Slashdot Mirror
Intel Next-Gen CPU Has Memory Controller and GPU
Many readers wrote in with news of Intel's revelations yesterday about its upcoming Penryn and Nehalem cores. Information has been trickling out about Penryn, but the big news concerns Nehalem — the "tock" to Penryn's "tick." Nehalem will be a scalable architecture with some products having on-board memory controller, "on-package" GPU, and up to 16 threads per chip. From Ars Technica's coverage: "...Intel's Pat Gelsinger also made a number of high-level disclosures about the successor to Penryn, the 45nm Nehalem core. Unlike Penryn, which is a shrink/derivative of Core 2 Duo (Merom), Nehalem is architected from the ground up for 45nm. This is a major new design, and Gelsinger revealed some truly tantalizing details about it. Nehalem has its roots in the four-issue Core 2 Duo architecture, but the direction that it will take Intel is apparent in Gelsinger's insistence that, 'we view Nehalem as the first true dynamically scalable microarchitecture.' What Gelsinger means by this is that Nehalem is not only designed to take Intel up to eight cores on a single die, but those cores are meant to be mixed and matched with varied amounts of cache and different features in order to produce processors that are tailored to specific market segments." More details, including Intel's slideware, appear at PC Perspectives and HotHardware.
It seems that AMD has lost, and I'm not trying to troll. It just seems that fortunes have truly reversed and that AMD is being beaten by 5 steps everywhere by AMD. Anybody have an opposing viewpoint? (Being an AMD fan, I am depressed.)
...they're taking AMD's on-die memory controller, AMD/ATi's on-die GPU and Sun's multi-thread handling and putting them on one chip?
Have Intel come up with anything genuinely new recently?
Goten Xiao
Intel has a lot of cash, and the ability to invest in expensive processes earlier than most. Certainly, earlier than AMD.
However, it's worth noting, that these are clearly AMD ideas.
* On die memory controller - AMD's idea - and it's been in use for quite a while now
* Embedded GPU - a rip off of the AMD fusion idea, announced shortly after the aquisition of AMD.
Intel is no longer leading as they have in yeas past - they are copying and looting their competition shamelessly. It appears that they are "leading" when point in fact it's simply not the case - had AMD not realeased the Athlon64 we would all still be using single processor NetBurst processors.
I can't wait for the Frodo and Samwise chips
In the meantime, you can get an AMD X2 3600 (65nm Brisbane core) for around $85 now, and probably in the $60 range well before these new products hit. The high end is one thing, but who actually buys it? Very few. I don't know anyone that bought the latest FX when it came out, or an Opteron 185 when they hit, or even a Core2Duo Extreme. All this does is push the mid- to low-end products down, and a ~$65 dual core that overclocks like crazy (some are getting 3 GHz on stock volts on the 3600) would seem like the best price/performance to me.
AMD's not out because they don't control the high end. Remember, you can get the X2 3600 w/ a Biostar TForce 550 motherboard at Newegg for the same price as an E4300 CPU (no mobo), and that's the board folks are using to get it up to crazy clock speeds.
It seems that AMD has lost, and I'm not trying to troll. It just seems that fortunes have truly reversed and that AMD is being beaten by 5 steps everywhere by AMD. Anybody have an opposing viewpoint? (Being an AMD fan, I am depressed.)
Look at the title of this thread: Intel Next-Gen CPU Has Memory Controller and GPU.
The on-board memory controller was pretty much the defining architectural feature of the Opteron family of CPUs, especially as Opteron interacted with the HyperTransport bus. The Opteron architecture was introduced in April of 2003, and the HyperTransport architecture was introduced way back in April of 2001!!! As for the GPU, AMD purchased ATI in July of 2006 precisely so that they could integrate a GPU into their Opteron/Hypertransport package.
So from an intellectual property point of view, it's Intel that's furiously trying to claw their way back into the game.
But ultimately all of this will be decided by implementation - if AMD releases a first-rate implementation of their intellectual property, at a competitive price, then they'll be fine.
The Tao of math: The numbers you can count are not the real numbers.
1. Putting a GPU on the processor immediately divides the market for it. Unless this is only going to be a laptop processor it probably won't sell well on desktops.
2. Hyperthreading only works well in an idle pipeline. The core 2 duo (like the AMD64) have fairly high IPC counts, and hence, low amount of bubbles (as compared to say the P4). And even on the P4 the benefit is marginal at best and in some cases it hurts performance.
The memory controller makes sense as it lowers the latency to memory.
if Intel wants to spend gates, why not put in more accelerators for things like the variants of the DCT used by MPEG, JPEG and MPEG audio? or how about crypto accelerators for things like AES and bignum math?
Tom
Someday, I'll have a real sig.
I can see those being quite hot for servers, where running "many small" tasks is where the game is.
On a desktop PC you often need the focused application (say, some sort of graphical/audio editor, game, or just a very fancy flash web site even) to get most of the power of the CPU to render well.
If you split the speed potential in 16, would desktop users see actual speed benefit? They'll see increased responsiveness from the smoother multitasking of the more and more background tasks running on our everyday OS-es, but can a mostly single-task focused desktop usage really benefit?
How of course, we're witnessing ways to split concerns of a single task application into multiple threads: the new interface of Windows runs in a separate CPU thread and on the GPU, never mind if the app itself is single threaded or not. That's helping.
Still, serial programming is, and is going to be, prevalent for many many years to come, as most tasks a casual / consumer applications performs are inherently serial and not "paralelizable" or whatever that would be called.
My point being, I hope we'll still be getting *faster* threads, not just *more* threads. The situation now is that i's harder harder to communicate "hey we have only 1000 threads/cores unlike the competition which has 1 million, but we're faster!". It's just like AMD's tough position in the past, explaining their chips are faster despite having slower clock-rate.
1: Integrated sells very well on the desktop almost every single machine in your big box shops has integrated graphics. I am sure it is outsells machines with separate graphics cards in the desktop. Gamers are not the market.
2: I am skeptical about hyperthreading, but it all depends on the implementation. I don't think this is something they are pursuing just for marketing. They must have found a way to eek out even better loading of all execution units by doing this. I can't imagine this being done if it actually performs worse than hyperthreading in P4. We have to wait and see.
Not quite. Intel projects are usually named after local geographical features, not all of them rivers. For example, Banias, Dothan, Yonah, and Merom (Centrino/core2 duo project names) are not rivers in Israel. Also, the first PIII project was done in Folsom and named "Katmai" -- again, there is no Katmai river in Northern California.
It's quite common in the industry to give projects names that don't mean anything, and each company uses a different scheme for generating the monikers. One interesting story is what happened when Apple used an internal project name of "Sagan". Carl Sagan took exception to this use of his name and threatened a lawsuit. Apple responded by changing the project name to "BHA", a TLA for "Butt-Head Astronomer". Sagan filed a lawsuit over this but it was thrown out of court when the judge ruled the new name was a generic one since Sagan was probably not the world's only butthead astronomer. (As least that's what I recall of it. Perhaps someone who worked at Apple during this time can add more detail?)
Looks like you've got one of those new computers that runs faster based on originality. I bet those Lian Li cases really make it scream then!
OK, these new parallel chips aren't even out yet, and software has to get the hardware before SW can improve to exploit the HW. But the HW has all the momentum, as usual. SW for parallel computing is as rudimentary as a 16bit microprocessor.
What we need is new models of computing that programmers can use, not just new tools. Languages that specify purely sequential operations on specific virtual hardware (like scalar variables that merely represent specific allocated memory hardware), or metaphors for info management that computing killed in the last century ("file cabinets", trashcans of unique items and universal "documents" are going extinct) are like speaking Latin about quantum physics.
There's already a way forward. Compiler geeks should be incorporating features of VHDL and VeriLog, inherently parallel languages, into gcc. And better "languages", like flowchart diagrams and other modes of expressing info flow, that aren't constrained by the procedural roots of those HW synthesis old guard, should spring up on these new chips like mushrooms on dewy morning lawns.
The hardware is always ahead of the software - as instructions for hardware to do what it does, software cannot do more. But now the HW is growing capacity literally geometrically, even arguably exponentially, in power and complexity beyond our ability to even articulate what it should do within what it can. Let's see some better ways to talk the walk.
--
make install -not war
http://www.anandtech.com/cpuchipsets/intel/showdoc .aspx?i=2955 provides a much more detailed look at the new processor architectures coming from Intel. A little better than the PR blurb at ars'.
It is interesting to note that Intel has now decided to put the memory controller on the die, after AMD showed the advantages of doing so.
However, I'm a little dismayed that Intel hasn't yet addressed the number one bottleneck for system throughput: the (shared) memory bus itself.
In the 90's, researchers at MIT were putting memory on the same die as the processor. These processors had unrestricted access to its own, internal RAM. There was no waiting on a relatively slow IDE drive or Ethernet card to complete a DMA transaction; no stalls during memory access, etc...
What is really needed is a redesign of the basic PC memory architecture. We really need dual ported RAM, so that a memory transfer to or from a peripheral doesn't take over the memory bus used by the processor. Having an onboard memory controller helps, but it doesn't address the fundamental issue that a 10 ms IDE DMA transfer effectively stalls the CPU for those 10 milliseconds. In this regard, the PC of today is no more efficient than the PC of 20 years ago.
The society for a thought-free internet welcomes you.
When I was your age, we used floppy disks as swap space.
:P (try THAT on for size!)
N 0C+N0CC+D+E8D+8C+4P2D+F+A2G+2F+2E2EG+2G+2F+EC+P4"
When I was your age, we overclocked our floppy drives.
When I was your age, memory upgrades came with a soldering iron and a hand-written instruction sheet.
When I was your age, L1 cache was just a really long wire loop with high capacitance.
When I was your age, computers booted in about 2/10ths of a second.
When I was your age, Compuserve was the world's biggest dial-up network
When I was your age, we didn't let teenagers post crap on the internets, we beat them into discipline with a belt and sent them off to bible school!
When I was your age, music piracy was a big problem: PLAY "T240MNO2 G+N0C+N0CC+D+C+P2P4D+F+EP2P4F+G+F+EP2P4G+F+EC+CG+
When I was your age, there was still money to be made in the computer industry.
-Billco, Fnarg.com
Redundant?
C'mon, modders, you can do better than that. Troll, Flamebait, Overrated, I'd understand; they're applicable. But redundant??
Besides, I was serious. When am I going to see some serious RAM on-chip?
110100 1101000 1101000 1100110 0 1101111 1101000 1100011 1
Yup.
"AMD has released their next generation of slow crap. With Intel breathing down their neck, can they survive for even the next few hours with such horrible products?"
vs
"Intel may release a product in the next 2-7 years. Can AMD survive?"
AMD has had on die memory controllers for how long now? Athlon 64 cornerstone was this feature. AMD has also developed and successfully integrated hypertransport in today's machines. AMD is also working on the same type of development, but they already has two key pieces already in place. Now with their acquisition of ATI i see them in a much better situation to implement this form of technology than intel. Intel can do it, but they have much more research to conduct and test before their chips will be ready. I believe AMD will quietly work on this, and drop it a year or so before intel. Then intel will be in the catchup phase again. This whole thing works in cycles, AMD and intel will constantly be swapping places on top of the mountain.
A lot of amd fanboys here. PErformance wise intels latest chips are winning versus amds chips. Without an integrated memory controller.
Now intel will have that also. This means that amds only advantage they will lose. AMD needs something huge to stay in the game.
Also baiscally this news sounds like Intels Version of the Cell processor.(they stated configurable cores that each core can be set to do something else.)
Generally for a type of memory, the larger its capacity, the larger its latency becomes and the smaller the throughput you'll get from it. A memory hierarchy is sometimes seen as a solution to reduce memory system cost, but more fundamentally, as silicon technologies evolve, it also reflects an inherent characteristic of memories - either large or fast, you can't have it both ways.
People who dislike China tend to mention Tiananmen Square a lot, but they always forget the Tank Man is also a Chinese.