I stand corrected. Here's an article from Y2000 stating the original Celeron were in fact good for gaming!

"Early models of the Celeron did not include any level 2 processor cache (L2 cache), and as such their application performance was extremely poor. Games wise they were very good though, as the all important Pentium II core with its powerful FPU formed the basic design."
Source https://www.eurogamer.net/p3-800

I'd like to see some benchmarks though because Y2000 is also working from memory - not a reflection of live lived experiences.

There were Anandtech articles from 1998 that also claimed the L2 cache didn't matter for gaming, which was complete BS.

Here are some benchmarks I ran with a Celeron 300A comparing results in frames per second with L2 cache on vs off.

Quake 2: 88 vs 62
Half-Life: 25 vs 17
Unreal: 31 vs 25
Quake 3: 47 vs 31
Descent 3: 50 vs 32
NFS 3: 36 vs 21
NFS Porsche: 21 vs 11
Serious Sam: 30 vs 21

Tbf you're not comparing apples.
The p2 was half speed cache.

The 266 MHz Covington is on par with a 200 MHz PII Klamath or a 292 MHz Pentium MMX in terms of gaming performance, so it's not as bad as some might claim and given its OC potential plus the solid chipsets it can operate on, makes it a worthwhile choice over the K6-2 or even the K6-III based systems even today. For context, the 266 Covington clocked at 448 MHz (4x112) is just tad bit slower than a 350 MHz PII Deschutes (tested on a 440BX based M/B) and keep in mind, that both CPUs were launched at the same time, except the 350 MHz PII's price was at least 3x higher than the 266 MHz Covington's.

myne wrote on 2025-12-02, 14:04:

Tbf you're not comparing apples.
The p2 was half speed cache.

The half speed cache was also 4x larger. Anyways, here's the results of a P3-550 Katmai vs Celeron 366 @550:

Unreal... 54.0 vs 54.7
Quake 3... 44.1 vs 45.6

That's almost equal performance.

Perfect. Now, just with the katmai, drop to 366, and run.
Then disable l2, and run.

That's a lot closer to the original situation anand commented.

SSTV2 wrote on 2025-12-02, 14:15:

The 266 MHz Covington is on par with a 200 MHz PII Klamath or a 292 MHz Pentium MMX in terms of gaming performance, so it's not as bad as some might claim and given its OC potential plus the solid chipsets it can operate on, makes it a worthwhile choice over the K6-2 or even the K6-III based systems even today. For context, the 266 Covington clocked at 448 MHz (4x112) is just tad bit slower than a 350 MHz PII Deschutes (tested on a 440BX based M/B) and keep in mind, that both CPUs were launched at the same time, except the 350 MHz PII's price was at least 3x higher than the 266 MHz Covington's.

It's a neat / interesting cpu especially now that the dust has long settled.
Back in the day magazines used to rip it a new one though.

There was a lot of hype around 3DNow and the K6-3D aka K6-2. I think in addition to being relatively cheap that made K6-2 popular. K6-2 apparently launched like 1 month after Covington Celeron.

Celeron didn't have anything new but did have the highly publicized gimping with the removed cache.

I bought a Pentium II 233 in 97. Upgraded from an AMD 5x86. I was very enthusiastic about Pentium Pro and having one without 16-bit performance issues was my calling.

Here we go, Pentium II 266 benchmarks, with L2 cache ON vs OFF.
All tests ran at 640x480 16-bit, with a TNT2 Pro graphics card, driver 7.76.
Different graphics card and driver than my previous posts, so result are not comparable.

Quake 2... 60 vs 43
Half Life... 29 vs 17
Unreal... 34 vs 24
NFS 3... 47 vs 29
Descent 3... 56 vs 39
Quake 3... 30 vs 17

Subjective feel is also noticeably slower with cache off, in situations like Windows boot time or loading times for games.
Minimum framerates are also about 30-50% lower.

Something like ZD WinBench 98 or Winstone 98 would be interesting too. Probably not worth the effort though.

pixel_workbench wrote on 2025-12-02, 19:59:

Here we go, Pentium II 266 benchmarks, with L2 cache ON vs OFF. All tests ran at 640x480 16-bit, with a TNT2 Pro graphics card, […]
Show full quote

Here we go, Pentium II 266 benchmarks, with L2 cache ON vs OFF.
All tests ran at 640x480 16-bit, with a TNT2 Pro graphics card, driver 7.76.
Different graphics card and driver than my previous posts, so result are not comparable.

Quake 2... 60 vs 43
Half Life... 29 vs 17
Unreal... 34 vs 24
NFS 3... 47 vs 29
Descent 3... 56 vs 39
Quake 3... 30 vs 17

Subjective feel is also noticeably slower with cache off, in situations like Windows boot time or loading times for games.
Minimum framerates are also about 30-50% lower.

Hmm. I seem to recall there being next to nothing in it at the time.
Certainly not 30%+

Guess the thing is, that what made the Celeron 266 popular for gaming was the overclockability - i.e. you bought a 400 MHz CPU and paid for, well a 266 gimped CPU. But at 400 the fast FPU made it run circles around Pentium II 266-300, at least for a fraction of the cost.

H3nrik V! wrote on 2025-12-03, 09:14:

Guess the thing is, that what made the Celeron 266 popular for gaming was the overclockability - i.e. you bought a 400 MHz CPU and paid for, well a 266 gimped CPU. But at 400 the fast FPU made it run circles around Pentium II 266-300, at least for a fraction of the cost.

The cache generates heat, and removing the L2 cache increased the thermal headroom. Also, the 0.25 µm die ran cooler than the previous generation P.MMX 0.28 µm die.

P.MMX could overclock max 25%. Covington could overclock ~50%. That was a real measurable differentiator.

MattRocks wrote on 2025-12-03, 09:55:

H3nrik V! wrote on 2025-12-03, 09:14:

Guess the thing is, that what made the Celeron 266 popular for gaming was the overclockability - i.e. you bought a 400 MHz CPU and paid for, well a 266 gimped CPU. But at 400 the fast FPU made it run circles around Pentium II 266-300, at least for a fraction of the cost.

The cache generates heat, and removing the L2 cache increased the thermal headroom. Also, the 0.25 µm die ran cooler than the previous generation P.MMX 0.28 µm die.

P.MMX could overclock max 25%. Covington could overclock ~50%. That was a real measurable differentiator.

The heat of the package was not the problem with the L2 cache. The problem was that the L2 cache had a fixed 1:2 clock divider and the L2 cache just could not be as well overclocked as the CPU itself.
Guess why the Mendocino became so popular!
To omit the problem AMD later introduced a 1:3 clock divider for the L2 cache on its higher clocked Slot A processors.

Disruptor wrote on 2025-12-03, 11:24:

The heat of the package was not the problem with the L2 cache. The problem was that the L2 cache had a fixed 1:2 clock divider a […]
Show full quote

MattRocks wrote on 2025-12-03, 09:55:

H3nrik V! wrote on 2025-12-03, 09:14:

Guess the thing is, that what made the Celeron 266 popular for gaming was the overclockability - i.e. you bought a 400 MHz CPU and paid for, well a 266 gimped CPU. But at 400 the fast FPU made it run circles around Pentium II 266-300, at least for a fraction of the cost.

The cache generates heat, and removing the L2 cache increased the thermal headroom. Also, the 0.25 µm die ran cooler than the previous generation P.MMX 0.28 µm die.

P.MMX could overclock max 25%. Covington could overclock ~50%. That was a real measurable differentiator.

The heat of the package was not the problem with the L2 cache. The problem was that the L2 cache had a fixed 1:2 clock divider and the L2 cache just could not be as well overclocked as the CPU itself.
Guess why the Mendocino became so popular!
To omit the problem AMD later introduced a 1:3 clock divider for the L2 cache on its higher clocked Slot A processors.

There was certainly a perception that L2 heat was a limiting factor because Pentium III Coppermine (with twice as much L2) did not overclock as much as the Celeron Coppermine (with same core) - the only difference visible at the time was L2 cache size. But, with hindsight, both actually top out at ~1.1GHz so the cache is unlikely to be an actual differentiator.

I concede.

MattRocks wrote on 2025-12-03, 13:35:

Disruptor wrote on 2025-12-03, 11:24:

The heat of the package was not the problem with the L2 cache. The problem was that the L2 cache had a fixed 1:2 clock divider a […]
Show full quote

MattRocks wrote on 2025-12-03, 09:55:

The cache generates heat, and removing the L2 cache increased the thermal headroom. Also, the 0.25 µm die ran cooler than the previous generation P.MMX 0.28 µm die.

P.MMX could overclock max 25%. Covington could overclock ~50%. That was a real measurable differentiator.

The heat of the package was not the problem with the L2 cache. The problem was that the L2 cache had a fixed 1:2 clock divider and the L2 cache just could not be as well overclocked as the CPU itself.
Guess why the Mendocino became so popular!
To omit the problem AMD later introduced a 1:3 clock divider for the L2 cache on its higher clocked Slot A processors.

If L2 heat was not a problem, why does Pentium III Coppermine (with twice as much L2) not overclock as much as the Celeron Coppermine?

Apples and Pears.
I was referring to external cache chips.
Coppermine L2 cache is on-die and ofc has a 1:1 divider.

MattRocks wrote on 2025-12-03, 13:35:

Disruptor wrote on 2025-12-03, 11:24:

The heat of the package was not the problem with the L2 cache. The problem was that the L2 cache had a fixed 1:2 clock divider a […]
Show full quote

MattRocks wrote on 2025-12-03, 09:55:

The cache generates heat, and removing the L2 cache increased the thermal headroom. Also, the 0.25 µm die ran cooler than the previous generation P.MMX 0.28 µm die.

P.MMX could overclock max 25%. Covington could overclock ~50%. That was a real measurable differentiator.

The heat of the package was not the problem with the L2 cache. The problem was that the L2 cache had a fixed 1:2 clock divider and the L2 cache just could not be as well overclocked as the CPU itself.
Guess why the Mendocino became so popular!
To omit the problem AMD later introduced a 1:3 clock divider for the L2 cache on its higher clocked Slot A processors.

If L2 heat was not a problem, why does Pentium III Coppermine (with twice as much L2) not overclock as much as the Celeron Coppermine?

As far as I know, to get 200 MHz FSB on TUSL-2C, overclockers are looking for specific stepping of 566 P3 Coppermine 100MHz.
Not a Celeron, but a Pentium
In my experience, most CPUs based on the Coppermine core operate at 1.133 MHz, regardless of the size of the L2 cache.
The rest is a silicon lottery.

MattRocks wrote on 2025-12-03, 09:55:

The cache generates heat, and removing the L2 cache increased the thermal headroom.

The external cache itself is limited by the microchips available on the market.
Plus, there are problems with long high-speed buses on the cartridge.
You won't get much more than the nominal value from this solution.
The Pentium 2 was clearly not limited by heat.

If the problem were heat dissipation from the L2 cache, Intel would simply place soft thermal pads between the main heat sink and these chips.
This technology is not new.

I have not used Celeron 266 or 300 non a, but i have used Celeron 300a up to Celeron 533 non a and i dont have Celeron 533a.

Many a and non a-s here.

marxveix wrote on 2025-12-05, 12:12:

I have not used Celeron 266 or 300 non a, but i have used Celeron 300a up to Celeron 533 non a and i dont have Celeron 533a.

Many a and non a-s here.

Well, there are only 2 "A" models of the Celeron. The 300A to distinguish it from the cache less 300, and the 533A which is a Coppermine based, to distinguish it from the Mendocino based 533.

1000a too isn't there? Or was that a p3?