Slashdot Mirror
AMD and Intel Update CPU Roadmaps
vincecate writes "Recently
AMD updated their processor roadmap. It shows their move to 90 nm and has a range of new processors over the next 1.5 years, including dual-core chips. An
unofficial AMD roadmap shows speeds and performance increasing.
Intel also recently updated their roadmap.
Intel does not show anything faster than
the current 3.6 Ghz in the next 11 months, including the recently delayed 4 Ghz chip, except to say '3.6 Ghz or greater.' Strangely, some of the recent SPEC benchmark results show the 3.6 Ghz chip to be slower than the 3.4 Ghz chip. One possible explanation for this is that the 3.6 Ghz chips will slow down due to 'thermal throttling' if you are not very careful to keep them cool. So it seems like heat may be the reason Intel's roadmap does now show much improvement."
Color harmony
The reason the 3.6GHz processor runs slower than the 3.4GHz processor is because they're different processors, not the same processor running at different clockspeeds. Just look at the die photos (www.chiparchitect.com) and you'll see what I mean. The idea is that the new processor will scale to higher clockspeeds which it, uh, already has. (Just look at the "OC records": nobody got an old Pentium 4 beyond 4GHz with standard HSF cooling - nobody. On the other hand, this is more or less straightforward with the new Pentium 4s.
What I don't understand is why more people aren't building Pentium M desktops.
Actually, during a review session with the 3.6GHz LGA775, we experienced so high heat production that we had to use water cooling to ensure that the thermal throttling was not enabled. When using regular air cooling the processor would reach temperatures > 70 degrees during load, and from the results at this load we saw that at some times the processor had to use thermal throttling.
Another disadvantage with this high heat production is that other core components in the computer (such as the mainboard) will be exposed to more heat as well, hence the durability of these components will be lower.
If Intel and AMD continues to approach Itaniums heat production, water-cooling or similiar technologies will become mandatory for high end processors.
GPUs are not powerful, they are just very efficient at matrix manipulations and calculations related to graphics operations. You can use this to increase performance in some situations, but 99.99% of the time you'd just bottleneck yourself.
Last week I benchmarked the 2.2Ghz Opteron on 64 bit Linux 2.6 and Java. I got almost three times the performance of a 3Ghz Xeon. For details see http://gregluck.com/blog/space/start/2004-07-29/1# AMD64,_JDK1.5.0_and_Linux_2.6_rock!/
- Get yourself first cheap Athlon XP you can see
- downclock and undervolt
- result: cool CPU
I've got a 2GHz Celeron laptop that is much slower at compile than my Athlon 1.47GHz.
The Celeron is a severly crippled chip, unlike the Duron, which is a respectably performing budget processor. It only has 128KB cache, which is CPU sucide on a P4 core. The P4 needs large amounts of cache to keep its long pipeline filled. People who buy high clock speed Celeron, thinking they're getting a fast CPU are getting massively screwed by Intel. So much so it borders on being an unethical and immoral business practice. The chips are not near as fast as their clockspeed would indicate. One would be much better of with an Athlon XP, Duron, or a slow P4 as a budget processor.
In addition to this, until we start seeing widespread use of PCIe, the downstream AGP bus is still a serious bottleneck as well. Uploading data to the GPU is really fast, downloading maxes out at ~133MB/s.
I haven't had the chance to play with a Pixel Shader 3.0 card yet, so I don't know how useful for generic computation they are. It usually helps if you're trying to process many sets of the same kind of data, rather than evolving one calculation through a long or iterative algorithm.
~phil
You can have those, just not at the same time. Via Eden runs fanless. But it's still 32bit! And it doesn't run in SMP-configurations (yet. there has been some info about SMP-solutions).
I think you could buy an Opteron 2xx-machine, underclock it to around 1GHz so it might run fanless. Then you would have your fanless 64bit SMP-machine,
Lesbian Nazi Hookers Abducted by UFOs and Forced Into Weight Loss Programs - -all next week on Town Talk.
A dual 3.2 GHz P4 would host a web site quite a bit larger than "medium sized." Consider this: for a web server, the vast majority of the computational resources are spent in bandwidth. If you had a number of static pages, you could probably serve up web pages from a single 1Ghz to millions of visitors a day[1]. Anything more than that, you'd need for running scripted pages. But even then, we're not talking about an awful lot of horsepower for web development (unless an incredibly bad programmer doesn't realize you need to max out the CPU to post a form)
[1] Watch when a site gets /.ed. With the exceptions of the articles where someone is running apache on their wristwatch, you usually can still get a small trickle of bandwidth from the web site. Depending on how the web server handles things, it may drop your connection randomly to prevent DOS attacks (OpenBSD). But you'll see the server is still up, it's the bandwidth that is the problem.
Also, a technical point: the overwhelming majority of heat sinks do not weigh more than a pound, or even terribly close to that. The specification calls for a HSF combo of at most 450g. I know for a fact there are HSFs that weigh more than that, because I have one and it came with a warning about it; but even among aftermarket HSFs targeting overclockers it's very rare. I realize you're exaggerating to better illustrate your point - and I fully expect that within a year or two we'll have passed the 1lb figure - but I wanted to make sure everyone else realized that as well.
If you want to complain about weight, heat, and power consumption, set your sights on video cards. Mine is terrifically heavy and has its own goddamn power connector. The 6800 Ultras need two power connectors. My entire 7-drive backplane needs two power connectors!
Just FYI, 1 lb = 454 g
Prescott is a massive CPU with a 31 stage pipeline, compared to the older P4's 20 and the Athlon XP's 12. I'm not sure off the top of my head how many stages the Athlon 64 has.
I thought XP had 10 and A64 12..
I was talking to my friend about this the other day, and we think that eventually they cannot go that much faster (well, maybe have a SMALL core of the chip that can go faster), and they'll start stacking in parallel instead. Ie, massively hyperthreaded processor cores. So maybe in a few years we'll see 6 GHz chips with 8 or 16 hyperthreaded processors?
We're physicists, though, not engineers, maybe there are some other clever ways to keep pushing the envelope?
make world, not war
FYI - I have an Athlon 64 and heat hasn't been a problem at all. I just have the retail processor with the el-cheapo heatsink that it comes with (nothing fancy - just thermal compound and a reasonably-sized sink). I haven't seen it exceed 55C under heavy load. Granted, my case is fairly well-ventilated, but nothing excessive (well, the case was excessive, but I unplugged about half the fans). Oh, did I mention that I overclocked it by about 8% or so?
AMD used to have a high-heat reputation and used to be known for difficult-to-overclock processors. Honestly, I don't think that is nearly as much the case with their newer processors. The Athlon64 seems to run fairly cool, plus it supports frequency scaling when it isn't busy (note - the 55C figure I gave was under heavy load for considerable time - no scaling in effect). Right now, I'm typing on the machine and the CPU is reading 37C - only 1.5C higher than case temperature.
I think AMD is actually passing Intel in this respect. Intel had better watch out if they expect year-long delays - eventually AMD will be releasing 3-4GHz Athlon 64's and they'll be FAR faster than anything Intel currently has...
Round wafers are easier to manufacture. Keep in mind that the the Si chips are made of is monocrystalline. When growing the wafer it is grown from the center outwards. When it hits the wall of the reactor vessel, it will probably break the crystal structure, and whatever growth that continues after that is not usable as there is a grain boundary. With a circular wafer you hit the edge at the same time.
Also, the wafers used today are what, 300 mm in diameter while the chips are something like 10x10 mm, so there's not much material lost anyway. And the leftovers are simply sent back to the wafer factory to be remanufactured into new wafers, so there's no material lost. Not that it would matter anyway, since Si is among the most abundant materials on earth.
That's the problem Intel has right now, really. Marketing seems to say, "Make it sound faster", only looking for good warrior CPUs in the Mega Hertz Wars. IBM/Apple and AMD have not been trying to go for faster clock speeds but instead for faster CPUs.
Such long pipelines as the Prescott line may help achieving higher clock speeds, but 31 stages means that you'll see more pipeline stalls, so your CPU is happily running at higher clock rates, doing nothing. Of course, not all instructions actually have to go through all 31 stages, but still, it's impractical to have so many stages in an architecture when you know that every so-many-but-fewer-than-31 instructions you're going to hit a branch. Not to mention the additional complication for the on-die dependency tracking that you need in out-of-order cores like Prescott.
Of course in-order architectures with full predication ISAs would solve some of the problems with longer pipelines, but I guess we can't say that this other Intel architecture, ia64, is such a great success ;-)
So what kind of apps (I mean, apart from Doom 3) do end users need this kind of grunt for? 3GHz? 3.6GHz? 4Ghz?! If Architects could use AutoCAD 2000 on a 950MHz cpu, without complaint, what has changed? Obviously a speed increase is nice, but three or four times that?
IDL, IRAF, Mathematica, Matlab, etc. In other words, physicsists and astrophysicists can always use faster computers for their everyday work. Even more so, (astro)physicists running fluid dynamical systems of galaxies need every bit of speed possible.
Granted going from 8GB to 16 GB of RAM is the largest benefit but increasing a CPU from 2 GHz to 4 GHz would also help out a lot (though 3.2->3.4 would not be worth it).
I also know some molecular biologists that run detailed simulations about, umm, molecular biology (excuse me but I am not a biologist) and are always looking to get all the speed they can as well.
While scientists certainly do not represent the average computer user, there is demand for fast computers in research.
I hope someone appreciates this as I am giving up my right to moderate but felt that I should mention it because I didn't see any mention of it.
Photons are the mediating particles of electromagnetic force, and it's definitely this force that couples two electrons together, or the electrons to the 'holes' in the doped semiconductors, etc etc. An elementary description of current in a wire is akin to a tube filled with marbles, you push one in, and one comes out at the far end. This interaction between the 'marbles' would be mediated by photons. Of course metals and semiconductors are far more complicated than this picture, but it's a rough start.
It might sound weird to you (it did to me at first), but when you send a 100 MHz signal down a coax cable, you are really sending photons. They're rather low-frequency photons confined to a waveguide, but they're definitely photons.
make world, not war