Sitting in a seemingly endless number of meetings at CES, you quickly realize that keeping up with all of the product releases, socket flavors, chipsets to wait for and GPUs to lust after is virtually a full time job.

We try and ease the burden as much as possible by providing you with coverage as soon as we get the information, but what we're seeing at the start of this new year is that with so much changing it's necessary to work a little harder to simplify things. The Intel world is pretty easy to follow; Prescott is due out soon in a Socket-478 flavor, followed by a LGA-775 version. By the end of this year Intel may begin sampling Tejas, the successor to Prescott but very little is known about the chip. Tejas has been rumored to be a multicore desktop chip, however we're now receiving information contrary to what we had once thought.

On the chipset side, we are all waiting for Grantsdale and Alderwood from Intel, the successors to the 865 and 875 platforms. But where Intel's 2004 roadmap is pretty straight forward, AMD's is a bit more cluttered with the confusion of varying cache sizes, Socket-939 and model numbers. Although most of the information about AMD's future plans have already been leaked on the net, we decided to put together a quick reference page to AMD's 2004 CPU roadmap to go along with our recent coverage of chipset plans.

The roadmaps are divided according to CPU socket and the rest is self explanatory - we hope the information is concise and useful:

Socket-754 Roadmap for 2004
  Clock Speed Cache Size Release Date
AMD Athlon 64 3700+
2.4GHz
1MB
Q2 '04
AMD Athlon 64 3400+
2.4GHz
512KB
Q2 '04
AMD Athlon 64 3400+
2.2GHz
1MB
Already Available
AMD Athlon 64 3200+
2.0GHz
1MB
Already Available
AMD Athlon 64 3000+
2.0GHz
512KB
Already Available
AMD Athlon XP 3000+
TBD
256KB
Q4 '04
AMD Athlon XP 2800+
TBD
256KB
Q3 '04

As of now, the fastest Socket-754 CPU will be the Athlon 64 3700+; current motherboard owners looking for an upgrade path can look no further than the 3700+ running at 2.4GHz.

What's interesting is that the model numbers seem to scale more than linearly with clock speed, something we noticed with the Athlon XP that eventually led to its misleading model numbers. In the case of the 3400+, a 10% increase in clock speed over the 3200+ resulted in a 6.25% increase in the model number (resulting in the 3400+). However, if we look at the 3700+, a 9.1% gain in clock speed results in an 8.8% increase in model number. Where we cannot draw a parallel to the Athlon XP model number situation is in the fact that the Athlon 64's on-die memory controller does allow the CPU to scale much better with clock speed. It could very well be that AMD is counting on the Athlon 64's performance scaling much better with clock speed than Prescott, thus justifying the increased model numbers.

What's also worth mentioning is that the 3700+ will be the last 1MB L2 cache Athlon 64 to hit the market, every other Athlon 64 will feature a 512KB L2 cache. The smaller cache size brings us to the difference between the two Athlon 64 3400+ processors listed in the chart above; the current 2.2GHz 3400+, as you know, features a 1MB L2 cache. In the next quarter, AMD will bump the clock speed of the 3400+ to 2.4GHz and cut the cache in half in order to maintain the performance rating.

The last thing to note are the two Socket-754 Athlon XPs on the chart; although clock speeds have yet to be determined, you can expect these two processors to feature an on-die memory controller just like their Athlon 64 brothers but have their 64-bit support disabled. The reduction in cache size to 256KB will decrease 32-bit performance by a noticeable amount, but they will make a good successor to the current Socket-A Athlon XPs.

Socket-939 Roadmap for 2004
  Clock Speed Cache Size Release Date
AMD Athlon 64 FX-55
2.6GHz
1MB
Q4 '04
AMD Athlon 64 FX-53
2.4GHz
1MB
Q2 '04
AMD Athlon 64 4000+
2.6GHz
512KB
Q4 '04
AMD Athlon 64 3700+
2.4GHz
512KB
Q4 '04
AMD Athlon 64 3700+
2.4GHz
512KB
Q2 '04
AMD Athlon 64 3400+
2.2GHz
512KB
Q2 '04

The new Socket-939 platform will bring dual channel support to the entire line of 64-bit AMD processors, and it will also take away half the cache from all but the FX processors. Although the Socket-754 3700+ features a full 1MB L2 cache, the Socket-939 version will not be given more than 512KB. What's interesting is that the two processors carry the same 3700+ model number, running at the same 2.4GHz, while the Socket-754 version features twice the cache. AMD seems to be indicating that the advantage of a 128-bit memory interface will offset any performance loss incurred by halving the cache. We have not seen much data to support this theory, but we'll keep a close watch on it.

If you'll notice, there are two 3700+ processors on the roadmap above and since we're trying to clear up confusion, here's an explanation: the Q4 Socket-939 processors are supposed to be based on AMD's 90nm process (that includes the 4000+ and the FX-55). AMD has demoed 90nm Opterons already, but they were running at 800MHz so there's no indication of how well AMD will be able to stick to this roadmap. If we were to expect any deviation from the roadmap as it stands it would be in Q4; new manufacturing processes are not easy to ramp up as we've seen time and time again from both AMD and Intel (Prescott anyone?).

Socket-940 Roadmap for 2004
  Clock Speed Cache Size Release Date
AMD Athlon 64 FX-53
2.4GHz
1MB
Q1 '04
AMD Athlon 64 FX-51
2.2GHz
1MB
Already Available

Finally we have the quickly demising (at least on the desktop) Socket-940 platform, the FX-53 looks like the end of the line.

We hope this has been helpful; if you'd like to see similar quick reference articles about other companies let us know.

POST A COMMENT

60 Comments

View All Comments

  • cyberphant0m - Monday, February 16, 2004 - link

    I think that both AMD and Intel will have to play their cards carefully, lest the competition swoop up the market and gain ground... If anything, Intel will need to play catch up for a while, seeing as how the Prescott release is quite weak, and it doesn't seem as if they have much left to release for a while. On the other side of the battlefield, AMD had a great release with the Athlon 64 / FX... plus they're readying the next group of CPUs to release...

    Things we get really heated when DDRII ram becomes supported, PCI-X devices are release, and we see major changes to the architecture...
    Reply
  • Icewind - Tuesday, January 20, 2004 - link

    History repeats itself.

    Need I say more? This is the same kinda arguing I saw back in the days of Pentium 4's launch and AMD's rising Athlon star.

    An uninformed rich person building a computer and his money will be soon parted. Do your research before you build or be part of AMD's marketing team or buy a dell and stop whining.

    #58
    Actually, intial tests show the dual channel to be less reliant on lat timings and even more overclockable, but I lost the damn links again.
    Reply
  • Pumpkinierre - Tuesday, January 20, 2004 - link

    My understanding, #57, from an X-bit article is that the A64 memory controller runs at processor speed so it would have to drop cycles to interface with slower memory. However AMD would tune the controller to the specified memory speed, so it will respond better with this particular memory speed. With the a64 memory ie RAM latency is paramount so faster memory is better. In the absence of this, cache can decrease memory access latency for applications with high data/command succession predictability. eg 2D apps. So if the DDR 333 is low latency it might be worth buying a 1Mb cache A64 for the extra ~$60 rather than DDR400 slow latency RAM for ~$100 (512K) if the main use is other than gaming.

    Anantech in their conclusion of the 3400+ Pt.2 article promise an article with varying clock speed probably using the below stock variability of the a64 multiplier. This will of course vary the memory speed so you may be able to glean the information on memory/performance/cache variability you seek from this.

    Aktering the multiplier varies the cpu speed but not the RAM speed. So you go asynchronous and out of step with the AMD memory controller tuning. So certain combinations of clock speed and multiplier will be more stable than others. This is not just true for a64 but also A-XP and even P4 (when using lower speed RAM). However these have a northbridge to smooth the RAM-CPU transition. This is not present for the a64 so multiplier affected speed changes are going to have a greater affect on stability.

    I havent posted much particularly on 256K L2 A64s (much to Reflex's bewilderment) as I am away on holidays using a friend's integrated graphics 845GLV (I think?) and forgotten my password. Dont worry about the switched off 64 bit, they'll turn it back on once WIN64 appears and it will be a great little chip (all you need). Pity about the dual bank controller but the single bank might be faster on the latency front and more overclockable.
    Reply
  • johnsonx - Tuesday, January 20, 2004 - link

    TrogdorJW,

    You make several very interesting points. I hadn't sat and thought about the fact that the various delay cycles may be the same in absolute cycles, but the cycles themselves are correspondingly longer at the lower clock frequencies. Excellent observation. You are correct also that DDR does not help with the cycle delays - those are full cycle delays, even though actual data is transferred twice per cycle.

    Regarding the memory controller, here is one of several places I've read that the DRAM interface is actually clocked using a divisor of the CPU core frequency:

    http://www.hexus.net/content/reviews/review.php?dX...

    This implies in turn that using DDR-266 RAM is no more or less sycnhronous than using DDR-400 RAM. In a sense, it's all asynchronous, but I think the new design really makes the old distinctions we've gotten so used to obsolete.
    Reply
  • TrogdorJW - Monday, January 19, 2004 - link

    JuohnsonX, I believe that you are mistaken on the point of whether or not the memory speeds are synchronous/asynchronous on the Athlon 64. It's the same as any other platform.

    (Incidentally, I looked around a bit for results showing how an Athlon 64 with PC2100/PC2700 compares to a PC3200 setup, but I couldn't find anything useful. Opteron results showed a 3% to 25% increase with PC3200 instead of PC2700, but the Opteron is slightly different. Plus, you're talking about PC2100. And the set of benchmarks was not very useful, since there were no games or office applications in what I found - it was mostly theoretical benchmarks.)

    The reason running PC2100 (266 MHz or 133 MHz DDR) memory is ansynchronous is because the system bus is still running at 200 MHz (double-pumped yields 400 MHz). So if you run 166 or 133 MHz DDR, your memory and system bus are not at the same speed. That's where the big penalties come from.

    Of course, as you've mentioned, the question is how *much* of a penalty you incur by running with asynchronous timings. My guess is that, as always, it will affect real application performance by anywhere from 3% to 33%, with the average probably being around 10%. (Just make sure none of the applications you depend on are in the 20+ percent penalty group!)

    Some other thoughts: The SDRAM cycle time is based off of the *memory* clock speed. So when you read memory timings of 2-3-2-6, for example, that's CAS 2.0, RAS-to-CAS 3.0, RAS Precharge 2.0, and Active to Precharge of 6.0, all at the memory clock. Two examples follow, one with 133 MHz DDR (PC2100) and the other with 200 MHz DDR (PC3200), both with the same timings of 2-3-2-6.

    PC2100: (133 MHz clock)
    CAS 2.0 = 15 ns = 30 CPU clock cycles.
    tRCD 3.0 = 23 ns = 46 CPU clock cycles.
    tRP 2.= 15 ns = 30 CPU clock cycles.
    tRAS 6.0 = 45 ns = 90 CPU clock cycles.

    PC3200: (200 MHz clock)
    CAS 2.0 = 10 ns = 20 CPU clock cycles.
    tRCD 3.0 = 15 ns = 30 CPU clock cycles.
    tRP 2.= 10 ns = 20 CPU clock cycles.
    tRAS 6.0 = 30 ns = 60 CPU clock cycles.

    So, not only is memory bandwidth reduced, but latency is increased. At least, that's my understanding of things. (Anyone who knows better can feel free to correct me. Maybe DDR halves all of those? I don't think so, though. DDR only doubles the bandwidth, I thought?)

    Now, how would more cache affect things? Well, the real question here is how much more efficient the 1 MB cache is in comparison to the 512K cache. I would guess that the 512K cache would have an efficiency of something like 92% (meaning 8% of the time there is a cache miss). While I'm guessing, I would venture that average hit rates with 1MB of cache are probably only 94% or something. So the miss rate is 6%.

    These are only guesses, based on caching analysis that I did way back in the Pentium 133 days. I don't really know if modern caching schemes are better, or if the change in software has made the hit rate decrease relative to eight years ago, or what. Probably a combination of the two.

    How does this all work out, then? To me, it looks like the Athlon 64 3200+ is only around 1 to 5% faster than the 3000+ in the majority of benchmarks, so the increased cache hit rate isn't helping that much. Going with PC2100 RAM is probably going to be a pretty decent hit on system performance, and at best the 3200+ might be 5 to 10% faster than the 3000+ when using both with the slower RAM.

    Overall, I would think it might be okay short term to go with PC2100, but you would want to upgrade that ASAP. It might even be more cost effective to buy new PC3200 RAM and a 3000+ than to use PC2100 and a 3200+.

    Maybe Anandtech will do an article for you, but I douby it. They don't often do non-standard benchmark setups.
    Reply
  • johnsonx - Monday, January 19, 2004 - link

    MoronBasher,

    I appreciate that you took the time to respond, but I think you've missed the point of my post. In short, I'd like to see AT test what effect cache size has when an A64 is paired with slower RAM. In a similar vein, I'd like to see how dependant the A64 is on memory bandwidth in general. The very slight performance difference between the A64 3400+ and FX-51 suggest that the 3.2Gb/s of single DDR-400 is enough to feed the beast (as the 6.4Gb/s available from dual DDR-400 makes little difference in performance); but no one has tested whether 2.7Gb/s or 2.1Gb/s is also enough.

    Also, I think you may be mistaken about running in async mode, and extra performance penalties due to that. From what I've read and understand about the A64 memory controller, it operates from a divider of the core frequency. On a 2.0Ghz chip, it uses a divider of 20 for DDR-200, 15 for DDR-266, 12 for DDR-333, and 10 for DDR-400 (each of which results in an exact frequency match). Therefore running DDR-266 on an A64 obviously results in less bandwidth than DDR-400, but incurs no async latency penalty like we are used to with AthlonXP.

    Interestingly, this also means that a 2.2Ghz A64 can't precisely match some of the DDR frequencies, and therefore runs some modules at less than rated speed. The 11 and 22 dividers yeild exactly 200 and 100 Mhz respectively, for DDR-400 and DDR-200. But DDR-333 has to run at 2200/14=157Mhz, and DDR-266 runs at 2200/17=129Mhz.

    This information makes sense to me, and I've read it several places, but that doesn't make it right.
    Reply
  • MoronBasher - Monday, January 19, 2004 - link

    he could put it in async mode . but that would have severe performace penalties. if someone is getting a pretty high end machine, it wouldn't make sense to skimp on a key component such as ram.

    *watch out for ddr400 though, ddr400 has ver lax latencies compared to ddr333 and 266, if you are not going to run it at ddr400 speeds, latencies will kill you performance wise. best to get corsairs xms pc3200c2 if you are getting ddr400 ram (or ram with timings of 2-3-3-6)
    Reply
  • johnsonx - Monday, January 19, 2004 - link

    The model number system and this roadmap have got me wondering again... what memory must one use to get perfomance equal to the model number? I assume that a 2.0Ghz 1Mb cache Athlon64 only gives "3200+" performance when paired with DDR-400 memory. Likewise I assume the 512K cache Athlon64 at 2.0Ghz gives "3000+" performance paired with DDR-400 as well. In other words, as long as you use DDR-400, then halving the cache of the 2.0Ghz A64 only causes a loss of "200+" worth of perfomance, or about 6%... hardly worth mentioning.

    But, what happens if you use DDR-266 or DDR-333 RAM? Of course the performance of both chips will be lower, but won't the 512k cache chip (3000+) take a much larger hit, losing more than 6% relative to the 1Mb cache chip (3200+)?

    I wonder because my brother-in-law has a AthlonXP 1800+ on a KT-266A board with a gig of DDR-266 RAM. He wants to upgrade his processor and motherboard, but might like to avoid buying new DDR-400 RAM if his existing RAM will work (leaving that for a future upgrade). Thinking aloud, I suggested that the extra $50 spent on a big-cache A64 3200+ would be worth it to lessen the performance impact of DDR-266 RAM. But I've seen no tests on this, so I really don't know.

    Surely other upgraders are in a similar situation; not everyone wants to build a completely new box every time. In a similar vein, with 3 different speed/cache/socket variations being called "3400+", implying roughly equal performance, would they still be equal if paired with slower memory than the presumed DDR-400 standard? Probably not, so which keeps more of it's rated perfomance when paired with slower memory?

    This is something I'd really like to see AT test out...
    Reply
  • TrogdorJW - Friday, January 16, 2004 - link

    So you actually went out and purchased an FX-51 system, and yet you say my $2000 price estimate was too high? That doesn't work very well. Unless by "system" you mean you bought a CPU and motherboard and RAM. In that case, you could probably put together an FX-51 with 512 MB registered RAM for around $1100. But that's with CURRENT prices, and you're talking like you bought the system a couple months ago, so you would need to add another $100 to $200.

    For a well-built system, you would want something like the following:

    1 GB RAM, which would be around $275 or more, depending on timings.
    The FX-51 is $750 or more.
    The motherboard should cost between $200 and $250, depending on brand.
    No point getting a lousy graphics card with an FX-51, so at least $350 for the Radeon 9800 Pro or XT.
    Toss in either a couple of hard drives for RAID 0 or maybe just one larger drive, which would be $125 minimum and up to $500 if you wanted a pair of 250GB drives.
    You would need a good case, so at least $100 there, and potentially $350 if you wanted really high-end.
    Might as well throw in a Lite On DVD+/-RW for $150 and a DVD-ROM for $30.
    Keyboard and mouse for $35 if you go cheap.

    So, rough estimates of prices for a complete Athlon FX-51 *system* come out at $2015 and could easily go up to $3000 or more when you throw in a good monitor and speakers. But clearly a $2000 price tag for a PC using the FX-51 as a basis is WAY off. (Yes, that was sarcasm.) Either that, or you can't do math worth a damn. And since those are Newegg prices and don’t include shipping or costs of putting the system together, $2000 is the bare minimum for what I would consider paying for an FX-51.

    If you paid less than that (and that’s a really big IF, since I still don’t buy your claims of owning an FX system), then go ahead and list your specs and the overall price. Either you’ve got a poorly designed system, you don’t have an FX, or you paid at least $2000.

    --------- End Price Talks ---------

    Now, as to the whole dropping of support for a platform, I never said that was a good tactic. However, it is very common with brand new technology. At the very least, you run serious risks of paying an early-entry premium and getting hardware that will has bugs or unexpected problems. Registered memory required, a forthcoming switch to socket 939, etc.

    If you never heard about socket 939 and went out and bought an Athlon FX, then you are clearly whacked out. I read about the future socket 939 when the Athlon FX/64 were first released, if not before them. It has always been a caveat in every review I've seen of the FX. Like the Anandtech review. I mean, you’re posting here, so you read the review, right? How about this, then: “Socket-939: Athlon 64 FX DOA?” (http://www.anandtech.com/cpu/showdoc.html?i=1884&a...

    Finally, as to the comment that "it's the consumer's fault for saving up and buying a top of the line processor", I never said that. It *is* the customers fault if they run out and buy a brand new technology without doing any research and then find it is not all that they were dreaming of. I feel no remorse in saying that anyone that went out and purchased an FX-51 and expected it to be the future platform of AMD top-end CPUs for the next two years is stupid. Again, socket 939 has been talked about for four months at the least, so if you didn't read about it, that's your fault, not mine and not AMD's.

    AMD should NOT give any customers that purchased an FX-51 motherboard and CPU the option to switch to 939. That would be like me asking Intel to give people a free upgrade to the 2.4/2.6/2.8/3.0/3.2 "C" processors. Or asking AMD to allow people who went out and bought the 2800+ 2.20 GHz "Thoroughbred" CPU and an early Nforce2 to get a free switch to a 2.03 GHz 2800+ "Barton" and an Nforce 2 Ultra 400.

    Hell, in the last two years, how many different chipset have come out from Intel, Nvidia, Via, SiS, etc? Every one of those was an upgrade to the companies' previous "best" chipset. Same could be said of CPUs - and due to bus and voltage requirements, quite a few users could not upgrade both Athlon/Athlon XP and Pentium 4 beyond a certain point.

    If you go out and buy bleeding edge components, you WILL get screwed in some sense. If you have the money, then that doesn't really matter. If you don't have the money, or if you feel that money DOES matter, then you should be buying stuff that's 3 to 6 months old and being more intelligent. Whining about your own mistakes doesn't make them anybody's fault but your own. Especially when web sites like Anandtech.com specifically recommended waiting. From the last page of their review: “We cannot recommend the FX until AMD does release a version with unbuffered memory support and we would strongly suggest waiting until the Socket-939 version is released if you are considering the FX.”

    So, either: A) You’re stupid for not researching. B) You’re stupid for going directly against the recommendation of several reputable web sites. C) You’re full of shit and don’t own an FX (you troll).

    I’m still betting on the last option, since you have given no clear evidence of owning an FX. I mean, if you already have an FX, why the hell should you care that in two or three months there’s going to be something faster? At this point, you’ve already been “leader of the pissing contest” for three or four months with the FX-51, which is as much as you can ask for from any bleeding edge technology.

    Way too much time for this response. Oh, well. It’s written now; might as well post....
    Reply
  • MoronBasher - Friday, January 16, 2004 - link

    Also, the sk8v also works with the opteron. and nvidia's nforce 3 250 will still work with the socket 940, given that a motherboard maker opts to make socket 940s Reply

Log in

Don't have an account? Sign up now