First Round of AMD Athlon 64 Reviews In

wrinkledshirt writes "Here's a bunch of AMD Athlon 64 reviews, courtesy of 8Dimensional." AcesHardware and HardOCP match the Athlon 64 line against the Pentium 4 Extreme Edition. amdmb, FiringSquad, and SharkyExtreme take a closer look at the FX-51. AthlonXP and PCStats have glowing reviews of the chips. Digit-Life compares the new Athlon 64 with Opteron and a Pentium 4. LegitReviews and Overclockers.com.au also both have succinct reviews of the FX-51. Overall the reviews speak very highly of the Athlon 64 and the FX version of the chip, with the only downside being the cost, especially of the FX chip.

Comments

Hmm..

[darkov]

Is it possible to have a first round of reviews after an earlier story about Athlon 64 reviews?

fp

Hang on..

[madprof]

Give it a bit more time. The motherboards can still be tweaked to get more performance out of the system. I think we should wait a bit before really making decisions, and get in a whole load more real world testing too.
Benchmarks are not always entirely, although often can be, illuminating.

Re:Hang on..

[FatherOfONe]

We have a dual Opteron 1.8GH and a 2GH on order. We found that with RedHat and Oracle the Opteron in 32 bit mode beat the crap out of a quad Xeon for the stuff we do.

Just an F.Y.I. Again this was with 32 bit code. I tried the RedHat BETA and it wouldn't even boot up without locking.

So given that Oracle cost us over $20k a processor, we saved over 40 grand!

Re:Hang on..

[WoTG]

I don't have any firsthand experience with this stuff, but I have read a lot of the reviews and whatnot. I gather that the difference is in scaling to 2 way (and 4 way) boxes. While a single 3GHz Xeon and a single Opteron are pretty competitive, the Opterons are built for dual and quad processor work - especially with memory intensive applications.

Each Opteron has it's own memory channels in multiprocessor boxes. All memory is still shared throughout the system, it's just that there is more total memory bandwidth to go around as you add Opterons. In comparison, Xeon systems have the same amount of system wide memory bandwidth from 1 CPU all the way to 4 CPU's. The net result is that in many cases a second Opteron processor nets a gain 80% or more performance - which is a LOT better scaling than Xeons. This will probably be even more evident in future comparative reviews of quad CPU boxes since the Xeons will be sputtering on memory bandwidth fumes (relative to the Opterons).

don't bother with the FX yet

[Tumbleweed]

The socket format will be changing soon, and once the upcoming changes happen, things will be much better. You'll then be able to use non-ECC memory, and the motherboards will be less expensive.

Until then, yeah, the FX is freaking fast, but waaaaay overpriced, so don't bother.

Re:Memory mapped disk?

[Anonymous Coward]

If you have lot of RAM or so programs will tend to exhaust the 32-bit address space before actually consuming all the physical memory. One example is that creating a thread might represent the reservation of 64KB of address space for its stack, while only 8KB of physical memory is actually committed.

A 64-bit address space is probably a good thing once a program is allocating 2GB or more of address space.

Betas Of Athlon64 Optimized Linux

[Anonymous Coward]

Are avalible from SuSE, Gentoo, and Debian!

SO if you are complaining "theres no 64 bit os yet", stop complaining, leave the evil empire behind and see the REAL power of opensource.

Re:Pentium 4 Emergency Edition

[MarcoAtWork]

while it's very likely that this is the case for the reviewing samples (probably rushed out as fast as possible, I can't believe that Intel had plans for the EE all along given that there were 0 leaks) given that this is a paper launch by the time the EE will actually be available to customers I'm sure the MP support will disappear.

Re:Quote from the article

[The One KEA]

Not quite.

The AMD64 core uses a 40-bit physical memory address space, which is 1 Terabyte. It also uses a 48-bit virtual memory address space, which is 256 Terabytes.

A full 64-bit physical memory address would allow for 16 Exabytes of memory.

Re:Pentium 4 Emergency Edition

[Unoriginal Nick]

It was either the Inquirer or The Register that had an interesting article saying that these CPUs (which are MP Xeons) still have their multi-cpu support enabled, thus saving astute customers thousands of dollars over their full-priced, $3,900 Xeon counterparts.

The multi-cpu support may or may not be still enabled, but the P4 EE has a different pin count than the XeonMP, so you wouldn't be able to use it in the XeonMP boards anyway.

more motherboard reviews, please (esp. w/ Linux)

[_|()|\|]

I haven't made all the rounds, but it seems like everyone is using the same two motherboards: Asus SK8N and MSI 8KT. I really like the looks of Monarch's Hornet 64, with a uATX Gigabyte GA-K8VT800M. I'd like to see some reviews first, especially regarding chipset support under Linux. I'd also like to hear more about video drivers. I've heard that NVIDIA's drivers need some work. (Does ATI even have any?)

We've got a couple of Opterons at work, one for 32-bit compatibility testing, and another for the AMD64 port. It's pretty cool to see this in Python on SuSE Linux 8.2 beta:

>>> type( 9223372036854775807 )
<type 'int'>

SuSE Linux 9.0 for AMD64 is supposed to ship next month. Hopefully, it will be a little cheaper than RHEL 3.0 for AMD64, which will be more than twice the price of RHEL 2.1 for x86!

Proper benchmarks

[NitroPye]

Untill I see proper benchmarks done with applications built for 64bit chips I am going to go pffffft.

I am still blown away that the FX did better then then 3.2 P4.

Proper benchmarks include not using a 64bit beta stupid o/s like windows, a properly optimized linux (suse 64 or gentoo) and applications built for the chip. Openoffice, kde and kde apps, mozilla, some miscellaneous 3d engines running some impresive demos, maybe tenebrae quake. Tenebrae quake is great being that its open source and takes a huge amount of gfx and proc power.

Most Only 32 Bits...

[MBCook]

The reviews are basically all 32 bit, which is a shame. Linux is out there. I've only seen one review that did anything really 64 bit (running win32 programs on win64 doesn't count). The only 64 bit test they did under Linux was MP3 encoding. The test was the Athlon 64 running a 32 bit version they compiled of the MP3 encoder vs running a 64 bit version of the same program. The "bitness" was the ONLY thing that was changed.

The results? The 64 bit version took nearly HALF THE TIME of the 32 bit version. This is the kind of thing we have to look forward to in some things (MP3s, video encoding, etc).

The Athlon 64 is fast in 32 bit mode, and can beat a P4 many times. But when the 64 bit code comes along, the P4 will be taking one hell of a beating.

PS: Sorry I don't remember which review had this test. I don't have time to go hunting for it right now.

Re:Most Only 32 Bits...

[Sangui5]

The test was the Athlon 64 running a 32 bit version they compiled of the MP3 encoder vs running a 64 bit version of the same program. The "bitness" was the ONLY thing that was changed.

Not really. While the "bitness" changes, what also changes is the number of registers visible to the compiler. The x86 ISA has been dealing with internal register rename as a nasty hack to deal with a sever shortage of programmer visible registers for a long time. This goes to show that the compiler is much smarter about register allocation than a hardware renamer can ever be. I'm interested in seeing the performance of common multimedia applications once hand-written core loops are available.

And a note to those who are pointing to improved SSE2 support as the reason for the performance gain: they are comparing an AMD64 in 32 bit mode vs one in 64 bit mode. Unless GCC is being bass-ackwards, the SSE2 support should be benefiting the 32 bit mode as well. It appears that the only variables in this benchmark are the 64 bit math and the additional registers.

Re:Most Only 32 Bits... FOUND IT

[MBCook]

Found it! It was Anandtech. Check out the bottom of the 32 bit vs 64 bit page of the review.

Re:"organic" ICs

[cosmo7]

Organic integrated circuits are produced without using fertilizer or pesticide.

Re:Memory mapped disk?

[ianezz]

You will not see consumer systems with more then 4GB any time soon

The point about mmap(2) is to let the system (the VM subsystem of the kernel) manage the caching for the userland processes using it, avoiding extra copies to/from buffers in userland and eliminating in several cases the need for custom caching code (processes don't have to worry about data being available in RAM: the kernel automatically takes care of that when needed).

You don't need gobs of memory to do this, but in order to work on large amount of data you need a large address space, which is what 64 bit architectures provide. Of course, the more physical memory you have, the less the kernel has to swap pages in and out, but the main point is not that.

A little example to clarify: in order to keep things simple (instead of needing two intermixed caching systems, one for the VM and one for disk accesses), the Hurd just mmaps the whole partition. This means that the maximum size of a partition has an upper limit given by the size of the addressing space, which is 4GB on 32bit architectures (actually less, since in that address space you have to keep also the code that uses the mmapped data, so it's more like 2GB/3GB). A 64bit architecture comes very handy here, given the current size of hard disks.