Industry Update - Q4-2004: AMD adds SSE3 Support, Intel's 925/915 not selling and more
by Anand Lal Shimpi on November 3, 2004 1:00 AM EST- Posted in
- Motherboards
AMD Athlon 64 Revision E adds SSE3 Support
As we discovered most recently, AMD's 90nm CPUs are of a brand new revision that featured a number of bug fixes and performance improvements. Internally AMD refers to this CPU revision as Revision D. However we noticed something very interesting when looking at the latest AMD roadmaps while in Taiwan: the 90nm chips listed as San Diego and Venice (Athlon 64 FX and Athlon 64 respectively) claimed SSE3 support as a feature, but the current 90nm chips do not have SSE3 support.
We went around to quite a few manufacturers asking what the difference was between the 90nm chips shipping today and the 90nm chips that supposedly feature SSE3 support, and unfortunately we were left with very little information - until we learned to ask for the right thing.
Internally, AMD's San Diego and Venice CPUs are referred to as nothing more than Athlon 64 Revision E chips. Revision E includes even more bug fixes and performance improvements over those we found in Revision D, including support for the 13 new instructions that were added with Prescott, more commonly known as SSE3.
The performance enhancements that go along with Revision E chips include some optimizations to the Athlon 64's memory controller. The more optimized memory controller improves bandwidth efficiency with regards to unified graphics memory accesses; given that the only type of graphics that uses system memory (and thus the on-die memory controller) is integrated graphics, it's safe to say that the Rev E chips will offer better integrated graphics performance.
The first Revision E CPUs will begin shipping in early 2005. AMD also has plans to introduce an Athlon 64 4200+ towards the middle of 2005; they are not listing whether the part will feature a 512KB or 1MB L2 cache, but it will most likely run at 2.6GHz. The Athlon 64 FX-57 is also listed on the roadmap as a 2H-05 part, it's specifications are also unclear but we'd expect it to be a 2.8GHz part with a 1MB L2 cache. Both the Athlon 64 FX-57 and Athlon 64 4200+ appear to be 90nm parts built on the San Diego and Venice cores, respectively.
AMD's 2005 roadmap did not specifically list anything faster than the 4200+, although the classic "> 4200+" was present on the roadmap to indicate potentially faster parts.
On the Intel side of things we heard the name Cedar Mill thrown around a bit more. In the past some have referred to Intel's dual core Pentium 4 as Cedar Mill, but we now know that is Smithfield. This time we heard Cedar Mill referred to as the first 1066MHz FSB Pentium 4s that aren't Extreme Edition chips. Cedar Mill is supposed to be out in the 2nd half of 2005, which supports a claim we made in our recent review of the Pentium 4 3.46EE: "the 3.46EE will be followed up by the 3.73EE as the only two chips to support the faster FSB for almost a year."
So what does having no mainstream 1066MHz FSB chips until the end of next year mean? It means that Intel is holding out for Glenwood and Lakeport and that Sunday's launch of the 1066MHz FSB was merely a gimmick. The 925XE will be buried as the chipset that offered support for two CPUs that no one bought, while Glenwood and Lakeport will be the knights in shining armor that brought 1066MHz FSB support to all.
Intel seems to have learned from their 925X and 915 chipset launches - multiple fundamental technology changes without performance gains don't go over well. Glenwood and Lakeport will be able to support faster DDR2 (DDR2-667), 1066MHz FSB and dual core support upon their release, not to mention that lower latency DDR2 should be available by then (mid next year) and PCI Express graphics should be much more prevalent as well.
Facebook
Twitter
RSS
61 Comments
View All Comments
tennesota - Sunday, November 7, 2004 - link
Thanks for the informative article. I'm sure many would agree that "we appreciate your hard work work". Thanks!ThePlagiarmaster - Friday, November 5, 2004 - link
#59 I agree pcie will be much better for boardmakers (I don't think we'll see price cuts...not much anyway). Boardmakers will just take the profits and I can't say I blame them with razor thin margins. At least they have a good excuse to take it. I also agree that pcie will be great for OTHER cards. I may have misunderstood your msg. It seemed to me you were extolling the virtues of it as a new video card bus. In this regard I think AGP was fine. It's not even used. We haven't even tapped 256mb cards yet (witness the 6600gt's scores @128mb onboard). So plenty of room to grow on agp if you ask me as everybody fears going to the 'dreaded' bus :) But as long as it keeps board costs down I guess I see the point.What I meant by not needing ddr2 is Intel should have waited for a technology (ddr3?) that would have actually made a difference over todays memory. To release a product that doesn't outperform yesterdays product is stupid (the P4 vs. P3 comes to mind, 3 die shrinks on the P3 would have made one hell of a chip). DDR2 is [and will be] slower than DDR1 until 667 (barely a victory if at all) and probably 800mhz. Perhaps they'll be able to get DDR2 to dizzy heights like ddr1 (DDR2 1200?) and finally make it really worthwhile.
A dualcore could probably do the same as a dual cpu does now. Each opteron has it's own bank of memory. You can get one of these boards for only $216 or so (and you wouldn't need all the routes for the 2nd cpu socket so it should be cheaper still). So one bank for the right side of the chip and one for the left :) Nevermind, technology marches on and I accept it LOL. I shouldn't even be talking I guess, I have an A64 939p (3200+ probably), Koolance Exos, 6600GT PCIe, and the K8N Neo Platinum/SLI on the books as an xmas gift to myself...ROFL. Or I mean a gift to my EARS! OK the performance will be nice too, but my ears will love me for sure.
slashbinslashbash - Friday, November 5, 2004 - link
#57 I'm fully aware that PCI-E offers no current benefit for graphics cards in terms of bandwidth. My point was about the underlying PCI-E architecture. You're forgetting that PCI-E is *not just for graphics cards*. Normal PCI (and its bastard offspring AGP) have been around for too long. Like Parallel ATA is being removed in favor of Serial ATA, PCI (a parallel architecture) is being removed in favor of PCI-E (a serial architecture). Look at all the traces on your motherboard that go to a single PCI slot. A x1 PCI-E slot (analogous to a regular PCI slot) needs 4 wires. This clears up huge issues for motherboard manufacturers and will ultimately make motherboards cheaper. It's also point-to-point, and the bandwidth is much more scalable. PCI hardly even has the bandwidth for gigabit ethernet. A new technology is needed to replace PCI/AGP. Even though graphics cards will see no benefit today, the fact that there are now lots of motherboards on the market with PCI-E slots (both x16 and x1) means that the peripheral manufacturers can move to PCI-E and dump the outdated PCI. Yes, I know about 64-bit and 66MHz PCI. Those are better performing than normal PCI, but they're even more of a pain to route and implement on a motherboard. That's why you only see them on expensive server motherboards.I also question your statement "So we could have just skipped DDR2 and waited for DDR3." What makes you think that we'll get to DDR3 if nobody uses DDR2? Surely the same sorts of issues that now cause headaches with DDR2 will also be in DDR3. I agree, it doesn't seem to make much sense for Intel in the long run since they're likely moving to a new architecture that won't be as bandwidth starved. However, don't forget that the upcoming dual core processors will likely need more bandwidth even if they run at a lower clock speed. Dual-core Athlon64s might even benefit from DDR2, even if single-core A64s don't need the bandwidth.
Ivo - Friday, November 5, 2004 - link
Hi Anand, great review! Thanks!As for what I want: I'd like to build an upgradeable, micro-ATX size, cool and quite home-PC (something like DTR, but with separate screen). Will anybody produce a micro-ATX (or BTX) Socket 939 motherboard in the near future?
ThePlagiarmaster - Thursday, November 4, 2004 - link
#53 PCIe is NO better than AGP. Yes it has more bandwidth BUT it's the same thing as before. If you never go onto the bus who cares? We'll keep seeing bigger and bigger amounts of memory on the cards (because memory makers keep shrinking the chips like everyone else, thus making bigger amounts cost the same as the last smaller revs) thereby killing the very point of more bandwidth for the bus. I don't forsee ANY time in the future where ANY game maker will want to go out onto the bus when there is a ton of memory on the card that runs circles around main memory with to-and-fro bus access. It's just a completely BAD idea. It will never be good. The only purpose that would serve is CHEAP systems playing an old game or two. Or business systems that want to keep costs down. It will never work for gamers. Even crap cards come with 128meg today.AGP could have been good (and was a decent idea before cheap graphics memory killed it) if memory would have stayed expensive. But it was already getting really cheap before AGP even hatched. So the whole thing was a moot idea forever all the way through 8x. Just a BS line Intel could say they were always ahead of AMD in. Marketing hype and thats about it. It got cards off the PCI bus, but Micron showed you could do that with a 66mhz PCI slot.
You could argue it raised the voltage/watts to the cards (agp and pcie), but anybody could have pulled that off with a 66mhz pci slot also. However Intel is always looking for a way to claim some technology advantage so we keep getting all these BS changes (DDR2 anybody?). Let's face it, AMD could run on DDR1 until at LEAST 533 (1066 hypertransport and DDR 533 perfect match). You could even argue for DDR1 600 as it's around now (no doubt a die shrink again would make this memory cheap too). So thats two more revs of A64's, a stop at 1066 and 1200. Neither of them would starve the A64 either. It's nowhere near starved at DDR400.
So we could have just skipped DDR2 and waited for DDR3. DDR2 won't show anything for the P4 until 800mhz really and Intel will be talking DDR3 by then (ROFL - AMD doesn't have it but we do - nana nana nana). Just more marketing hype again. Though I like the idea of the power savings for notebooks. But not much more than that. We haven't even proven how many extra minutes that gets you yet (may be a puny savings, who knows). By the time they get DDR2 rolling the P4 will be dead and a P3 dual core (eh, I mean Pentium M, heh) that isn't bandwidth starved won't need the extra bandwidth Intel will be telling us we need. I see joke after joke after joke coming. The joke's on the consumer who buys into this load of BS every time they tell us we need it.
quasisnig - Thursday, November 4, 2004 - link
Anand,Excellent, I can only imagine the number of industry analyst that read your articles.
jiulemoigt - Thursday, November 4, 2004 - link
#47 people switched to the NForce2 chipset because of soundstorm... plus if Nvidia simply pkged it as both digital and analog PCI-E 1x riser card it would have the bandwidth it needed and met the needs of those you want an add on card and those who just want cheap sound. Personaly I'd love to see SoundStorm2 be dug out of what ever file they left it in and made it avalibe and a chipset that could be added to mobos, the way the SATA controllers were added and the AC97 chips are added, that cheap boards could have AC97 and mobo mkg who wanted to offer high end boards with it could. I paid $250 for my DFI Nforce2 board despite having an Asus dlx already because i liked it better. I have seen people all across the net who have wanted better audio without giving into creative which makes horible drivers.quanta - Thursday, November 4, 2004 - link
#41, nowhere I have claimed that Intel processors will stop supporting the rest of SSE3 if HyperThreading is turned off. However, those workstation application developers can write (and have written) boneheaded pro-Intel detection routines that don't turn on advanced AMD features that are supported by CPUs of both companies (eg: SSE in Windows Media Encoder 4 through 7.1), which will be used by Intel to fool uninformed buyers (though it won't be as bad as BAPCo SysMark 2002).Even without Intel's 'help' in marketing department, AMD does not work as tightly with complier builder as Intel does (after all, Intel does makes its own compiler). That can mean those dumb CPU detection routines are likely to cripple AMD CPUs a lot more than Intel's will.
slashbinslashbash - Thursday, November 4, 2004 - link
Great article. I'd just like to comment on the people who are hating on Intel for forcing the move to DDR2 and PCI-E. It's just one of those things that more or less had to happen. Dell is going to sell whatever Intel tells them to. Dell sells more computers than anybody else in the world. That means that Nvidia and ATI will have to make PCI-E cards available in mass quantities, and memory makers will have to make DDR2 chips available in mass quantities. PCI-E in particular is a much better technology than the hack that AGP was. I know it doesn't perform better today; neither did AGP8X or even 4X. The underlying bus is the big improvement with PCI-E; it makes motherboards easier and simpler to lay out, as well as offering higher performance.So, Intel was using its position of power as the market leader to break the chicken-and-egg problem of PCI-E and DDR2. Now that there are PCI-E graphics cards available, ATI/Nvidia/VIA are all starting to come out with PCI-E chipsets. Once DDR2 matures a bit more, we are certain to see a move to it as well. On the AMD side, we may not see a move to DDR2 very soon because the A64 is not bandwidth limited like the P4. But one day, it will happen, because DDR2 will be better-performing, and RAM manufacturers will hit the wall with what plain ol' DDR can do.
Really, Intel has done us all a big favor by forcing the move to DDR2 and PCI-E. It's perhaps not in Intel's best interests, but it's definitely in our (the consumers') best interests. BTW I mostly use AMD procs, so I'm not an Intel fanboy by any means -- I run AXP desktops and dual Opteron servers.
bhtooefr - Thursday, November 4, 2004 - link
#22, the processor codename was Shelton. It was a 1GHz Banias without any cache (in other words, the REALLY fscked up ones). The chipset was a bog-standard 845GV. Also, it wasn't QUITE Mini-ITX.