Athelon64 & Opteron vs. P4

[wkulecz]

I'd jump to 64 bits when I see the performance gains to justify the price. These early media encoding tests I found show sub P4 performance so far! (for what I care about most).

Sure 4-8GB RAM systems will be good for Micron and therefore Boise, but PC3200 RAM seems to be running about $250-300/GB. Gonna take some serious performace gains to justify the price along with ~$700 CPUs.

Only my largest photoshop projects showed any worthwhile gains in going from 768MB to 1.5GB RAM (which was why I did it). Noticed no improvement in MSP6.52a.

--wally.

[Flying dutchman]

@Ray, Jerrold, Marshmallowman:

I quote a few lines from this week's CT magazine (issue 20/2003), who tested and benchmarked a few 64-bit systems, including Athlon-64 and Powermac G5:


(p. 106)
... often the CPU core has to handle much smaller data types (char, short etc) which means clear disadvantages with respect to performance and power consumption. When calling subroutines and handling task switches, it has to push unnecessary large amounts of data on the stack, arithmetic operations (multiplication, division) take longer than needed (five instead of three for multiplication on the athlon 64) and far too often the data is written into the caches and the main memory in much too large packages.

...Even the 32-bit world already wastes a lot of memory, over 50% of the resources are used for the integer "sign" alone, because in the majority of cases the upper 15 or 16 bits are all zero or, in the case of negative numbers, 1. The transition to 64 bits drastically increases this memory waste to over 75 percent.

( p. 110)

...FPU, MMX and 3dNOW are not available under all 64-bit operating systems. At least they aren't under Windows. Here SSE/SSE2 have to handle everything by themselves. This appears to be a political decision and not a limitation of the processor architecture (according to the cpu manual, the FPU, MMX etc work in all modes).

(p.110)

...Conclusion: the mere transition to 64-bit registers will only result in better performance in a few exceptional cases, more often the opposite will be the case. Sure, there are a few examples such as RSA encryption using the Montgomery algorithm where 64-bit really shines. But in many cases the current 64- or 128-bit multimedia extensions can handle the job equally well. The same applies to string handling.



I tend to agree with CT here.

The kind of data we process in video applications are either 16 (YUV) or 24/32 bits (RGB) by their very nature so there really isn't much to be gained by packing this stuff into 64-bit registers.
Sure, one could pack two pixels into one 64-bit memory location but I doubt if anyone will implement that - it would mean severe changes to the Windows GDI API.

The new integrated MMU in the Athlon 64 is of course very welcome, but it would be equally welcome in a 32-bit CPU. It is not a merit of 64-bit architecture itself.


And let's not forget one important fact: AMD-64 is simply NOT binary compatible to Intel-64. Will hardware and software manufacturers really develop software and drivers for two totally different 64-bit beasts? How committed is Microsoft to AMD-64? Remember what happened to Windows-2000 for the DEC Alpha!

[Jerry Jones]

Wally wrote:

"I'd jump to 64 bits when I see the performance gains to justify the price. These early media encoding tests I found show sub P4 performance so far! (for what I care about most)."

I'd recommend re-reading that article.

The performance testing in the article was based on computer configurations not relevant to the actual development path of the AMD 64-Bit platform.

As I've mentioned, three factors need to be in place:

1. 64-Bit chip

2. 64-Bit operating system

3. 64-Bit application

Numbers 1 & 2 will arrive by the end of the first quarter of next year.

Number 3 will follow within 6 to 12 months.

Why would anybody want to buy outdated 32-bit technology - now - when an entire paradigm shift is coming so soon??

Clearly, the article you cited didn't test either chip in connection with video editing applications.

As this article...

http://news.com.com/2100-1006_3-5080217.html?tag=rn

...explains:

"The payoff from upgrading to 64-bit software involves higher performance from applications such as video-editing programs, along with more "cinematic" graphics for games. The improvements come mainly from a 64-bit computer's ability to use much more RAM than 32-bit computers can."

So when you're running your Pentium Extreme chip on 1gb of memory, the person who has upgraded to a state-of-the-art 64-bit workstation running 8gb of memory will be able to work in heavy-duty compositing applications such as BorisFX and After Effects and 3D applications such as Ulead Cool 3D Production Studio and enjoy no excuses real time performance.

To me, it boils down to this question:

"Do you want real time or not?"

If you do, then a 64-Bit system is the answer.

Jerry Jones

http://www.jonesgroup.net

[Jerry Jones]

Also - I just re-read Wally's article to confirm what I thought.

The test systems only had two 512mb ram sticks installed.

The test systems didn't feature the 8gb of ram that will unlock the potential of 64-Bit processing on 64-Bit Windows using 64-Bit applications.

As mentioned before:

http://news.com.com/2100-1016_3-5080...l?tag=mainstry

"While most performance PCs sold today come with between 512MB and 1GB of RAM, systems with between 4GB and 8GB will become more commonplace as 64-bit software works its way onto the desktop and as RAM module sizes gain more data storage capacity, AMD has said."

Since only the 64-Bit workstations will be able to *directly address* more than 4gb of memory, they'll make 32-Bit workstations look like the under-powered notebooks that used to run Windows 3.1.

Jerry Jones

http://www.jonesgroup.net

[Jerry Jones]

More interesting articles:

http://news.com.com/2009-1006-5080659.html?tag=nl

Jerry Jones

http://www.jonesgroup.net

[Ray Austin]

Oh, just stop it Jerry!

you're burning a hole in my pocket!

[Flying dutchman]

@Jerrold:

>Since only the 64-Bit workstations will be able to *directly address*....

I agree that having a larger address space is great. Great in the respect that applications can now map memory-mapped-files greater than 2 gigabytes into an application's virtual address space so they can use memory- and pointer techniques to navigate rather than use direct file I/O.

Other than that, most new PC's come with only 0,5 gigs of memory today. High-end machines have 1 or 2 gigs so few people have actually hit the limit of 4 gigabytes RAM (of which the bottom 2 are reserved for Windows).

I'm afraid that much of the gain that the integrated MMU of the Athlon 64 brings will be eaten up by the inefficiency of having to handle 8/16/24 bit data types in registers and memory locations that are twice as wide.

And let me make some remarks from the software-engineering side of the business (which is, after all, my profession since 14 years, I have some notion of what's going on under the hood).

You mention that three factors need to be in place (64 bit OS, applications and chip) but you forget the two most important ones:

4) - one needs 64-bit DRIVERS for EVERYTHING. Haven't you noticed that none of the Athlon-64 benchmarks had anything to do with audio/video capturing and processing, which happens to be one of the few terrains that really manipulate HUGE amounts of data? Guess why? No one has been able to capture anything yet!

5) - One needs a 64-bit optimizing compiler and development environment for the processor in order to develop software.

We are now at the threshold of a big change in software engineering. Windows XP will soon be on the market in three flavours for Itanium, Athlon64/Opteron and Intel-32. And these three platforms are binary incompatible as far as native code is concerned (everything can be emulated, of course).

So far, the development environment of choice has been MS Visual C++ for most companies, which comes with a good optimizing C++ compiler and which can also use Intel's own compiler which produces even faster, tighter code.
Intel sure as hell won't bring out a compiler for the AMD platform. Microsoft might do that, (they must have one, or how else would they have developed Windows XP for the Athlon 64?) but they have declared "native" machine code obsolete. The Microsoft compiler future is a beast called ".NET" which relies on an intermediate virtual processor code rather than native machine code. This intermediate code, a thing called "MSIL", is then compiled onto your system by a just-in-time compiler when you run the application. Write once, run everywhere, and no more DLL hell.
We've heard that all before, only then the beast was called Java and in my eyes Java happens to be the sluggiest slug on earth.

Now Microsoft had the luck that they could hire Anders "Turbo" Hejlsberg, one of the most brilliant compiler makers on earth, to do the development of .NET so we may safely assume that it will be faster than Java will ever be. And already many companies have jumped on the .NET bandwagon, even for low-level stuff (Powerquest Drive Image is a good example) so it must be a workable system.

So my guess is, that .NET is going to be THE way for companies to support all Windows platforms on the market, be they 32, 64 or no matter how many bits wide. But no matter how good the just-in-time compiler is, it's never as fast as hand-written hand-optimized machine code. So my fear is that much of the advantages of the new fast 64-bit CPU's will simply melt away.

I'd recommend you to hang on a bit to your 32-bit versions of MSPRO or Premiere, it may be a quite while before you get something faster!

[Jerry Jones]

Agree.

Now is not the time to buy anything new.

Wait a few more months for these developments to take shape.

Wise words.

Jerry Jones

http://www.jonesgroup.net

[wkulecz]

When you show me a 64-bit system where the price times the rendering time is less than for the current P4 I'll get intrested. Until then I've got better things to do with my money.

If you need 8GB of DDR RAM toget the performance gains its gonna be tough to compete in price performance. There is always as niche market for the "best" performance at any price, but in today's economy the patrons for this are getting few and far between. I work for a government agency that used to be one, but cheaper better faster is the new mantra.

--wally.

[Jerry Jones]

Interesting post to another forum from Martin Atkinson-Barr today:

"Most of the Ulead product will be written in a high level language so its just a matter of recompiling, the compilers will get better as time progresses."

"The critical parts will be hand-crafter assembler which will indeed be optimal for AMD64."

"There will be parts of the C "backbone" that could be better in 64-bits but they should be identified, if they are important, in the tracing stage."

"Most likely they will not be relevant to run-time (i.e rendering)."

"No matter how fast the chips run the way of the future is 64 bits."

"You see right now a 3GB file has to be mapped into sections and the virtual memory manager has to track those."

"With a 64-bit address space (actually you only need a part of that and the AMD64 Athlon64 implementation only uses I think
52 bits) you just map the whole file into virtual space so there's no "pauses" as the VMM remaps."

"The processor just sees the video file as one big, long part of memory - much simpler (and therefore faster) - than trying
to do it in a 4GB, 32-bit space."

"You'll see fewer glitches."

"Benchmarks may be done on small files but what counts is that everything runs smoothly on meaningful video files and AMD64 gives you that."

"Bottom line is - think when you process a video file."

"It starts OK and you leave it when you go to bed or whatever."

"In 32-bit systems at some point the OS has to remap portions and things go wrong, stresses the hardware."

"In AMD64 the first 5 minutes is just the same as 3+ hours."

"It will be more reliable, and faster."

"People into numeric processing have used 64-bits for years - expensive but necessary."

" Now anyone can buy it and that's great."

"It's not new technology as such, just better and cheaper."

"On top of that AMD has made an amazing chip from a technological point of view (apart from just the 64-bit stuff),
it does everything right."

"Just a few years ago we would have paid $$$$ for
this and now its what, less than a $1000."

"Just amazing."

[Flying dutchman]

Yup, just as I said, the larger memory-mapped files are THE one feature that may really make a difference. BTW: the Athlon-64 has a 48-bit address space, not 52, but that's still perfectly adequate.

I wonder where he gets his compiler, though, my version of Visual Studio 2003 only contains native Win32 support. And as I read in Ct, in Windows-64 API the integers (longs) are 32 bit wide, not 64, so "just recompiling" won't improve the application.

We'll see. The problem remains: No drivers=no show.