Intel Opens Its Front-Side Bus 185
vivin writes "The Inquirer is reporting that Intel has opened up its FSB. Intel did this during IDF 07. What this means is that you can plug non-Intel things into the Intel CPU socket. The article says 'This shows that Intel is willing to take AMD seriously as a competitive threat, and is prepared to act upon it. In addition to this breaking one of the most sacred taboos at Intel, it also hints that engineering now has the upper hand over bureaucracy.'"
Re:Not the first time (Score:5, Informative)
IIRC, the socket-7 issue was not that Intel *wanted* others to use the technology, but rather that their license agreements with various other manufacturers allowed the rest of the industry to use it.
The only reason that Intel is opening up their FSB this time around is because they will be forced to use HyperTransport [zdnet.com] if they *don't* open it up (a royalty-free deal, to boot).
Their already using AMD64 and with AMD's new processors showing promise, Intel are really scratching and clawing here. I don't have the knowledge to pick a bus based on merit but, from what I've read, Hypertransport is better. Can anyone with experience here chime in?
Do we want Hypertransport or Intel's bus? What about licensing?
Not at ALL what you are thinking: (Score:5, Informative)
Intel is not trying to open their bus up to AMD. That is not at all the goal. First of all, access to the the Intel bus requires a license. I'm not sure Intel would even grand AMD one for a sane price. Second of all, AMD would in no way want Intel's bus. As has been the hot topic of discussion for over a year, AMD's HT (HyperTransport) point-2-point links are faster both in terms of bandwidth, and latency than Intel's FSB. HT uses less pins than Intel's bus, and HT devices are simpler, cost less, and use less power. HT is a pretty neat and effective technology. Intel's FSB on the other hand, is much the same as it was around 10 years ago. To answer your question, AMD would take a massive hit by going to Intel's POS bus. It's funny, ATM, AMD has the better bus/platform and Intel has the better core. No one here seems to realize that AMD would never be willing to throw out their main advantage right now... AM2 isn't the issue. The issue is HT. Hell, even IBM announced that Power7 will use AMD's HT links. No one will be dropping HT for the POSFSB any time soon.
Intel/AMD are only opening their sockets/buses in an attempt to get third party developers to make FPGAs, JAVAics, and other accelerators. AMD has had some luck with this, and one can buy co-processors that drop into an AMD socket today. Intel is trying to get the same benifits, but I don't really see the point until Intel can get CSI working and drop the antiquated FSB.
AMD vs. Intel, but not so literally. (Score:5, Informative)
That was a while ago, but I suspect it's coming to fruition or perhaps gaining more traction, if only now Intel is saying "me too."
http://news.zdnet.com/2100-9595_22-528221.html [zdnet.com]
Re:Trend (Score:3, Informative)
The reason it's a response to AMD (Score:4, Informative)
And yes, the bus speed matters. I've seen neural net tests in which Woodcrest, for example, does much better at 1333MHz using four cores than you'd see at 1066MHz. That's the same architecture except for bus speed. AMD's memory bandwidth is still better, though they lag in other areas.
I don't know whether, or how much, you'll see that bus bandwidth matter in the typical slashdotter workload (games).
Re:didn't it used to be this way? (Score:3, Informative)
chip speed (bus speed @ multiplier)
386dx 40 (40 @ 1x)
486dx 50 (50 @ 1x)
486dx2 66 (33 @ 2x)
486dx2 80 (40 @ 2x)
486dx4 100 (25 @ 4x, 33.3 @ 3x, even 50 @ 2x with proper cooling)
486dx4 120 (40 @ 3x)
I used to have this one but sold it:
k6-2 350 (100Mhz @ 3.5x)
AMD had more chips than this, including the k5, k6, and k6-3. I never owned any of those, so I don't remember the specs off the top of my head. After the k6-2 and k6-3 came the Socket A and Slot A Athlons and Durons. I won't get into history that recent.
I have a Cyrix 6x86 150+ which was a 120Mhz chip running on a 60Mhz bus at 2x multiplier. It really would keep up with a Pentium 150 on stuff written for a 486. However, it wouldn't run a lot of software optimized for the Pentium because it wasn't fully compatible. Like the original Pentiums, it didn't have MMX, either. The 6x86MX line did. These were also known as the M1 (6x86) and M2 (6x86MX) lines of chips. Cyrix is now part of Via.
Many older motherboards (socket 3 and socket 7, for instance) often let you change your bus speed, voltage, and multiplier with jumpers on the board. It didn't keep your chip safe, but if you could figure out a way to overclock without burning it up you were free to do so.
Intel also had the dx50, BTW. Lots of my friends have or had it. I also know people who used to run the Intel dx4-100 at 50 @ 2x (I know I did) even though Intel advised against it. Socket 7 for Intel was followed by Slot 1 and Socket 370.
Eh, um, no. (Score:3, Informative)
Re:Not the first time (Score:5, Informative)
Do we want Hypertransport or Intel's bus? What about licensing?
HT can run with approximately twice the number of transfers per second per pin as current-generation Intel FSBs. HT is also more readily expandible to use more pins, because it's an autonegotiating variable-width bus, similar to PCI-express. It also wastes fewer pins on control signals. HT is clearly the best, technologically.
Licensing wise, HT is licensed "royalty-free" for an annual fee. I don't believe the fee is particularly large. Many chip producers have already licensed it and will license modules to connect your own chip design to it for very small fees. Such modules exist on some modern FPGAs. This is not currently true of the Intel FSB spec.
Re:A lot of it depends how quickly you change CPUs (Score:5, Informative)
100GHz is probably pushing it. You'll note that we haven't seen a huge increase in clock speeds recently, but rather continuing increases in instructions per cycle. I'd guess we'll reach a plateau somewhere around the 10GHz mark.
Moore's Law will soon hit a much more fundamental law: physics. You can't keep shrinking transistors like they are at the moment; it was predicted that we'd reach the limit years ago (yes, I too remember the advent of 200MHz desktop processors, and thinking they couldn't get much faster), but the fact we haven't so far doesn't mean we won't. Moore's Law demands a shrinking by a factor of 1.4 every 18 months. We're currently on 45nm. This gives us the following trend:
end 2008 - 32nm
start 2010 - 22nm
end 2011 - 16nm
start 2013 - 12nm
end 2014 - 8nm
start 2016 - 6nm
end 2017 - 4nm
4 nanometres is only 38 atomic radii of silicon. It seems unlikely that a transistor this small could be produced. Therefore, as long as we continue to use silicon transistors (and no promising alternative that solves this issue exists right now) we will see the end of Moore's Law within the next 10 years. I'm sure of it.
And an end of Moore's Law will not only slow GHz increases, but also will slow the adoption of larger numbers of cores, because without shrinking transistors the only way to increase number of cores is by having a larger die size, which is more expensive and requires larger chip size, which requires larger system board size, which requires larger case size, which consumers don't like.
Re:wow (Score:1, Informative)
Wrong again. Current HT does not scale well past 4 sockets.
It uses broadcasting cache coherency,
while 8-64 sockets system use directory based cache coherency system.
Your 8-64 socket Pentium has a whole lot of extra glue logic "on the motherboard" to cope with this. Have you seen the diminishing returns you get on 2,4 and 8-way pentium (Xeon or Core or whatever) systems? You can put extra glue in an Opteron system too.
AMD on the other hand can show very nice performance on synthetic memory benchmarks.
I've seen 4-way Opterons boxen smoke 4-way Pentium boxes with much higher clock frequencies. Mind you, they were running Solaris, not Windows.
I really shouldn't feed the trolls.
Re:Does this really make sense? (Score:3, Informative)
That's just wrong. AM2 was simply AMD switching to DDR2 RAM. It didn't cause a performance drop, just no immediate performance improvement over socket 939 with DDR, and there's nothing they could have done to change that, except trying to force manufacturers around the world to produce faster DDR RAM.
Even with the higher latency of DDR2, AMD still has a much faster bus, and lower latency, than Intel. And even if the opposite were the case, there's no benefit to AMD of switching... Their on-board memory controller is a big benefit of AMD64, and switching to the standard FSB model would be a serious step backwards for them in performance.
Except for the fact that you haven't the slightest idea what you're talking about...