It's not that simple. For example when it come to 1997 the Celeron A series usually mopped the floor with the full Pentium 2 running at the same clocks while the Celeron A had a quarter of the L2 cache compared to the Pentium. The trick was that unlike with the Pentium, the cache on the Celeron ran at full speed.

Older games and programs still benefit from L2 cache, but they typically will use less of it. For example, I tested a Celeron 366 @550mhz (128k full speed L2) vs a Pentium 3 Katmai 550mhz (512k half speed L2). In 1990s 3D games like Quake2, HalfLife, and Unreal the Celeron was slightly faster (NOT mopping the floor). In later games like Max Payne and NOLF the P3 was faster because the Celeron's measly 128k L2 cache was no longer adequate.

I also tested a Coppermine Celeron 800 with 128k L2 vs a Coppermine P3 800 with 256k L2, both full speed, 100mhz FSB. This was where mopping the floor happened, only not in Celeron's favor. Even in 1990s games like HalfLife, Unreal, Descent 3, Quake 3 - the P3 was easily 10-20% faster, and even a P3 running 100mhz slower was still pulling ahead of the Celeron. So yeah, L2 cache is pretty important for 3D games.

In your case I'd stick with the P4. For 1990s games both are overkill and the difference from 1.6 to 1.7 is negligible. There is no benefit from a smaller L2 cache if both are the same speed and architecture. Also the Willamette Celeron was infamous for terrible performance, so in no case would it be a good choice.

pixel_workbench wrote on 2022-08-14, 15:20:

In 1990s 3D games like Quake2, HalfLife, and Unreal the Celeron was slightly faster (NOT mopping the floor).

If your cheapest $150 budget CPU keeps up with or beats your $850+ flagship, I call that mopping the floor (also that P3 is not an early 1999 CPU, though to be truthful the Celeron A came out in 1998).

RandomStranger wrote on 2022-08-14, 15:30:

pixel_workbench wrote on 2022-08-14, 15:20:

In 1990s 3D games like Quake2, HalfLife, and Unreal the Celeron was slightly faster (NOT mopping the floor).

If your cheapest $150 budget CPU keeps up with or beats your $850+ flagship, I call that mopping the floor (also that P3 is not an early 1999 CPU, though to be truthful the Celeron A came out in 1998).

I wouldn't. Mopping the floor to me implies a significant performance lead, not a bang/buck lead.

pixel_workbench wrote:

In later games like Max Payne and NOLF the P3 was faster because the Celeron's measly 128k L2 cache was no longer adequate.

More likely SSE is a thing.

pixel_workbench wrote on 2022-08-14, 15:36:

RandomStranger wrote on 2022-08-14, 15:30:

pixel_workbench wrote on 2022-08-14, 15:20:

In 1990s 3D games like Quake2, HalfLife, and Unreal the Celeron was slightly faster (NOT mopping the floor).

If your cheapest $150 budget CPU keeps up with or beats your $850+ flagship, I call that mopping the floor (also that P3 is not an early 1999 CPU, though to be truthful the Celeron A came out in 1998).

I wouldn't. Mopping the floor to me implies a significant performance lead, not a bang/buck lead.

Today yes. Today it doesn't make sense to buy a Celeron A to do a Pentium 2's job, since they sell the same. Back then, when you had to spend your hard earned 1998 dollars it mattered.

Actually it doesn't matter. They perform almost identical and you could argue that lack of external L2 cache, which is not cooled on stock heatsinks, is a plus.

pixel_workbench wrote on 2022-08-14, 15:20:

[...]

I also tested a Coppermine Celeron 800 with 128k L2 vs a Coppermine P3 800 with 256k L2, both full speed, 100mhz FSB. This was where mopping the floor happened, only not in Celeron's favor. Even in 1990s games like HalfLife, Unreal, Descent 3, Quake 3 - the P3 was easily 10-20% faster, and even a P3 running 100mhz slower was still pulling ahead of the Celeron. So yeah, L2 cache is pretty important for 3D games.

All other things - including cache speed - being equal. Which was the case here, but most definitely not with Mendocino vs Deschutes/Katmai. Katmai had SSE and struggled to pull ahead in titles that used it. Deschutes failed to pull ahead of a Mendocino at 333/66 or 450/100 in most games simply because the 2x faster cache with lower latencies due to being on-die usually more than compensated for there only being 1/4 of it.

This had actually been noted earlier with the P2 Mobile Dixon that Mendocino was derived from. Its 256kB on-die L2 easily beat 512kB off-die on Klamath and Deschutes.

Similar things were relevant during the various early Athlon designs, with 512kB off-die L2 cache at 50-33% of core clock on Argon/Pluto/Orion - and 256kB on-die L2 on Thunderbird beating all of them soundly.

In your case I'd stick with the P4. For 1990s games both are overkill and the difference from 1.6 to 1.7 is negligible. There is no benefit from a smaller L2 cache if both are the same speed and architecture. Also the Willamette Celeron was infamous for terrible performance, so in no case would it be a good choice.

Fully agreed here. The CPU architecture determines how sensitive it is to cache size. Due to the incredibly deep pipeline on Netburst, a cache miss was hugely costly, far more so than on 686 or K7 architectures, so cache size mattered far more. None of the WIllamette or Northwood Celerons were ever interesting as they were always outperformed by older P4 models at the same price point.

dionb wrote:

Similar things were relevant during the various early Athlon designs

K7 does not behave like P6. Firstly, all K7 had huge L1 cache with exclusive access, i.e. no data was required to be copied to L2 cache. Secondly, Thunderbird (and newer cores) cache was somewhat slow for full speed, which resulted in very marginal difference between different cores, including Duron with laughably small 64kb L2.

In my case, i just want to have the slowest 478 cpu. My main question is, if it will benefit older games or is it still an overkill.

Btw what would be a good cpu compromise to play late dos era games and also late 90s win games. My p4 is working with 90 % of all games fine, but there are some which make some problems like top gun or wipeout. My gpu is a gf2 ti. After long testing it is the ideal option. Still fast enough and also compatible with old drivers

You'll need to throttle it anyway. Optimally some CPU which multiplier can be divided by 4 or 8.