.09 Athlon 64: Value, Speed and Overclocking
by Wesley Fink on October 14, 2004 12:05 AM EST- Posted in
- CPUs
Final Words
We were impressed with the performance and overclocking that we found with the 90nm 3500+. However, it is very difficult to draw any conclusions based on a single sample, so we also bought a 90nm 3000+ from another vendor. These two processors, at different speeds and from different sources, performed similarly enough to allow us to draw some broad conclusions about the performance of the new 90nm Athlon 64 processors. The new 3500+, 3200+, and 3000+ perform from 1% to 7% faster than comparable 130nm parts. We still don't know if this is the result of the die-shrink or the new DH8-D0 revision. We checked all recent Athlon 64 in the lab and we could not find a D0 A64 for comparison. We have asked AMD to shed some light on what we found in our testing, and we will report what AMD says about the performance improvements as soon as we receive the information from AMD.Based on the Performance tests alone, there is reason to be pleased with what we know of the AMD die-shrink so far. If you also consider the fact that AMD appears to have accomplished the shrink to 90nm without an increase in heat, the process move should be considered a big success. We won't know this for sure until we see the fastest AMD chips in 90nm clothes - that is where the thermal impact of the shrink will be most visible.
The other side of the equation is headroom or overclocking. AMD enthusiasts have always seemed to flock to value chips in the AMD family. More than Intel users, AMD enthusiasts seem to always want something for nothing. It's not because they are cheap necessarily, but because AMD won them over by providing outstanding value points in their product line. In fact, many AMD users have hung on to Socket A technology long after it was significantly outperformed in the market because they could buy the Athlon XP cheap and overclock the heck out of it. Those users will love the new 90nm chips in general - and the 90nm 3000+ in particular.
They will love the new 90nm chips because they can buy a 3000+ running at 1.8GHz for less than $200 and still have a good chance of reaching 2.6GHz with very little effort with the same chip. 2.6GHz is faster than any current Athlon 64, and it is, in fact, the speed that we expect from the upcoming FX55 - the new Athlon 64 top-of-the-line. It's been a while since we've seen this kind of headroom on an AMD chip, and those who were waiting for 90nm to get a magic overclocker will get in line to buy the new 90nm 3000+.
The major impact of the new 90nm Athlon 64 chips may not be quite as obvious. Prior to the new 3000+, 3200+, and 3500+ 90nm chips, the entry point to dual-channel 939 was the $400 3500+. As a result, buyers saw the Socket 754 as the value solution for Athlon 64 shoppers, where they could buy a 754 Sempron 3100+ for about $120 or a full 64-bit 2800+ for around $140. The new 3000+ should cost about the same as the 130nm 3000+ once the new settles into the market. That will make the cost of entry for the top 939 chipset well below $200 for the CPU. Many buyers who would have bought 939 if it had been cheaper will now be able to buy 939.
This leads us to future directions for 754 and 939. Roadmaps show 754 ending in late 2005, but 5 quarters is still a lifetime in CPU sockets. AMD plans to discontinue Socket A and move all processors to Socket 754/939/940. This will likely mean that we will see even cheaper 754 processors to entice buyers who found Athlon XP prices attractive. 754 will likely move much lower before it goes away in a year or so, and 939 will also likely move down a bit further as 90nm is fully implemented and production costs go down. All-in-all, it's becoming a very good time to be in the market for an Athlon 64 processor.
Facebook
Twitter
RSS
89 Comments
View All Comments
Bugler - Thursday, October 14, 2004 - link
Newegg Model#: OCZ4001024ELDCPER2-KItem#: N82E16820146890
OCZ EL Platinum Revision 2 Dual Channel Kit 184-Pin 1GB(512MBx2) DDR PC-3200 - Retail $281
Araemo - Thursday, October 14, 2004 - link
nevermind...It is only for sale in 1 gig packs of 2x512 right now, different part #:
http://www.newegg.com/app/viewproductdesc.asp?DEPA...
Araemo - Thursday, October 14, 2004 - link
Also.. is that ram available in the retail channel? I wanted to look up the price, and found the part number(I believe) OCZ400512ELPER2However, this isn't on newegg, or pricewatch.
Araemo - Thursday, October 14, 2004 - link
#23 - WesleyIf you already have overclock results of a p4 from another article, how difficult would it be to include in the graphs? Or were those results using a different enough configuration that it is not an applicable comparison?(In which case, as a reader that loves Anandtech for your thoroughness, I would like to see an applicable comparison.)
All in all, good review. Not as overly wordy as some have been recently(Though I won't name names. ;P).
Bugler - Thursday, October 14, 2004 - link
With the 3500+ showing a 20% overclock and the 3000+ hitting a 45% overclock, it would be great to know how the 3200+ would overclock in this comparison.Wesley, thank you so much. Once again, another fine job.
PrinceGaz - Thursday, October 14, 2004 - link
Further to my earlier comment, the default core voltage of all the Winchester-core 90nm A64 parts currently available is 1.4V, not 1.5V as indicated in the review. Its important this is corrected on the Overclocking page of the review as it is very relevant to the obtained results.I now see that you didn't actually measure the temperature under full-load conditions. Other reports suggest that the 90nm parts do run cooler when idle than the equivalent 130nm parts, but are hotter under full-load conditions due to the higher thermal density. They have been measured as using less power under full-load than the 130nm parts, but run hotter because that power is concentrated in a smaller core.
I'd be very interested to know just how hot that 3000+ got under full-load conditions (eg. running Prime95) when you were feeding it 1.6V instead of 1.4V, and had it clocked at the maximum of 2610MHz. If you were using the standard retail HSF, it may have been rather hot :)
----
As for why the 90nm parts run a little faster than the 130nm parts, I found this post on the AMD forum. I don't know if the info is accurate, but it sounds reasonable:
Whether the 90nm process for the 3000+ to 3500+ runs cooler is still up for speculation to a degree. What will eventually be shown is that the TDP for these processors is lower than the current 130nm. (currently it is 89W TDP, the TDP for these three - when the information is released - is 67W).
In addition the 90nm A64 (DH8-D0) has these improvements over the 130nm (DH7-CG):
- improved DRAM page closing policy
- improved memory addressing with graphics cards using main memory (eg. integrated cards) as frame buffer
- memory controller power reductions (DDR receivers go off in default)
- memory power consumption reductions (CKE pins disconnect)
- second write combining buffer
- SAHF and LAHF instructions are now supported in 64bit mode
Wesley Fink - Thursday, October 14, 2004 - link
#22 - I appreciate your suggestion, and we did overclock the Pentium 4 775 in our "Intel 925X Roundup: Creative Engineering 101" at http://www.anandtech.com/mb/showdoc.aspx?i=2162.The highest stable overclock we could achieve with the P4 on air cooling was 3.92GHz (280x14) on the best overclocking 925X board. Others have achieved higher overclocks with water and phase-change cooling, and higher overclocks will also likely be achieved with those methods on the new 90nm Athlon 64 processors.
We will be looking at Pentium 4 overclocking again in the upcoming launch of some new and improved P4 processors.
thermalpaste - Thursday, October 14, 2004 - link
I am an AMD freak, and Im happy they launched the winchester. You should have, however overclocked the Pentium-4 also, just to compare the scalability of both the CPUs.I had read an article on somebody overclocking the pentium-4 to 6 Ghz. Though this was an unstable overclock, what this indirectly implies is that despite of have a 30-odd stage pipeline, intel may find it difficult to reach speeds in excess of 5Ghz using the 0.09u process...I expect a more thorough comparo soon.....cheers!
deathwalker - Thursday, October 14, 2004 - link
All the buzz in this article is about the O/C'ing capabilities of the new .90 die...personally im just as impressed or maybe even more so with the performance of the memory used in this testing. Having made that statement it is clear that the O/C'ing capability of the 3000+ version of this Proc. takes us back to the good old days of the Celery 300.Wesley Fink - Thursday, October 14, 2004 - link
#11 & #16 - The memory brand is identified in the "Performace Test Configuration" on p.4 and the timings are in Overclocking table on p.5.The OCZ PC3200 Platinum Rev. 2 and other top performing memory is tested on the Athlon 64 in "Athlon 64 Memory: Rewriting the Rules" at http://www.anandtech.com/memory/showdoc.aspx?i=222... Some memory in that review made it to DDR618 on A64, but DDR580 at 1T was the fastest 1T performance.