64Bit UT2K3 (saw this over at nvnews)

[Natoma]

http://biz.yahoo.com/ap/021119/comdex_amd_2.html

This should make a great benchmark for comparing code that executes in 32bit and code that executes in 64bit on the Hammer.

Can't wait!!

Any performance thoughts?

 

[Humus]

The 16 registers vs. 8 registers really should boost performance. Or really, it's more like 14 vs. 6 registers as we generally don't want to touch the stack pointer and frame pointer. From this alone I expect (to pull a number out of my ass) maybe 40% speed increase in normal applications. The flat floating point model (non-stack based) should boost performance of floating point operation quite a bit. Probably some 40-50% there too.

 

[Natoma]

The 16 registers vs. 8 registers really should boost performance. Or really, it's more like 14 vs. 6 registers as we generally don't want to touch the stack pointer and frame pointer. From this alone I expect (to pull a number out of my ass) maybe 40% speed increase in normal applications. The flat floating point model (non-stack based) should boost performance of floating point operation quite a bit. Probably some 40-50% there too.

Are you going to create any demos to test the performance differences between 32bit and 64bit on the Hammer?

While I no doubt believe the UT2K3 guys did the best job they could, I'm inclined to think that the game, while running in 64bit mode, isn't necessarily built from the ground up for 64bit performance, i.e., its tweaked for 32bit performance.

So the differential might actually be less than you're anticipating.

 

[borzwazie]

Ok, I'll pull some stuff out of my butt too, since I'm not the first

I remember reading quite a while back that the Hammer, unlike the Itanium, would really improve the speed of 32-bit apps, something like (forgive me if I'm really off here) there were 2 32-bit paths or something like that, so 32-bit users could expect a real improvement. I'm sorry I can't remember where I read it, but I remember that it was one of the things that was to make Hammer competitive with Itanium during the general move to 64-bit platforms, since the Itanium doesn't do 32bit programs very well.

hope I didn't reach too far pulling this stuff out

 

[Humus]

Anandtech on Hammer performance:

http://www.anandtech.com/showdoc.html?i=1752&p=3

The performance of the Athlon 64 at 1.4GHz was around the speed of a 2.2GHz Pentium 4 or Athlon XP 2200+

That would mean about 30% higher performance than a AthlonXP on a clock for clock basis. That's in 32bit applications, so it should really perform well in 64bit applications.

But the most exciting really is this:

The heatsink was lukewarm at best while the system was running 3DMark 2001 SE, clearly a testament to how beneficial Silicon on Insulator is

 

[Johnny Rotten]

I wouldnt expect a significant performance difference either way. In fact, John Carmack said a 64 bit version of Doom wouldnt have any tangible benefit over a 32 bit version.

 

[demalion]

I wouldnt expect a significant performance difference either way. In fact, John Carmack said a 64 bit version of Doom wouldnt have any tangible benefit over a 32 bit version.

Isn't Doom 3 going to be graphics card limited?

 

[Nagorak]

I wouldnt expect a significant performance difference either way. In fact, John Carmack said a 64 bit version of Doom wouldnt have any tangible benefit over a 32 bit version.

If there's no performance difference then why are they moving to 64 bit processors at all? For that matter why didn't we just stick with 8 bit DOS. Just because Carmack says something doesn't mean it's written in stone. Although that very well could be the case with Doom, that's probably just due to his programming path.

On the other hand that does bring up an important point. In the short term, I really wonder if 64 bit processors are really going to be that beneficial. It's going to take a long time to move away from the current 32 bit installed processor base. So it will be interesting to see how much attention they actually receive.

 

[antlers]

64-bit comes in handy once a significant fraction of your calculations need more than 32 bits. Obviously, servers running huge databases will benefit from 64 bit immediately. Doubles are generally 64-bit, so you might see some speed up if you are doing floating point--the whole value is in one word for the CPU.

Games are starting to get to the point where they might need to address more than 32-bits worth of memory (well, not yet, but I can see it happening in a few years).

If you are using FP16 for your graphics calculations, those fit nicely in 64-bits. I'm not sure when CPUs will need to access data in this format, though.

 

[KimB]

But the move to 64-bit isn't the only benefit here.

AMD's x86-64 also includes far more registers, meaning that you can do more complex calculations without having to move data in and out of the registers as much.

In other words, there are benefits of the x86-64 architecture other than just the move to a 64-bit instruction set.