Some of the hardware I have that was released in 2003. Hard to believe it's been 20 years since this stuff was released, I was 17. Athlon XP 2500+ Barton, Duron 1.8GHz, 3x Duron 1.6GHz, and a Radeon 9600 Pro Advantage 256MB

Will 64-bit architecture ever become obsolete once 96-bit or 128-bit arc takes the market, surely it's not as future-proof as it was in 2003. It's similar to how IPv4 is obsolete and we've moved on to IPv6.

This time in 2003 I had an Epox 8K3A+, an XP 1600 cooled by a Thermalright AX7, 512 megabytes of DDR, and a GeForce 4 Ti 4200 with 64 megabytes.

Skyscraper wrote on 2023-04-18, 08:39:

Orientation. […]
Show full quote

Orientation.

This post will be constantly updated.

(April 2023)
At this point in time 20 years ago the Intel Canterwood chipset (875P) just had been released on April 14 along with the Pentium 4 Northwood 3.0 GHz. This platform was also relevant for workstations and small servers with Xeon Prestonia and Gallatin CPUs.
On the AMD-side the Barton Athlon XP 3000+ was the top CPU and the nForce2 was the best chipset. For workstations AMD had the 760MPX chipset and Athlon MP (2600+) CPUs. This Saturday (22-04-23) marks the 20 year anniversary of the launch of AMDs 64-bit K8 architecture in the form of the Opteron 240, 242 and 244.

Relevant platforms during the spring of 2003.

Intel.

875P - Canterwood
845xx - Brookdale-xx
E720x - Granite Bay, Plumas and Placer. (workstation/server)
860 - Colusa
850(E) - Tehama(-E)
VIA P4X333/400
SiS 645DX/648

AMD.
AMD 760MP(X)
AMD 8000
VIA KT333/400, KT600 in May.
nForce2
SiS 746FX

Relevant video cards during the spring of 2003.

ATI Radeon 9xxx
nVidia FX series
(9800Pro launched in Mars, FX5900Ultra in May. 9800XT and FX5950Ultra are only minor clock-bumps released a few months later.)

Relevant reviews.

Barton 3000+ review
https://www.anandtech.com/show/1066
Pentium 4 3.0C
https://www.anandtech.com/show/1093
Intel Canterwood (875p)
https://www.anandtech.com/show/1094
nForce2
https://www.anandtech.com/show/1005
AMD Opteron (K8)
https://www.anandtech.com/show/1098
Intel Prestonia and Gallatin
https://www.anandtech.com/show/1032
ATI Radeon 9800/9600 Pro
https://www.anandtech.com/show/1077
nVidia Geforce FX5900 Ultra
https://www.anandtech.com/show/1104

A Toms Hardware video showing an Opteron system just before the release 2003
https://www.youtube.com/watch?v=xDZbhftYyNk

Level1Techs on Youtube made a video on K8 and x86-64
https://www.youtube.com/watch?v=rMvwHMuTLGw

Dell Dimension 8300 also use 875P which I picked up for free, recapped it and rebuilt with ebay parts and local sourced parts, and used it for years at previous work (around 2005 and later). Very good northwood HT machine even at 2.8GHz. (Northwood 2.8C) using G450 dual head VGA ports. Unfortunately didn't kept it.

Cheers,

From 2003 I have the following:
Abit motherboards
1.socket 478: IC7-G; IC7-MAX3; IS7; AI7
2. socket 462: KD7A; KV7; NF7-s 2.0
3. socket 754: KV8-MAX3
CPU: Pentium 4 2.8C, 3.0C, 3.2C; Athlon XP 2500, 2800, 3000, 3200; Duron 1800
GPU: Ati radeon 9600xt from Abit; 9800xt; nvidia geforce 5950 ultra
HDD: WD 36gb and many other
Period correct Enermax PSU: eg465ax; eg365ax; FSP 350w
Fully assembled is my all time favorite: Athlon XP 3200; Abit AN7; 2*1gb DDR; WD 36gb; geforce 7800gs; FSp 350w with strong 5V rail. The rest are in storage in the basement.

twiz11 wrote on 2023-04-22, 21:17:

Will 64-bit architecture ever become obsolete once 96-bit or 128-bit arc takes the market, surely it's not as future-proof as it was in 2003. It's similar to how IPv4 is obsolete and we've moved on to IPv6.

Once 16 exbibytes of memory isn't enough for everyone anymore...

On the GPU-side wider architectures are more common.

.

@Everyone. Good stuff.

.

As my Opteron 244 CPUs are still in the mail I thought that I would bench some Intel Canterwood + Pentium 4C during the coming week.

I found an Asus P4C800 Deluxe i875P board!

Or so I thought...

.

After some searching I found this heavy Fujitsu-Siemens steel case.

The board and box is not a perfect match but an Asus P4C800 is as good as another I guess. I have not started the system yet so I don't know what model of CPU that's in the socket, probably something too new but that's easily corrected.

.

Edit

I found a Pentium 4 3.0C with the first production S-spec code SL6WK so this setup with the Radeon 9800 Pro will be very period correct for April 2003.

Some more progress in getting a state of the art April 2003 Pentium 4 system going.

In my box of sweets I found a couple of Corsair XMS3200 512MB sticks that should be able to do Winbond BH5/BH6 like tight timings @200. These sticks have Samsung ICs but as it's a downgrade if anything they should be fine for testing. The few decent Winbond 512MB sticks I have seem to be hidden away in systems...

.

Indeed the memory could do tight timings. If a system can do a full PCMark05 run they are usually half way to stable. I skipped the extra HDD-tests because this disk sounds like medium sized rocks in a cement mixer.

P4 3.0C Hyperthreading off.

P4 3.0C Hyperthreading on.

.

In hindsight I should have checked if I needed to replace the HDD with something less horrible compared to what ever is in there now making all the rattle.

Now I need to figure out which benchmarks to run.

It's been a busy day but after chasing out some gremlins I managed to run a benchmark. 😁

.

The Pentium 4 3.0C locks rather good in 3dmark 2001. I kind of doubt the Opteron 244 can beat this running at stock 1.8GHz.

.

Is this Vogons fastest period correct 20 year old system as of now?

Skyscraper wrote on 2023-04-25, 18:58:

It's been a busy day but after chasing out some gremlins I managed to run a benchmark. :D […]
Show full quote

It's been a busy day but after chasing out some gremlins I managed to run a benchmark. 😁

.

The Pentium 4 3.0C locks rather good in 3dmark 2001. I kind of doubt the Opteron 244 can beat this running at stock 1.8GHz.

P4 3.0C 9800PRO 3DMark2001.jpg

.

Is this Vogons fastest period correct 20 year old system as of now?

it might well be, looking very fast indeed, optimal. i dont think my P4s would reach it

gerry wrote on 2023-04-25, 19:16:

it might well be, looking very fast indeed, optimal. i dont think my P4s would reach it

Hardware wise it's probably as fast as a P4 3.0C could be and still be considered running fully stock. Using the Asus P4C800 line of motherboards is pretty much like cheating.

Asus tight memory subtimings and chipset timings made them dominant performance wise on the Intel side at the time. Overclocking evens out the field and even some 865 chipset boards can compete.

.

.

Benchmarking old systems takes time when using benchmarks that scale well with faster hardware.

I got a few more done. In these kind of benchmarks the Opteron 244 likely will be able to compete.

.

P4 3.0C Frybench x86. Render time: 01h 06m 19s

.

P4 3.0C SuperPI 1M. 44.219s

P4 3.0C SuperPI 32M. 39m 20.297s

Today I got a package with reason or two to speed up the Pentium 4 testing somewhat! 😀

.

Cinebench 11.5. P4 3.0C.

.

I'm not sure there is a reason to run any (32bit) 3dmark benches other than 2001 and 06 but here are all of them.

3DMark2001 P4 3.0C + Radeon 9800 Pro

3DMark03 P4 3.0C + Radeon 9800 Pro

3DMark05 P4 3.0C + Radeon 9800 Pro

3DMark06 P4 3.0C + Radeon 9800 Pro

.

[Edit]

I replaced the 3DMark2001 result with one using Catalyst 7.12 in the Omega version. This driver perform about the same as the Catalyst 7.10 in most cases so I won't replace the other results but for some reason the 7.12 driver is a smidgen faster in 3dmark2001. The reason I transitioned over to the Omega driver package was that it includes tools for overclocking, it saves time when installing.

[/Edit]

To round off the 3dbenching with the Pentium 4 3.0C and the 9800 Pro I ran Doom 3.

Doom 3 1024*768 Ultra P4 3.0C + Radeon 9800 Pro

.

Another benchmark I had forgot was AIDA64. I find that PhotoWorxx is great as an indicator of a well running, responsive system.

AIDA64 Cache & Memory Benchmark plus PhotoWorxx P4 3.0C.

.

And finally the seeds to this thread have arrived. Notice the mid 2003 packing date.

Skyscraper wrote on 2023-04-23, 17:27:

On the GPU-side wider architectures are more common.

That's because they count differently.
'64-bit' in terms of x86 CPUs means the register size and address space. More specifically, it means register size for general purpose integer instructions.
x86 also supports SIMD extensions like MMX, SSE and AVX, and in that case they already go up to 512-bit.

Since GPUs are also SIMD processors, they are generally only marketed as such. Individual variables on GPUs tend to be either 32-bit or 64-bit floating point. What's relevant is how many you can pack into a SIMD register, and process in parallel.
The address space on GPUs is also not relevant, as the GPU can only access memory on the card. Since video cards rarely have more than 16 GB, we're nowhere near the limits of 64-bit address space yet.
It's not clear what the actual address space on today's GPUs is, but I wouldn't be surprised if it's far less than 64-bit.
After all, these are not general purpose processors. They only need to work within the restrictions of graphics and compute APIs, and on their own video boards, where GPU and RAM are fixed.

In short: you can't make direct comparisons on 'bitness' between CPUs and GPUs. It makes no sense.

Scali wrote on 2023-04-28, 17:34:

That's because they count differently. '64-bit' in terms of x86 CPUs means the register size and address space. More specificall […]
Show full quote

Skyscraper wrote on 2023-04-23, 17:27:

On the GPU-side wider architectures are more common.

That's because they count differently.
'64-bit' in terms of x86 CPUs means the register size and address space. More specifically, it means register size for general purpose integer instructions.
x86 also supports SIMD extensions like MMX, SSE and AVX, and in that case they already go up to 512-bit.

Since GPUs are also SIMD processors, they are generally only marketed as such. Individual variables on GPUs tend to be either 32-bit or 64-bit floating point. What's relevant is how many you can pack into a SIMD register, and process in parallel.
The address space on GPUs is also not relevant, as the GPU can only access memory on the card. Since video cards rarely have more than 16 GB, we're nowhere near the limits of 64-bit address space yet.
It's not clear what the actual address space on today's GPUs is, but I wouldn't be surprised if it's far less than 64-bit.
After all, these are not general purpose processors. They only need to work within the restrictions of graphics and compute APIs, and on their own video boards, where GPU and RAM are fixed.

In short: you can't make direct comparisons on 'bitness' between CPUs and GPUs. It makes no sense.

The second line in my answer that you are quoting wasn't really any more serious than the first. 😀

To me a wider architecture means doing more things in parallel, more pipelines more execute units and so on.

While I will not argue that there is a real point in comparing CPUs and GPUs I do consider GPUs in general to be wider architectures but as you say not necessarily in bits of addressing space. I think the Russians actually tried to make a CPU using some (not necessarily wide) VLIW design but I have no idea how it performed.

Skyscraper wrote on 2023-04-28, 18:03:

To me a wider architecture means doing more things in parallel, more pipelines more execute units and so on.

While I will not argue that there is a real point in comparing CPUs and GPUs I do consider GPUs in general to be wider architectures but as you say not necessarily in bits of addressing space. I think the Russians actually tried to make a CPU using some (not necessarily wide) VLIW design but I have no idea how it performed.

But as I said, x86 CPUs are also far wider than 64-bit in terms of pipelines, execution units and whatnot. Which explains why Intel wanted to use x86 as the basis for a GPU, with their Larrabee project.
The Larrabee was technically a 64-bit x86 CPU, if you were to count like that. But if you instead count 'the GPU-way', you count it way differently. Then you take the 512-bit AVX units, and multiply by the amount of threads you can run in parallel by means of cores and HyperThreading.

Yes things are getting way more blurry than it once was. (I know Larrabee wasn't yesterday but I'm quite old)

Some of us on this forum are a bit stuck in the old. 😁

Regardless if you count the CPUs 64 bits of addressing space or the 128 bits on some oldish GPUs memory interface once you could be pretty sure the latter did more stuff in parallel (but perhaps with an architecture that's actually rather narrow depending on how you look at it).

Remember, 32 bits meant 4GB addressable range. GPU processes 16 to 64 bits. Quardros addresses more than 4GB on some models even the equivalent cores is limited much less vram on GTX series, The Quardo FX 5800 based on GT200 is the first to get 4GB, using GPU die similar to GTX 280 1GB. First gen Titan and next gen GTX 980ti was the first to break the 6GB range. Then GTX 1060 made 5GB and 6GB commonplace and GTX 1070 started to use 8GB.

Cheers,

Skyscraper wrote on 2023-04-28, 18:03:

To me a wider architecture means doing more things in parallel, more pipelines more execute units and so on.

Have a nice comparison (K8 vs. Zen4 + Prescott vs. Golden cove [a.k.a. Alder Lake P-Core]) 😀
Source : https://chipsandcheese.com

agent_x007 wrote on 2023-04-29, 12:25:

Have a nice comparison (K8 vs. Zen4 + Prescott vs. Golden cove [a.k.a. Alder Lake P-Core]) 😀
Source : https://chipsandcheese.com
arch.png

Nice comparison! Things are indeed getting wider.

It can't be all fat in Golden Cove as I'm very happy with my 12900k that I bought on launch day. The small cores have (of course) been disabled so the cache can run at 4.7GHz with the P cores @5GHz, both at less than 1.2v. This big little crap can stay in laptops. I have a "hacked" Xeon E5-2696v3 18-core on a Rampage V doing 3350MHz (33x101.5) on heavy 36 thread loads with tight 2133 11.11.11 memory. The 12900K using only its 8 P-cores @5GHz beats the turbo hacked Xeon in many well threaded benchmarks. We are finally seeing progress! Thanks AMD! Thanks ARM!

.

.

I have finally done some testing with a stock Opteron 244 with the Radeon 9800 Pro.

.

The Opteron 244 gets beaten in PCMark05 as expected.

.

In 3DMark 2001 the Opteron 244 comes close enough to hint that a faster clocked part would be... fast.

.

In 3DMark03 the Opteron 244 gets beaten by a reasonable margin.

.

In 3DMark05 the result of the Opteron 244 is a bit worse than I would have expected.

.

In 3DMark06 on the other hand the Opteron 244 beats the P4 3.0C by a large margin.

.

.

I will add more benchmarks later.

At some point in the future after benchmarking a few more systems the results will be presented in a moderately more practical fashion.

[EDIT] Updated the results with a bit more representative (closer to expected) scores. [/EDIT]

does the opteron feel 'responsive' enough in use though? it might emulate the experiences of early adopters of the time