Slashdot Mirror
Nvidia Mulls Cheap, Integrated x86 Chip
CWmike writes "Nvidia is considering developing an integrated chip based on the x86 architecture for use in devices such as netbooks and mobile Internet devices, said Michael Hara, vice president of investor relations at Nvidia during a speech that was webcast from the Morgan Stanley Technology Conference this week. Nvidia has already developed an integrated chip called Tegra, which combines an Arm processor, a GeForce graphics core and other components on a single chip. The chips are aimed at small devices such as smartphones and MIDs, and will start shipping in the second half of this year. 'Tegra, by any definition, is a complete computer-on-chip, and the requirements of that market are such that you have to be very low power and very small but highly efficient,' Hara said. 'Someday, it's going to make sense to take the same approach in the x86 market as well.'"
For those of us who dealt with intel's "integrated" graphic cards on laptops for the past several years now... on their behalf I just want to say PLEASE FOR THE LOVE OF ALL THINGS SHINY AND SILICON, DON'T DO IT! Anything with the word "integrated" near it makes me want to cringe... it's a post traumatic stress response caused by watching a myriad of good video games shutter, blink, crash, and burn right in front of me. It's a black day indeed when Warcraft 3 can't run at full resolution on a laptop produced only a year ago.
#fuckbeta #iamslashdot #dicemustdie
Nvidia develops a very basic x86 CPU thats tightly coupled to one of their embedded GPUs that doesn't implement any x86 technology that's still currently patent-protected. The basic x86 CPU acts as a shim for software that expects to talk to an x86 CPU and offloads as much as possible to the significantly more advanced GPU running the bulk of the load. The end result? An x86-compatible embedded system that vastly outperforms anything currently on the market that doesn't violate anyone's active x86 patents.
read:
"Nvidia NULLS Cheap, Integrated x86 Chip "
Previously: "Linux... Toward the Sunrise..." Now: "Linux... Toward the-- No, now, part of Every Sunrise"
1. Microbenchmarking becomes too difficult when two or more cores of different types are used
2. Inter-core communication takes a hit because you'll end up designing new flit routers (or even newer protocols) that efficiently route packets within cores that have different communication topology
3. Failure of one core can render the chip useless, whereas in the case symmetric multicore design, failure of one core means other cores are still functional and the company can market them separately.
4. Production issues involving 1,2, and 3 above.
Eventually we'll end up with a dynamic multicore design which seems more promising than asymmetric designs. Some research has been done in this area (symmetric vs. asymmetric vs. dynamic via threading): http://www.cs.wisc.edu/multifacet/papers/tr1593_amdahl_multicore.pdf
FreeBSD bounties
not only that but what exactly do you think your handheld/phone is going to do with more than 4 gig of ram per process?
If you mod me down, I will become more powerful than you can imagine....
Why botther at all? Better go straight to x64, I mean, even the lowliest of nvidia GPUs is already 64 bits, why bother with 32 bits technology?
They day an embedded system's CPU needs to address more than 4 gigs of memory (which is essentially why you would shift from a 32-bit to 64-bit CPU) is the day my shit turns purple and smells like rainbow sherbet.
You aren't allowed to call them netbooks, didn't you get the subpoena?
Confucius say, "Find worm in apple - bad. Find half a worm - worse."
x64 is a Microsoft marketing term. Please stop using it. The architecture is x86-64.
Since when have you ever seen a phone that even has 1 gig of memory in it let alone over 4?
My Intel 855GM handles xterms very well, recently they have become very wobbly slimey when I drag them around in Gnome, other than that everything is fine with my integrated chip.
Daily?
http://na.blackberry.com/eng/devices/blackberrystorm/storm_specifications.jsp
Okay, but the point still stands that 64bit address space isn't needed on a phone at any point in the future. That is probably one of only a handful that even have 1 gig.
***WOOOOOOSHHHH***
Wow, did Chuck Norris just go by, or did you miss a joke?
not only that but what exactly do you think your handheld/phone is going to do with more than 4 gig of ram per process?
Nothing as 640k is enough for any phone.
From what I can find, that device has 196 MB RAM and 1 GB of onboard storage (this is in addition to the 128 MB of flash for the OS and the microSD slot).
This is different than having 1 GB of byte-addressable DRAM.
We're already a quarter of the way there. It won't be long now.
What was that quote about 640K? ;)
Surely a better design is to produce a series of very small, highly specialized, very fast cores on a single piece of silicon, and then have a layer on top of that which makes it appear to be an x86, ARM or whatever.
One reason for having a bunch of specialist cores is that you don't have one core per task (GPU, CPU or whatever), but rather one core per operation type (which means you can eliminate redundancy).
Another reason is that having a bunch of mini cores should make the hardware per mini core much simpler, which should improve reliability and speed.
Finally, such an approach means that the base layers can be the same whether the top layer is x86, ARM, PPC, Sparc or a walrus. NVidia could be free to innovate the stuff that matters, without having to care what architecture was fashionable that week for the market NVidia happens to care about.
This is not their approach, from everything I'm seeing. They seem to be wanting to build tightly integrated system-on-a-chip cores, rather than having a generic SoaC and an emulation layer. I would have thought this harder to architect, slower to develop and more costly to verify, but NVidia aren't idiots. They'll have looked at the options and chosen the one they're following for business and/or technical reasons they have carefully studied.
If I was as bright as them, why is it that they have the big cash and I only get the 4 digit UID? Ergo, their reasoning is probably very sound and very rational, and if presented with my thoughts could very likely produce an excellent counter-argument to show why their option is logically superior and will produce better returns on their investments.
The question then changes as follows: What reasoning could they have come up with to design a SoaC unit the way they are? If it's the "best" option, although demonstrably not the only option, then what makes it the best, and what is it the best at?
It's a small world and it smells funny; I'd buy another if it wasn't for the money; Take back what I paid (SoM)
Why botther at all? Better go straight to x64, I mean, even the lowliest of nvidia GPUs is already 64 bits, why bother with 32 bits technology?
They day an embedded system's CPU needs to address more than 4 gigs of memory (which is essentially why you would shift from a 32-bit to 64-bit CPU) is the day my shit turns purple and smells like rainbow sherbet.
Is 640K enough for you still?
The iPhone packs 128 meg, as does the BlackBerry Bold. A modern smartphone packs as much ram as the average desktop did a little under a decade ago. SODIMMs (laptop ram) look to be just over $10 per gig. But hey, 640k is all anyone really needs, right?
You may notice some changes in the appearance and fragrance of your excrement in less than a decade.
"640K should be enough for anybody", updated for 2009.
Plus, TFS specifically mentions netbooks as part of where the chips would be used, which are more likely to make that particular feature likely than handhelds/phones, as the use cases for those are pretty much those for a general purpose laptop -- but with more focus on battery life (hence, power consumption), weight, and cost. I suspect the same issues that Nvidia sees that would make it make sense to "at some point" go from ARM to x86 for integrated chips would also at some point make it make sense to go to x86-64 as well, though it might be at a later point than going to x86.
the day my shit turns purple and smells like rainbow sherbet
Anime styled shit sure sounds like a good way to earn a lot of money in Japan, by combining two of its most prominent fetishes into a product, that can be easily packaged and sold in vending machines.
A Beowu.... aww fuck it.
I will not be pushed, filed, stamped, indexed, briefed, debriefed or numbered. My life is my own.
They should just push ARM heavily. ARM is doing great right now. Companies like Texas Instruments are pushing the architecture heavily, and there's high demand.
Linux ARM support is blasting ahead, thanks to projects like the Beagleboard.
On top of that, a while ago Microsoft said they were developing an ARM version of Windows. Although we won't see it right away, in a couple years that'll open up even more options.
If they push ARM hardware heavily enough, software will follow. Heck, the software is already coming along, so they just have to market the hardware properly.
Most people won't know the difference between a linux MID and a windows MID. Both have "Email", "Instant Messenger", "Calendar", "Web Browser", etc., and if you need a new program you just download it... Nobody would even think of installing software off a CD, so most "Why won't this work?" scenarios won't even come up. It'll just look slightly different.
And once a couple game devs follow - or heck, a program like Google Earth - it won't be long before oodles of software is being ported, and the ARM-x86 barrier breaks down.
Since next year (or the year after or whenever) when 256 GB of DDR7 RAM costs $1.95. Take the long view, dude.
"Believe me!" -- Donald Trump
The demarcation of storage and RAM is a legacy constraint forced by hardware limitations. Ubiquitous 64-bit and SSD will blur and eventually totally eliminate this separation.
POKE 36879,8
x64 is a Microsoft marketing term. Please stop using it. The architecture is x86-64.
Pray tell, why not amd64 then - the way it was originally marketed by the inventor when it was released?
Last I checked, there's no definite established term for this, anyway, and x64 is the shortest while still being vendor-neutral. Even if Microsoft came up with it first (and are you sure they did, really?), so what? I don't understand how it is a "marketing term" for them, as they don't market it.
So, instead of forging ahead with novel parallel processor technology, Nvidia thinks that the way to go is to copy last century's dinosaur CPU? It's enough to make a grown man cry. Whoever is in charge of research at Nvidia should be given the boot. What a waste of talented engineers! But it's not too late, Nvidia. Click on the link below and do the right thing. Otherwise, Otellini will tear you a new one and you know it.
How to Solve the Parallel Programming Crisis
On a different note, did not Nvidia recently say that the world is moving away form the CPU? I am beginning to think that Nvidia is either scared or bluffing. Otellini made a comment last week to the effect that Nvidia needs a CPU in order to build a GPGPU heterogeneous multicore processor and now there's all this talk about an Nvidia x86 coming out in a couple of years. Does Otellini call the shots at Nvidia? I am not so sure.
Nvidia is right that the days of the CPU are numbered but so are the days of the GPU. The reason is simple. Neither CPU nor GPU provides a universal or homogeneous solution to the parallel programming crisis. The heterogeneous route is pure folly too, if only because it does nothing to solve the crisis. In fact, it makes it worse because it combines two incompatible parallel models on the same dye. A match made in hell.
There is way to solve the crisis but it involves neither CPU nor GPU. Think pure MIMD vector processing. That's where Nvidia should invest its processor R&D resources, all of it. That is, if it wants to dominate the parallel computing industry for the next several decades. Intel would not know what hit it until it's too late. Big money is at stake. BIG.
Because flash based SSDs will be as fast as DRAM? In reading, writing and latency?
Or, while everyone is nitpicking...
Why not refer to it by "AMD's original designation for this processor architecture, 'x86-64'"?!
Tell me, if they announced an intention to do a SPARC core, would you assume they meant a 32-bit version? How about POWER?
x86 is just as 64-bit as they are.
Who said anything about flash-based SSD? I said SSD as in SOLID-STATE NON-VOLATILE STORAGE. Regardless, flash now is faster than CPU registers were not so long ago.
POKE 36879,8
x64 is a short, sweet, unambiguous term that is much less unwieldy than x86-64. Even if Microsoft was the first to use it (and AFAIK they certainly don't market it, I've only ever seen it in the filenames of MSDN ISOs) what about that makes it verboten for the rest of us?
I ran into a limit with 32 bits more than 10 years ago.
My Fortran compiler was Cray derived (Lahey, iirc), and I had dynamically allocated a huge array. They were in some way bit-addressed, leading to a crash.
I turns out that my adviser's machine had more memory (512Mb) than any of their own test machines.
The workaround at the time was static allocation, which made the code faster, anyway.
hawk
Bullshit. The difference between flash storage and RAM is hardly a legacy constraint; the hardware limitations are very significant.
Flash devices are many orders of magniture slower than DRAM. Flash devices have very limited write cycles, whereas DRAM does not. Flash devices operate on large (erase) block sizes - starting at 64KB. This is many orders of magnitude higher than RAM, which is addressed by the byte. The erase block size is also much larger than the 512 byte sectors used in hard drives. Note that most consumer flash-based devices present themselves as block devices with 512 byte sectors.
On flash, data can be initially written in chunks much smaller than the erase block size, but once written, an entire erase block must be erased at the same time.
It can be useful for data on secondary storage to be accessible as if it were in memory. Modern operating systems already provide this functionality - see mmap(2). However, the usefulness of exposing a large slow block device's entire address space is limited unless a high-performance application actually randomly uses all that data - and if that's the case the frequent page faults caused by the lack of DRAM for caching would result in lower performance than just using standard read/write operations.
Sorry, my bad. I read SSD and immediately thought about flash. While flash is faster than CPU register were some time ago, it is slower than DRAM (even SDR) is now. In the future, flash may become faster than PC66 RAM, but it will probably never become as fast as the DRAM technology of the day.
Also, if you addressed all storage using the same addresses as RAM, how would you handle removable media (especially the one that can have different sizes)?.
I never said that SSD would eliminate RAM either, which in turn will not eliminate on-die cache or registers in the forseeable future. When you've got a practically infinite address space you may as well use it.
POKE 36879,8
Sure is lack of imagination in here...
POKE 36879,8
x64 probably started as a verbal colloquialism. x64 rolls off the tongue much easier than x86-64.
-- If you try to fail and succeed, which have you done? - Uli's moose
I would be brave enough to say that in the not-too-distant future, 1GB storage will be about as common as 1KB RAM chips are now.
POKE 36879,8
True, but Level 1, 2, and 3 CPU caches are all much faster than DRAM too, and yet they share the same address space. Theoretically, you could just make the SSD byte addressable and have your "main memory" DRAM act as a Level N+1 cache for the SSD. If you want a file system you run a RAM disk :-) That would make systems come up much faster from a power-off mode, but system resets would be more challenging to pull off.
Alternatively, you could also think of it as the SSD being a big persistent swap partition for your main DRAM memory. That second approach might be easier to implement at first, since the first approach is more of a paradigm shift in how you organize your O/S. It might also make it easier to deal with flash mem's limits on repeated writes to the same location.
Laissez lire, et laissez danser; ces deux amusements ne feront jamais de mal au monde. - Voltaire
x86 is more than x64, so it's better right?
POKE 36879,8
This just in, nVidia announces world's first netbook to require not one but two separate AC adapters at all times. Other features include built-in vacuum cleaner noise generator, and thermal pubic hair remover.
I read the script, and I think it would help my character's motivation if he was on fire. -Bender
Or external PCIe. I've been waiting for that. The PCIE standard has it specified, just nobody wants to make stuff for it. Think of it this way, you come home, you plug in a box (with its own PSU) into your laptop, and you can now game on your laptop with whatever cards you had put in that box. When you're done, unplug everything, switch your resolution/drivers if necessary, and go.
While I don't have much interest in making my PCs boot faster (I reboot them when I want to change some non-hot-plug device or something goes bad, in both cases the whole process takes way longer than it takes for my PC to boot afterward), I don't see how it would make systems boot faster than they do now.
If you mean that on power-on the data is copied from the ssd/hdd to the RAM then we already have that function - it's called "hibernate". If the PC tried to use the ssd/hdd without copying the data first, it would be very slow until the data is cached to the RAM.
Alternatively, you could also think of it as the SSD being a big persistent swap partition for your main DRAM memory.
How does that work? Don't we create swap partitions on slower but bigger devices so that the PC can write data when it runs out of RAM?
What would be the difference from what we have now? Now the OS handles swapping to disk, then some chip on the motherboard would handle it? Except that the chip would have to be like the OS and know which processes to swap out. How it would even know that the system is running out of RAM? AFAIK RAM does not have a FAT or some other file system so as to be able to tell how much free space is there.
You can by SSD now which consists purely of the same RAM sticks as on your motherboard. This is the idea that I'm getting at. So how do we handle it? Seems a bit wasteful to limit that massive bandwidth through legacy storage interfaces, so put it on the RAM bus. Also a bit redundant to double-handle it via antiquated filesystem IO, so map it directly.
See where it's going...
POKE 36879,8
SSD, that consists of RAM sticks, that is put on a RAM bus, that is accessed by the same address and has no file system. Wouldn't that just be a RAMdisk?
And if you want it not to lose data while the PC is turned off - do not turn off the PC, just leave it on stand-by (with an internal battery).
*sigh*
You know, I get the same, or similar, response every time I talk to a proponent of any type of bullshit.
"You don't see how Elvis could still alive? Gee, you sure are closed minded ..."
Accusing people of being closed-minded or unimaginative seems to be the equivalent of saying "I have no evidence and no clue what I'm talking about, so I'm going to attack you just so you'll shut up".
What about a NAS box? I'm considering building one using commodity PC parts, and when you can have a 4GB disk cache for $40, it's tempting. Though on the other hand, my desktops are kind of dated, and having a NAS box with twice or more ram of any computer it would be serving files to would be completely redicilious.
Back it up to floppy.
POKE 36879,8
Look man, extrapolating SSD into an argument about flash vs SDRAM just shows a lack of imagination. I'll see you in ten years.
POKE 36879,8
It seems that everything is moving in the direction of operational efficiency. More instructions per cycle, less power draw, faster and more efficient buses among processor, memory and peripheral devices are among important issues being focused on.
But what ever happened to Moore's law? Are we already outside of its prediction? Has the chain been broken? I thought we would all have 5GHz machines running ice-cold by now but some of the latest and greatest stuff is a mere 1.6GHz atom processor based sub-notebook. Is power and speed out of vogue or is it just not as possible as it once was and so people are drawing attention in other directions hoping people have forgotten about Moore?
Well if you're talking about SPARC it would be the x64%4 architecture.
Easy - you already have a separation between volatile and non-volatile storage in all operating system kernels. It doesn't matter whether the underlying technology is the same - in fact you can deliberately blur the line by using part of your RAM as a virtual disk (tmpfs + file + losetup in Linux, memory-backed vnode in FreeBSD).
Sam ty sig.
That's not quite true - the CPU's MMU handles some housekeeping on used RAM. However that's none of the motherboard's business, especially with memory controllers being integrated into the CPU these days. And even that information is far too limited to implement page replacement, which is still the responsibility of a sophisticated kernel.
Sam ty sig.
So you say that the OS would still have to do swapping whether the hdd is mapped to RAM address or not. So what would be the point of mapping if the OS (and the CPU) still has to swap pages in and out of the hard disk?
No, the number represents how far it is from perfection. Perfection is when there is nothing left to take away. The ultimate CPU will be x0. I'm going to trademark it ahead of time.
Sam ty sig.
Dude, external PCIe is available in laptop for years, it is called ExpressCard. And suprise, it's even used for external graphics: http://www.tomshardware.com/reviews/vidock-expresscard-graphics,1933.html
:wq
These go to eleven.
"It's such a fine line between stupid and clever" -- David St. Hubbins, Spinal Tap
Problem is that more and more netbooks are sold with linux, and NVIDIA drivers integration in any distro is less than stellar. Contrast that with Intel hardware where everything is well supported by all vendors.
Unless they open their drivers, this platform will be Windows-only so even their lower-end models will be hampered by the Windows Tax.
That won't go very far.
No that was last time, when his shit turned its current color.
Get back to him in a couple of years for something really weird.
No, it shows that the poster that you are arguing with understands the different physical constraints between volatile and persistent storage. He also understands that we can make external storage look like RAM already, and the reason that we don't do so is that abstraction would kill performance for most applications that need to be aware of the distinction.
You have nothing in response other than "blah blah imagination I'll be right in ten years". Yeah, course you will sunshine. As we progress down the fab sizes I'm sure that the distinction that volatile memory only needs to store a charge for a tiny amount of time (true looking back through DRAM all the way to mercury delay lines) where-as persistent storage needs more power to flip a state in some sort of field, ie magnetic.
But of course, that's lack of imagination and physics will be drastically altered in the next ten years. I can't understand why nobody has seen you for the visionary that you are...
Slashdot: where don knuth is an idiot because he cant grasp the awesome power of php
Would NVIDIA be able to do a Transmeta? Seeing that their GPUs are supposed to be more and more CPU-like, maybe they could translate the whole x86 instruction set in software?
Intel recently announced it was making the Atom CPU core available for SoCs made in TSMC. NVIDIA has dealings with TSMC, as they only do design and outsource manufacturing. So theoretically NVIDIA could just use an Atom core as a base and slap a GPU on it, much like they did with Tegra with an ARM core.
From what I understand, most interest in external PCI Express has been related to server environments and virtualization. I worked on this product a couple of years ago but unfortunately it was canned. Not really something you can use with a laptop though anyway :-)
were you expecting to see a sig here? perhaps you'd rather see the inside of an ambulance!
Take a look at this: http://www.nvidia.com/page/quadroplex.html These things hold graphics cards externally, though they're only for desktops. There are models that hold as many as four external graphics cards. They're not cheap though.
The demarcation of storage and RAM is a legacy constraint forced by hardware limitations. Ubiquitous 64-bit and SSD will blur and eventually totally eliminate this separation.
SSD is still orders of magnitude slower than RAM. The only way the distinction is going to go away is if someone comes up with persistent storage that's as fast as RAM but as cheap (per gigabyte) as disk. There's no reason to think such storage exists even in principle, unless you're living in a future where everything is made for free by solar-powered nanorobotic universal assemblers or something (and therefore don't have to worry about price).
Fast, cheap, persistent: pick any two.
MediaWiki developer, Total War Center sysadmin
Yes, but at 1x PCIe 1.0, it's a mere shadow of what video demands today. The 1x lane on the ExpressCard slot (2Gbps after 8/10 encoding) was intended to be an inexpensive, nice upgrade over PCI (Cardbus), and nothing more. It makes me wish thay had planned-ahead, and pushed 2x into the slot for a little more cost.
It has one other serious problem: due to bandwidth limitatations, you can't route the images from the video card expansion back to the laptop LCD, so this is not a "portable" solution. Even though the expansion box is very small, you must connect it to an external display, and that is NOT small.
The good news: ExpressCard 2.0 will be released in 2010, and will feature USB 3.0 and PCIe 2.0, so the data rate will jump to 4.0 Gbps. That might actually be enough to make this concept work.
Man is the animal that laughs.
And occasionally whores for Karma.
It's a black day indeed when Warcraft 3 can't run at full resolution on a laptop produced only a year ago.
Yes it can! I have the screenshots to prove it!
(Had to turn the detail down slightly though...)
"Slashdot - News and Chat Sites Deviant". (Click "homepage" link above for details).
Yes, I want to have to plug an external PCIe device into my cell phone in order to see the caller ID number on the screen as it rings.
I wish that some company would combine ARM and x86 into one chip already, so that a complete migration to a more efficient line of processors becomes possible.
DISCLAIMER: I know nothing about the feasibility of this.
I am not devoid of humor.