I've been very quiet about AMD's Socket-AM2, and here's why:

Back in January we got our hands on a shiny new Socket-AM2 motherboard and a Socket-AM2 Athlon 64 4800+. I didn't pay too much attention to the fact that the 4800+ AM2 chip we had was a 2.4GHz/1MB part, just like the Socket-939 variant, I was too busy being happy about having an AM2 platform.

Needless to say my excitement vanished after I ran the first performance numbers and it offered about half the memory bandwidth of an average Socket-939 platform. Remember that the major change with AM2 is the migration to DDR2-667 (and DDR2-800) over DDR-400, so memory bandwidth should go up - not be cut in half.

I chalked it up to being an early board with an early CPU, but honestly I had no idea whether it was the motherboard or CPU that was at fault here. Usually when I get my hands on a chip from Intel < 6 months before its release, its performance is pretty close to final. I had Prescott about that amount of time before its release and its performance didn't change. With AMD CPUs it's a little more difficult since they're equipped with an on-die memory controller; so what used to be easy to trace back to the motherboard could now be either the board or the CPU at fault.

I checked with AMD's partners and they were seeing the same poor results that I was, and I also checked with AMD, to see if they had seen anything different. Of course AMD's response was that they were seeing better performance than I was. I figured that AMD wasn't going to launch AM2 6 months later and cut performance in half, so I held off on publishing any numbers. AMD isn't the only company to receive this sort of treatment - I actually did the same thing with Prescott when I first tested it. I figured I had an early version of the CPU since performance was actually lower than an identically clocked Northwood; of course the outcome of that situation was much different as Intel actually did reduce performance with Prescott :)

I eventually got more boards and BIOS updates that improved performance with AM2, but it was still lower than any of the current Socket-939 systems and I felt that it wouldn't really help anyone to put out numbers that weren't representative of what we'll be seeing in June. Today performance is a lot better than when I first looked at the AM2 platform, however it is still slower than a similarly configured Socket-939 platform. We're now around 3 months away from AM2's official launch and I'm beginning to worry. But it's what I saw in my most recent tests that may shed some light on what the AM2 story will be.

AMD's Socket-AM2 Roadmap got leaked a few weeks ago, but the roadmap only illustrated model numbers, there were no actual clock speeds reported. According to the most current roadmaps, the AM2 platform will launch alongside the Athlon 64 X2 5000+ processor, which is currently a 2.6GHz CPU with a 512KB L2 cache per core. The current Socket-939 4800+ is a 2.4GHz CPU with a 1MB L2 cache per core, and the 4600+ below it is a 2.4GHz CPU with a 512KB L2 cache per core. Going one more step down you've got the 4400+ which is 2.2GHz/1MB and then the 4200+, a 2.2GHz/512KB chip. See a pattern? With the X2s, AMD alternates between increasing clock speed by 200MHz and increasing L2 cache size. If a 5000+ were to debut on the Socket-939 platform it would be a 2.6GHz chip with a 512KB L2 cache. A 5200+ would be a 2.6GHz chip with a 1MB L2 cache. But the 5000+ isn't a 939 chip, it's an AM2 chip, meaning that its model number implies the move to AM2 doesn't actually offer any performance benefit. AMD has been fairly silent on what to expect from AM2, but could it be that they aren't expecting to see any tangible performance benefit at all? If the move to DDR2-800 was going to result in some sort of a performance gain I'd expect the 5000+ to be a 2.4GHz/1MB chip just like the current 4800+, or for a 2.6GHz/512KB part to be a higher rated part.

Now clearly I'm still talking about an unannounced platform and an unannounced CPU, but assuming the specs on the 5000+ don't change it may very well be that AM2 does nothing for AMD.

Obviously the one area I still would expect AM2 to offer some tangible performance increase would be in heavy multitasking environments. Basically any areas where we saw an increase with DDR-500 I'd expect something similar with DDR2.

AMD will be allowing their partners to show off AM2 boards at CeBIT and where there's a platform, there are usually benchmarks. I'm not sure where AM2 will go between now and its launch, but hopefully this will set your expectations at least in the short term a bit better.

Off to IDF, more info from the show this week.
POST A COMMENT

19 Comments

View All Comments

  • kleinwl - Thursday, March 30, 2006 - link

    According to the EE times, AMD is coming out with Hyper Link 3.0 around June. Hyper Link 3.0 is supposed to have 2x the bandwidth of Hyper Link 2.0. This could be responsible for some gains that AMD expect out of AM2, without improvements in the actual processor (as Anandtech has reported). Reply
  • Questar - Friday, March 31, 2006 - link

    I assume you mean Hypertransport?

    I'm going to also assume that you know that HT has nothing to do with memory?
    Reply
  • PaulDriver - Sunday, April 9, 2006 - link

    bzzzzt.

    HyperTransport is not used for memory access in a single socket configuration, in multiple socket configurations, memory requests can travel across HyperTrasport links.
    Reply
  • microAmp - Monday, March 27, 2006 - link

    Any chance for some benchmarks on TES:O? Reply
  • creathir - Monday, March 27, 2006 - link

    Anand is once again MIA... pushing 3 weeks now...
    Is this just...
    "noanandtech.com"???

    Hope all is well...

    - Creathir
    Reply
  • Nimbo - Thursday, April 6, 2006 - link

    Officially one month without trace of Anand in the site that carries his name. I feel like an orphan. Reply
  • Houdani - Wednesday, April 12, 2006 - link

    Not true. Anand's been posting articles, just not sharing about the quirks of his new house which should probably be done by now. Reply
  • creathir - Tuesday, April 4, 2006 - link

    2 days and it will be a full month since his last contact with us...
    Anand, I know you are busy and all, but you are the driving force of this site. Without you, it will die...
    Where in the world is... Carm- Anand Lal Shimpi?

    - Creathir
    Reply
  • lopri - Tuesday, March 7, 2006 - link

    Whose fault is it? CPU itself? Integrated Memory Controller? Or Motherboard implementation? It's somewhat relieving - I just invested heavily on S939 platform. ;) - but at the same time worrying, of course.

    Will we see many enthusiasts moving to conroe/merom platforms at the end of the year?
    Reply
  • Calin - Thursday, March 23, 2006 - link

    Integrated memory controller, and maybe the fact that typical DDR2 667 memory has a higher bandwidth but a longer delay (latency) time until data can be fetched.
    Tests (by Anandtech and others) have shown that Athlon64 is fairing better with lower latencies and less bandwitdh than the reverse. The performance "gap" between Socket 754 and Socket 939 (where you can see a doubling of bandwidth for a 10% some improvement) is proof enough.
    DDR2 could prove usefull in multitasking/multicore scenarios
    Reply

Log in

Don't have an account? Sign up now