. It would basically be its L1 or L2 cache. Each core would have a wide bus to a central hub (or hubs) on die for sharing memory, or cache snooping (the difference will blur). So for example a 4x4 matrix of processor cores would have 4GB of total memory plus perhaps a pool of shared memory that may be slower but can be more densely packed in the range of 128GB or so. How do we get so much on a processor core? Well CMOS process size is going down to accound for some of the density, but also many advances are being made in 3D memory architecture, basically wafers of memory cells being stacked upon each other. Instead of disk drives we would use solid state MRAM (Magnetic RAM; holds its bits even after the power has been turned off) drives that IBM has been dumping a lot of research into. The interface would not be a puny ATA or S-ATA port or even something as mundane as a Super Duper PCI. At this point it would make more sence to have these 3D MRAM communicate with the processor core via a direct memory bus to the processor. On separate sockets, this chip(s) would be able to hold many GBs of permanent storage, around 500-1000GB, maybe more!
Yes... I had seen the link posted on Ars but hadn't followed it through. It is exactly the kind of thing I'm talking about, and it is probably quite similar to the "Cell" project (if it is in fact different than the Cell project... it might be the same thing). Notice that they are using enhanced 440 cores w/ FPU & SIMD functionality added. No mention of VMX/AltiVec though, which would be a shame -- hopefully they apply VMX to such a setup eventually.
it would be nice if they made everything modular so that you could add various processing units as needed, which would definitely be (and needs to be) the way to go.
Of course we are. All we have now is a G4 that is in dire need of an update. And Apple is really starting to lag in this Ghz gap, so we are all pretty desperate and want something to brag about. I sure know that I want a performance hit, but I don't know if you do.
You'll notice that there is little discussion of timeframe -- these massively parallel machines with large amounts of embedded RAM are a few years off. 970 this year, 970+ next year, 980 (i.e. POWER5 based) in 2005, 990 in 2006 at the earliest. Cell chips might start to show up in 2005, after all that's when Sony is targeting.
Maybe just a little, but here is FACT, the 970 HAS been announced and is in process of production. It has Altivec? when IBM does not require it. Optimism?, no just simple math.
I do believe your relevance of optimism is more to the point of a "roadmap", that Apple and its customers can hang their hat on. Fact, IBM have discussed the 980 and 990. Proving, that the 970 is not the only Intel competative chip we will see from IBM.
Further, the problem with the G4 and its lack of incremental clock rates, restricts Apple's ability to announce faster iMacs, eMacs, PowerBooks. So right now the PowerMac sales have been seriously affected by the performance gap (and Apple's poor financial results were attributed to this product - Fred Anderson's words not mine), imagine what would happen to Apple financially if this spread to the the rest of the product line.
The 970 @ 1.2-2.5+ will be fine for about 18-24mths, before the 980 is released.
Now Amorph, some of use here are geeks and some of use are not. I would consider a person who uses a Mac not a geek because Macs are for the normal person. Now look at the PC users, I just have to go into the Computer Science faculty at my University and I can spot a PC user to a Mac user in a flash. Go and yell that the Macintosh sucks and you will have at least 5 guys coming up to you right away to debate it. For a PC user in that faculty, you just have to see if they are wearing huge glasses and wearing pants that come up to their chest. Kinda like Steven Q Erkel.
Comments
Originally posted by OverToasty
Where's a good place to learn more about NUMA and how it might migrate down to the desktop?
I don't know.
Originally posted by Mac OS X Addict
Just one question. I've heard of the 980 being based on the Power 5, but where did you get all the other info from??
Rumour, supposition, and clever guesswork.
Originally posted by Derrick 61
Moto who??
I understand the feeling. However, competition is good. Multiple suppliers is good. Keeping options open is good. Keep 'em coming, Motorola. Let's see what you've got.
BTW, three cheers to Outsider for a truly superb thread. I also would like to see more links to some of this info.
Originally Posted By Programmer:
Rumour, supposition, and clever guesswork.
Lets hope that at least some of this comes true.
Programmer brings up some good point on embedded RAM. Eventually the difference between processor core and memory will become so blurry that we will have to rethink our attitudes on expansion. An idea is to intricately link each core with a sizable memory chunk, like 256MB. It would basically be its L1 or L2 cache. Each core would have a wide bus to a central hub (or hubs) on die for sharing memory, or cache snooping (the difference will blur). So for example a 4x4 matrix of processor cores would have 4GB of total memory plus perhaps a pool of shared memory that may be slower but can be more densely packed in the range of 128GB or so. How do we get so much on a processor core? Well CMOS process size is going down to accound for some of the density, but also many advances are being made in 3D memory architecture, basically wafers of memory cells being stacked upon each other. Instead of disk drives we would use solid state MRAM (Magnetic RAM; holds its bits even after the power has been turned off) drives that IBM has been dumping a lot of research into. The interface would not be a puny ATA or S-ATA port or even something as mundane as a Super Duper PCI. At this point it would make more sence to have these 3D MRAM communicate with the processor core via a direct memory bus to the processor. On separate sockets, this chip(s) would be able to hold many GBs of permanent storage, around 500-1000GB, maybe more!
From MacArs.
Originally posted by wmf
Wow, y'all sure are optimistic.
Of course we are. All we have now is a G4 that is in dire need of an update. And Apple is really starting to lag in this Ghz gap, so we are all pretty desperate and want something to brag about. I sure know that I want a performance hit, but I don't know if you do.
Originally posted by Mac OS X Addict
I sure know that I want a performance hit, but I don't know if you do.
Maybe you meant a performance bump? We've taken enough performance hits already...
Originally posted by wmf
Wow, y'all sure are optimistic.
You'll notice that there is little discussion of timeframe -- these massively parallel machines with large amounts of embedded RAM are a few years off. 970 this year, 970+ next year, 980 (i.e. POWER5 based) in 2005, 990 in 2006 at the earliest. Cell chips might start to show up in 2005, after all that's when Sony is targeting.
Originally posted by CharlesS
Maybe you meant a performance bump? We've taken enough performance hits already...
That's what I meant, a performance bump. Sorry, doing too much work and not thinking when I am typing.
Originally posted by wmf
Wow, y'all sure are optimistic.
Maybe just a little, but here is FACT, the 970 HAS been announced and is in process of production. It has Altivec? when IBM does not require it. Optimism?, no just simple math.
I do believe your relevance of optimism is more to the point of a "roadmap", that Apple and its customers can hang their hat on. Fact, IBM have discussed the 980 and 990. Proving, that the 970 is not the only Intel competative chip we will see from IBM.
Further, the problem with the G4 and its lack of incremental clock rates, restricts Apple's ability to announce faster iMacs, eMacs, PowerBooks. So right now the PowerMac sales have been seriously affected by the performance gap (and Apple's poor financial results were attributed to this product - Fred Anderson's words not mine), imagine what would happen to Apple financially if this spread to the the rest of the product line.
The 970 @ 1.2-2.5+ will be fine for about 18-24mths, before the 980 is released.
Just my 2 cents....
Originally posted by Mac OS X Addict
I sure know that I want a performance hit, but I don't know if you do.
Performance: Heroin for Geeks.
Originally posted by Amorph
Performance: Heroin for Geeks.
Now Amorph, some of use here are geeks and some of use are not. I would consider a person who uses a Mac not a geek because Macs are for the normal person. Now look at the PC users, I just have to go into the Computer Science faculty at my University and I can spot a PC user to a Mac user in a flash. Go and yell that the Macintosh sucks and you will have at least 5 guys coming up to you right away to debate it. For a PC user in that faculty, you just have to see if they are wearing huge glasses and wearing pants that come up to their chest. Kinda like Steven Q Erkel.