AMD's Athlon 64 3400+: Death of the FX-51

38 Comments

View All Comments

• AtaStrumf - Friday, January 9, 2004 - link

  This is just a quick test not an in-depth review, since I'm sure Anand is rather short on time, but was swept up in all this A64 hysteria and just had to do the test.
  
  Is's an OK test though, and I really liked the BFBI (bang for buck index). I also like the conclusion, which is one of the more clear ones in recent history. bussines/games/compile --> AMD64, encoding, 3D rendering programs --> Intel. Then just pick your BFBI and you have your CPU.
  
  Gotta try to keep things simple or else all you get is confusion.
  
  Maybe reviews should be written like some books, which keep simple information in normal size font, and more detailed in fine print, since it is intended for the more informed and interested readers and would only confuse the rest.

• Evan Lieb - Friday, January 9, 2004 - link

  TrogdorJW,
  
  Intel's 875P motherboard is at worst 2% slower than ASUS, ABIT, Gigabyte, etc. 875P motherboards. Anand may have used Intel's 875P board merely to guarantee as best he could that no problems would arise during testing, as Intel motherboards are usually the most reliable on the market.
  
  Take care,
  
  Evan

• EZDEALZ - Thursday, January 8, 2004 - link

  Athlon 64 3200+ processor locked? I have my K8V running at 2.2gHz (220 x 10) with no problems..

• TrogdorJW - Wednesday, January 7, 2004 - link

  Still no comment on the use of the Intel motherboard? Oh, well. How about the actual memory timings that the tests ran at? I know the RAM says it's rated for 2-3-2-6 timings, but just because RAM is rated at a certain speed doesn't mean it was actually run at that speed. (Motherboard and chipset issues can often cause problems.)
  
  As for the suggestion of benchmarking with different memory speeds, johnsonx, I have to think that it would prove to be largely useless. Tests on numerous hardware sites have shown repeatedly that running the memory and system bus asynchronously seriously impairs performance. If you're running a 200 MHz FSB and then OC the memory to 233, you'll probably end up running slower in many applications. Synchronous clocking is the way to go right now, IMO.

• johnsonx - Wednesday, January 7, 2004 - link

  gee, screwed up again... in my corrected line I left out "controller", as in "memory controller is relatively inefficient"
  
  But again, I'm guessing you all knew what I meant. I just hate leaving out words...
  

• johnsonx - Wednesday, January 7, 2004 - link

  nuts... I didn't finish an edit before I posted. That line in the 6th paragraph should read "this suggests the memory is relatively inefficient in dual-channel mode".
  
  But I'm guessing you knew what I meant anyway...
  

• johnsonx - Wednesday, January 7, 2004 - link

  Seems to me that the extremely close scores between the single and dual channel 2.2Ghz A64 chips suggests:
  
  1. Any bandwidth benefit of dual-channel memory with the A64FX is offset by the latency penalty of registered memory. With the memory controller on-die, latency is apparently everything.
  
  2. The single-channel version of the A64 memory controller is extremely efficient.
  
  3. The dual-channel version of the A64FX memory controller is perhaps somewhat inefficient, or perhaps the cpu-core to memory controller interconnect is not as fast as it should be. Latency penalty aside, all that extra bandwidth should amount to a bigger performance gain.
  
  Our intrepid friends at Anandtech could shed some light on this by testing the 2.2Ghz chips with different memory speeds. I believe both chips will support memory speeds down to 133Mhz/PC-2100.
  
  This might provide interesting results. If the numbers increase fairly linearly with memory speed, but are relatively close between single and dual channel, this suggests the memory controller is fairly relatively in dual-channel mode. If the numbers for single channel scale much more with memory speed than with dual-channel (i.e. if single DDR-266 is much slower than single DDR-400, but dual DDR-266 (with roughly DDR-533 equivalent bandwidth) is not much slower than dual DDR-400), then that suggests that the A64 is architecurally limited somehow, and at current speeds is well fed by 3200Mb/s bandwidth, (or whatever the tests show), dual-channel or not.
  
  A related interesting test would be to repeat the same set of tests using an A64 and A64FX running at the same 2.2Ghz, but using a 220Mhz clock speed and a 10x multiplier. For this, obviously an A64 3200+ would be overclocked... This would tell us whether the memory controller's performance (and the performance of the on-die link between the core and memory controller) is dependant on the base clock speed.
  
  just my lofty ideas I can't test myself since I don't have any A64 at all....
  

• Icewind - Wednesday, January 7, 2004 - link

  Fudge.....I really wanted Dual Channel with my Athlon FX next summer, but god I just can't swing a $750 cost, I wanna get a R420 also. Damnitm why cant AMD give the A64 dual and just let the FX have the unlocked multiplier for ethusisasts?
  
  For some reason, Tejas is starting to sound pretty good now

• Wesley Fink - Wednesday, January 7, 2004 - link

  In my benchmarking of the FX versus the 3400+, the FX is the clear winner in Workstation Performance, Media Encoding, and some of the more demanding games. The point is not whether the FX is faster, because it is in some applications, it is whether it is now worth the 40% higher price asked for it.
  
  The P4 wants all the bandwidth you can give it, and there is a big difference in perfromance between Dual-Channel and Single-Channel memory. The A64 architecture is not as dependent on Memory Bandwidth for performance. Having said that, Dual-Channel on the A64 is faster in some applications and it may be even more important to 64-bit performance.
  
  The upcoming Socket 939 IS dual-Channel, so we will be seeing more Dual-Channel solutions from AMD.

• DAVIDS - Wednesday, January 7, 2004 - link

  Insomniac: yes, what you say is true. But is no consistency to the way Anandtech tests CPU's. Look at the benchmarks used for the 3000+ and the original Athlon64 review from 9/23. Also, the FX does not have a multiplier lock, a fact this article seems to downplay. Anyone who overclocks understands the importance of an unlocked CPU.