Slashdot Mirror
AMD Athlon 64 6000+ Launched And Tested
Spinnerbait writes "AMD officially launched their next speed bump in the Athlon 64 product line,
in the form of
a new 3GHz part branded the Athlon 64 6000+. This new dual-core Athlon
64 sports 1MB of on-chip cache per core and is designed for AMD's Socket AM2
platform. This chip is still built on AMD's 90nm fab node and is comprised
of some 227 million transistors. It also carries a thermal power profile
of about 125Watts. Unfortunately, in all the
benchmarks seen here, it was still unable to catch Intel's Core 2 Duo E6700
chip at 2.66GHz."
I've always wanted to try running DOS on a processor with 1MB of L2 cache...there's just something retro wicked about running an OS where the entire base memory fits in on-die cache.
I have to wonder if qemu and the kernel's kvm will allow me to dedicate an entire core to a DOS image.
tasks(723) drafts(105) languages(484) examples(29106)
In the full announcement they also mention new 45W single-core desktop processors: Athlon 64 3500+ for $88, and 3800+ for $93.
So they're about 2 years behind Intel on this? Prescott topped out around 135W IIRC, so AMD has 10W more to go...
Burn karma, burn.
-nB
whois gawk date unzip strip find touch finger mount join nice man top fsck grep eject more yes exit umount sleep dump
The problem with socialism is that they always run out of other people's money. - Margaret Thatcher
Nah, with cool n' quiet we'll be able to run it at half of its power and clock speed. Like an Athlon 3000, for example, only more expensive.
"I think it would be a good idea!"
Gandhi, about Internet Security
Better served? Yes, of course. Possible in the short term? No!
Both manufacturers hurry out minor iterations of their existing processor set while readying the next generation; it's a stop-loss tactic, since they can pop something like this out in the engineering equivalent of an afternoon, and it masks the fact that they're falling behind. Rather like the Pentium IV QRSTTurboMach5's that were coming out almost weekly back when Athlon was pantsing Intel. Intel knew they sucked just as much as we did -- but not releasing them would have terminated their share price.
Besides -- your average Dell buyer only sees "New Release", not benchmarks.
"Unfortunately, in all the benchmarks seen here, it was still unable to catch Intel's Core 2 Duo E6700 chip at 2.66GHz."
What's unfortunate about it? It's just a fact.
"Ask not what your country can do for you." --John F. Kennedy
From TFA: The OS used was Windows XP Pro SP2.
A 32 bit OS. The real strength of the AMD 64 architecture is running in 64 bit mode - benchmarking this chip compared to other 64 bit architectures would be far more helpful than running a 32bit Sandra tests and Photoshop tests on it.
Not a very helpful benchmark. I'd like to see these chips compared running 64 bit OS's - and compare the speed and throughput of applications like Apache, Oracle, PostgreSQL, MySQL, PHP / Perl scripting, and raw image processing - not Photoshop, where most of the time is spent waiting on the user to do something.
AMD has been skimping lately on its cache. I have a sneaking suspicion that the majority of AMD's current performance issues are related to cache and lack thereof.
The Intel chips carry 4 to 8 Mb of cache. The thing about the Intel architecture is that the cache is shared across both or all 4 cores. In contrast the AMD chips have a dedicated *tiny* 1 MB cache for the consumer chips and 2mb per core on the high-end parts.
With that said, the reality of dual core computing is that one core is used much more heavily than the other. In Intel's case this means that one core is basically given the entire cache for its use - a significant performance boost when running a few tasks. In AMD's case the idle cache is inaccessible to the heavily loaded core.
The reason that makes me think that the cache is the current bottleneck is that the memory controller on the AMD chip is significantly faster than Intel's. Given that fact one would conclude that in non disk-bound applications that require large amounts of memory (games) the AMD chips would pull ahead. This is not the case. Of course there is more than just cache at play here but the fact that the Intel chips has 4 to 8 times more cache available to it has to make a fairly significant difference.
Check out my AMD FX-70 at http://amd4x4.blogspot.com/
Several things:
1. They have several production lines. They make more than one CPU type at a time. They are capable of simultaneously producing/testing 65nm while making 90nm parts.
2. Idle time in a fab is a KNOWN COST HAZARD. I'm not making this up. It costs money to keep rooms clean, pay the interest on the debt, etc
3. Word on the street [when I was an AMD employee...] was the average processor cost ~60-80$ USD in raw materials/time/effort to make (assuming 100% yield). Yes, your opteron cost about the same to make (excluding yield problems) as that $50 sempron. So why make semprons if they lose money? Yes, I know I'm discounting yield which does contribute real cost to the processors. On the opteron side though, my take [personal op] was that most of the cost was to recoup the R&D not the production costs.
Point is, both AMD and Intel produce low end parts that cost money. Even in the Celeron line which they call "mistakes" (e.g. parts with broken caches) that's not entirely true and is misleading. Even if you made a defective cache, it costs more money to just throw the die out, then to package it as a celeron and sell it at a loss.
4. Intel cores are fast, but they're not the be-all. They still lack NUMA support which is handy in HPC environments (re: not your desktop). They're also not quite a strong in the ALU front (though from my crypto benchmarks are VERY VERY close).
I'm by no means an AMD fanboi. Hell, my desktop is a core2. But I still love my 2-way Opteron workstation and get it to do things that run circles around the core2 (like hosting 15 engineers running simulations/verifications/etc).
Buy what you need, not what some lame commercial on TV tells you. For many, the core2 is the best buy. It's fast, wicked low power and the cost isn't bad. For others, AMD is the better buy (cheaper) or simply more powerful (opterons).
Tom
Someday, I'll have a real sig.
The design and development of a processor has improved vastly since the days of borked multipliers. There are standard benchmark tests that engineers use to rate their designs in-house. If AMD chooses to go with smaller caches, I would imagine they have very good reasons.
Perhaps in order to keep good performance when communicating between caches they need to keep the number of memory addresses low so that the overhead stays low. They decided that separate caches was a better model, and they currently have to maximize performance with this design.
AMD might have favored their server market when choosing this design and separate cache works better for server machines. They may need to refine their architecture for the desktop market. Don't be so quick to accuse AMD of making cache mistakes without doing the math for find the theoretical best solution.
The X2 names are double the clock speed (in MHz) for 1MB cache parts, 200 less than that for 512kB cache parts, and 400 less for the 256kB cache part. It seems they've stopped looking at them as Intel cpu performance equivalence numbers. The single-core chips still seem to be named pretty much arbitrarily.
Complete list:
3000MHz dual-core 1MB = 3000x2 = 6000
2800MHz dual-core 1MB = 2800x2 = 5600
2800MHz dual-core 512kB = 2800x2 - 200 = 5400
2600MHz dual-core 1MB = 2600x2 = 5200
2600MHz dual-core 512kB = 2600x2 - 200 = 5000
2500MHz dual-core 512kB = 2500x2 - 200 = 4800
2400MHz dual-core 1MB = 2400x2 = 4800
2400MHz dual-core 512kB = 2400x2 - 200 = 4600
2300MHz dual-core 512kB = 2300x2 - 200 = 4400
2200MHz dual-core 1MB = 2200x2 = 4400
2200MHz dual-core 512kB = 2200x2 - 200 = 4200
2100MHz dual-core 512kB = 2100x2 - 200 = 4000
2000MHz dual-core 1MB = 2000x2 = 4000
2000MHz dual-core 512kB = 2000x2 - 200 = 3800
2000MHz dual-core 256kB = 2000x2 - 400 = 3600
1900MHz dual-core 512kB = 1900x2 - 200 = 3600