Slashdot Mirror
Dell Set to Introduce AMD's Triple-core Phenom CPU
An anonymous reader writes "AMD is set to launch what is considered its most important product against Intel's Core 2 Duo processors next week. TG Daily reports that the triple-core Phenoms — quad-core CPUs with one disabled core — will be launching on February 19. Oddly enough, the first company expected to announce systems with triple-core Phenoms will be Dell. Yes, that is the same company that was rumored to be dropping AMD just a few weeks ago. Now we are waiting for the hardware review sites to tell us whether three cores are actually better than two in real world applications and not just in marketing."
I don't understand why they disabled it in the first place. Anyone care to explain?
In multicore systems each core can only talk to two other cores.
With a quad core system, each core cant directly talk to the core diagonal to it which slows things down.
Three core systems can talk among the cores easily without any bottlenecks so they are faster than dual core and quad core.
I don't think so. Not in the case of triple-core processors that are just quad cores with one core disabled. In this triple-core configuration, one of the cores is connected to the other two, but there is no direct connection between those two that doesn't involve the third core. Visualize it as cores 1, 2, 3, and 4, cores 1 and 2 at the top, cores 3 and 4 at the bottom, 1 is connected to 2 and 3, 4 is connected to 2 and 3 but 1 is not directly connected to 4. Core 4 is the unlucky one, and there is no direct connection between core 2 and core 3, so if core 2 wants to talk to core 3 it has to go through core 1.
A native triple-core would have equal spacing between the three cores such that any core could talk to any other core without having to go through a middleman.
"There is much pleasure to be gained from useless knowledge." - Bertrand Russell.
Why?
If one is disabled, it would cycle 1,2,4,1,2,4 (assuming #3 is the bad one).
Moreover, if one of the cores isn't running, and you have a cooling system designed for four cores, it really doesn't matter. If it can handle four full-tilt cores, it can handle three. The zero heat production is a bigger benefit than a slightly uneven distribution. If it's truly a suitable medium, the heat generated will be spread throughout pretty well, even if the heat-production is only on one edge of the medium. Think of an electric stove burner--it only has heat applied at one end, but the opposite end heats up pretty well. Obviously it's not perfect, but it doesn't need to be.
Ironically, the main advantage of dual-core has nothing to do with applications taking advantage of that second core -- in fact, just the opposite.
Dual-core means that for most cases, I can run a video encode, a backup/compression process, a long-ish compilation (of the sort that doesn't like 'make -j2'), etc -- not so much all at once, as I can fire off any background process and not worry about it, as I have a whole other core to use. It's shameful -- Amarok will occasionally use 100% of one core, and I won't notice for hours.
Having more than two cores wouldn't benefit me a lot right now. I wouldn't mind it, certainly -- I've been playing a bit with things like Erlang, which should be able to scale arbitrarily -- but I think the real applications are only just catching on to the idea that threading is a good thing. I imagine it's still going to be a lot longer till a quad-core machine is useful for anything other than, say, running virtual machines, as most programming languages do not make threading easy. (Locks and semaphores are almost as bad as manual memory management.)
While I'm playing crystal ball, I'll predict that the first application of multicore will be things which were already running on multiple machines in the first place -- video rendering, for instance. Not encoding, rendering.
The second application for it will be gaming. This will take longer, and will only be the larger, higher-quality engines, who will simply throw manpower at the problem of squeezing the most out of whatever hardware is available.
I suspect that the old pattern will be very much in effect, though -- wherein gamers will buy a three-core system and unlock the fourth one (if possible), then use maybe one core, probably half of one, with the video card still being the most important purchase. If there's a perceptible improvement, it'll be because their spyware, IM, torrents, leftover Firefox with 20 MySpace pages and flash ads, etc, won't be able to quite fill the other three cores.
I'd like to add that for most people, including me, one core is plenty if you know how to manage your processes properly -- set priorities, kill Amarok when it gets stuck in that infinite loop, and get off my lawn!
Don't thank God, thank a doctor!
I haven't really looked at Phenom's design, but I highly doubt that it'll rotate between the cores while running. You can't really transfer the contents of registers and whats in the pipeline between cores in any sort of efficient manner (unless there is something about the Phenom I don't know about).
Why would the thermal design even matter that much? It'll be equivalent to having hotspots on the motherboard (though nowhere near as dramatic, the die is tiny and very conductive to heat). By simply having a heatsink on it that can handle four loads, it'll easily be able to handle 3 active (and 1 idling) core.
I thought that this was a "native" quad-core CPU? I know that Intel does the dual-dual-core thing, but it was my understanding that the reason AMD is touting their quad (and this triple) core as being better than Intel because of this fact.
Then again, I haven't been following CPU product lines in the past few months, so I could be mistaken.
In the end, this CPU will enable AMD to yield more CPU's and actually turn profit, but it won't be on market too long once AMD perfects the process and yields working quad-core chips most of the time.
- It's not the Macs I hate. It's Digg users. -
I think I remember reading an article on Tomshardwareguide where they tried running one dual core, and a single core CPU in the same system for 3 cores. While they got it to boot the OS, a lot of applications failed to run.
I'm guessing there is a lot of code out there that's looking for power of 2 number of cores. A program might run fine with 1,2,4,8, or 16 cores, but if you do some kind of odd number I wouldn't be surprised if several applications just refused to run. It will be interesting to see what kind of compatibility testing AMD has done with this new processor.
In the end though, this just seems like another last ditch attempt by AMD to marginally compete on the lower end market with Intel. Intel says they have no need for 3 core chips since their yields are so much higher.
There are a couple known problems with the first spin of the Phenom die (codename Agena).
The first (and less relevant) problem is the TLB errata. The second (and more relevant to this discussion), is a problem in which core #2 (out of [0,1,2,3]) is lower yielding than the first three. For example, on the same CPU die, cores [0,1,3] may work fine at 2.6Ghz, but core [2] yields only at 2.0GHz. This is a widespread problem, mostly found out through failed overclocking attempts.
Google it yourself and find out..
The distinguishing feature is often the number of tests done to certify the hardware and in some cases it is not a failure in a certain test but that the test required for the higher spec was not done at all. The rumor with the Celeron mentioned above was that they were rebadged after passing all the tests required for the Pentium II 450 spec but there were a lot of them in storage and more Celeron 300's were required - so they got the "A and circle" symbol to distinguish them from the other Celeron 300's.
This used to happen even in the 486 days already.
:-( )
486es with a working co-processor (Floating Point Unit) were sold as "DX" models, the ones where it was broken were sold as "SX".
Even better, it allowed a market for FPU co-pro upgrades where one would install a co-processor upgrade alongside their 486SX later on.
Once production yields improved, this practice was continued for a while maintaining a market for both "SX" and "DX" models, where the "SX" models would have their FPU deliberately disabled. What on earth moved AMD and Intel not to simply start selling the "DX" processors at a pricepoint closer to the "SX" ones, I don't know.
The DRAM market has been much the same for even longer. The ZX Spectrum (Timex in U.S.) 48K model had in fact 80K of possible RAM on board. The first 8K were a sort of memory swapping / paging bank, and the remaining 40K actually consisted of DRAM chips where only half of the chip worked, which were cheaper than even the half-size but fully working ones. Replace those DRAM chips with fully working full-size ones and you'd have a whopping 80k in your computer.
(This post outs me as a dinosaur fossil, doesn't it?
What about the original Athlons and Durons? The only difference between them was often a cut link on the top surface of the CPU that disabled moste of the cache. Have a Google and you'll find lots of instructions on how to remake the link and turn a Duron into a fully functional Athlon.
It's all about economics and "perceived value", not technology.
I've hacked a couple of graphics cards by moving a resistor on the top of the chip. One was a GeForce and it came up afterwards as a "Quadro". The other was an ATI 9500 which came up afterwards as a 9700 (more shaders). Both cards worked perfectly for years.
No sig today...
ok, I am an IC test engineer:
#1: you do test these chips before the saw step (chopping the wafer up into individual die)
#2: its hard to predict speeds/vcc/temp sensitive yields at that stage, but you do test all the die and usually check for full functionality (as much as the test coverage allows)
#3: once packaged, the chips are "binned" to functional fails, speed grades. etc, and are tested at temp, vcc limits for speed sorting. so you could have 1 core that fails at 30C with a high vcc, but the others are ok (this is should be rare since they all sit together on the wafer in close proximity, and thus shouldn't vary much from each other)
#4: nanoscopic defects occur and could take out one or two of the die. It would be advantageous to bin this out as a tri/dual core.
#5: I am 100% sure that if these become popular, there will be some chips that pass all tests fully, but have one core disabled. happens all the time.
JP
jp