Celeron M 'Shelton' performance analysed

CHIPZILLA HAS been rather secretive about the Shelton processor since its introduction in early 2005. This Celeron M, based on the 130nm Banias core but with no L2 cache, still receives no mention on its maker's developer pages. Oh the ignominy!

Originally reported as a low-cost challenger to Sempr0ns and VIA C3s in Asian markets, Shelton then shifted to target embedded designs.

The first boards that escaped into the wild used the 845GV chipset and featured a chip clocked at 1GHz it its pre-release configuration.

In release form, it was sold as a bundle of 600MHz, 0 L2 cache, 400MHz FSB, BGA-mounted processor, 852GM north bridge and an ICH4-M south bridge. This was priced competitively to the 1GHz VIA Eden and Geode NX. The question on our inquisitive minds was how much the lack of cache would have hurt this chip. Certainly a 600MHz Dothan would, on paper, have given a good shoeing to a C3, and maybe given the K7 Geode a run for its money, but how about with that cache missing?

Readers with long memories will remember the last time Intel produced an x86 CPU without Level 2 cache: the infamous Covington Celeron - launched in 1998. An experiment in cost reduction, Covingtons were thoroughly outperformed by, well, just about everything.

This sullied the word 'Celeron' for years, forcing Intel into the early launch of the ambitious Mendocino core, with its L2 cache integrated onto the processor die: par for the course now, but a big leap in 1998.

So what made Intel think that things would be any different now? After all the Banias and Shelton are very much members of the same P6 family as Covington.

Well, one thing would be the massive increase in bandwidth available to the processor from memory, and reduced latencies. The new bus, even with DDR266 memory (all that the 852 chipset will allow) offers 4x the bandwidth of old Covington, whereas the clockspeed has only doubled. Coupled with a much stronger hardware pre-fetch unit, the theoretical disadvantage should, on paper at least, have diminished.

Enough pondering, let's get to the numbers!

First: the competition. We tested the Shelton, or as it is otherwise known, the Celeron M ULV 600MHz (Zero Cache) against an AMD Geode NX 1GHz sat in a SiS741CX chipset, with DDR333; against the VIA Eden "Nehemiah" 1GHz with the CN400 and DDR400; and threw in a Dothan down-clocked to 600MHz on an 855GME board for good measure.

All chips (apart from the Pentium M) feature similar TDPs, are fanless and cost about the same. Sadly time constraints stopped us from performing all tests on all the boards, but here's what we came up with.

3DMark 2001SE
Internal Graphics
VIA Eden: 765
Geode NX: 1056
Shelton: 1123
Under-clocked Pentium M: 1831

First strike to Shelton. An impressive show with Shelton beating out the AMD and VIA options. The Intel graphics system proves to have the edge over the SiS. Adding 2MB cache and moving to the 855GME adds 63% to the score of the Shelton. The VIA is in last place.

To isolate the effect of the internal graphics, we put in an ATI Radeon 9000 card, and the results shifted around quite a bit:

Mobility Radeon 9000, 64MB
VIA Eden: 3737
Shelton: 4344
Geode NX: 5364
Under-clocked Pentium M: 5772

Even with the graphics removed from the equation, the Shelton retains a 16% lead over the VIA Eden clocked 400MHz higher, but now the extra horsepower of the K7-based Geode pulls it well clear with a 24% lead. The added cache of the Dothan still gives it a 33% boost over its baby brother.

Sciencemark 2.0 (32-bit)

	MolDyn	Primordia	Crypto	Stream	Memory	Blas
Shelton	109.78	162.78	233.14	303.83	489.17	162.5
Geode NX	321.55	324.46	436.06	359.26	359.26	456.9
AdvantageAMD	+193%	+99%	+87%	+18%	-27%	+181%

From this we can see that the Intel solution is efficient with its memory usage: even with DDR266, this shows a good advantage against the AMD with DDR333. However, despite this, the AMD is miles ahead on the real tests. We didn't test the VIA board, however the C3 has traditionally not scored well in Sciencemark.

Office Applications One thing that interested us was how the original Celeron's biggest weak point was in traditional office-type applications. How would Shelton perform here against the C3 with its higher clock speed and 64KB L2 cache?

We dug out some older benchmarks to do a modern day re-run:

	Eden1GHz	Shelton600	Advantage VIA
Business Winstone 2002	13.4	12.1	+11%
Business Winstone 2004	7.7	6.9	+12%
OfficeBench 2001	37.7	56.2	+49%

Well finally a result for VIA! It appears that the Shelton's old weakness is still there; and it runs "standard" software better than the Intel solution. Again, we didn't have a chance to run these tests on the AMD, but based on prior benchmarks would imagine it to have a strong lead over both solutions.

Video Playback On video, theoretically, VIA's CN400 with its integrated MPEG2/4 acceleration should have an advantage over the 852. This was borne out by our rough tests: playing back a DVD the C3 was averaging only around 15% CPU usage, whereas the Shelton was at 35%. A similar situation existed with an MPEG4 file, 40% for the C3 and 60% for the Shelton. However it's clear that neither of these chips will be ideal for your HDTV playback!

In Short
So there we have it, Shelton - a very mixed bag indeed. In some areas, it really does resemble the old Covington Celeron: terrible performance in office and scientific applications. However the strong chipset and higher memory bandwidth have lifted it up to be a real contender in graphics performance.

Overall our feeling is that AMD have a good little chip in the NX, offering a very well-balanced set of performance and power, and proving there's life in the K7 yet. Likewise, the C3 and identically performing "Luke" CoreFusion processor still has its strengths over the Shelton at 600MHz.

We'd have loved to have had a go at overclocking this, because without L2 cache, its low clock speed, and the famous low power of the Pentium M, we feel that it would have had some mammoth headroom, but our board's BIOS was locked down at the 400MHz bus, sadly.

On a final note, we've heard that Intel have just made a version at 800MHz available, though, while unlikely not enough to beat out the Geode NX, it should well match the C3 / CoreFusion in the weaker areas. µ