A few years ago I wrote about a strange NetBurst processor with SL7HY S-spec that landed at the OS/2 Museum. After renewed reader interest I pulled it out of the closet and tested the processor again. A collection of miscellaneous notes follows.
CPU-Z today is just as clueless as it was a few years ago (no surprise), thinking it’s a Socket 604 processor and that it’s an engineering sample:
The SL7HY processor was briefly tested in an ASUS P5PE-VM board. The BIOS complains at boot-up that it has no microcode for the processor, and clearly shows the brand string as
Intel(R) Xeon(TM) CPU 3.73GHz. That gives me confidence that this product string is really what’s burned into the processor.
The P5PE-VM does not support 1066 MHz FSB, but an Intel D975XBX2 (Bad Axe 2) does. In the D975XBX2 board, the SL7HY runs at the rated 3.73 GHz. It requires a sufficiently beefy cooler, but for example the standard Intel LGA775 cooler suitable for Core 2 Extreme processors does a good job.
A kind reader dug up a microcode update for the CPU. Note that the processor signature is 0F42h, which is a signature that is not included in Intel-supplied microcode update packages at all. But the recovered microcode update was tested with Linux and successfully applied; after update, the reported microcode revision changed to 3, clearly indicating that the microcode really was updated. Beyond that, it’s not at all obvious what the updated microcode does. The microcode is from April 21, 2005.
Considering the fact that the SL7HY processor is from late 2004 yielded several thoughts about the timeline and possible classification of the mystery CPU. First of all, it is a relatively early Socket 775 CPU (Socket 775 appeared in 2004), and it is a very early 1066 MHz FSB CPU, available at about the same time as the first chipsets/boards supporting 1066 MHz FSB (late 2004).
The SL7HY must also be one of the earliest Intel processors with NX bit support. NX bit capability also makes it definitely not a Nocona, and it is in fact a Prescott. In general it appears to be a very close relative of the 3.73 GHz Pentium 4 Extreme Edition CPU; the processor features appear to be identical (64-bit, NX, HT, enhanced halt, thermal monitor) and the only significant difference is 1MB L2 cache on the SL7HY vs. 2MB on the 3.73 GHz P4 EE. The SL7HY is also from late 2004 whereas the Prescott P4 EE is from February 2005. It is also noteworthy that the SL7HY CPUID signature is (as mentioned above) 0F42h, while the Prescott EE is 0F43h, hinting at the close relationship.
Note that there was a 3.46 GHz P4 EE with 1066 MHz FSB released in late 2004, but that was not based on the 90nm Prescott core. When the SL7HY shipped, it was possibly the fastest Pentium 4 available, at least for some definition of “available”.
The mystery CPU is also a very close relative of Pentium 4 550 S-spec SL7PZ which was released in November 2004. The SL7PZ ran at 3.4 GHz with 800 MHz FSB but was otherwise the same. If the SL7PZ were changed to run at 1066 instead of 800 MHz FSB and the multiplier reduced from 17 to 14, we’d end up with SL7HY.
Is it real or is it a fake? A reader suggested that someone could have re-lasered the label on the CPU. While that is true, I have difficulty imagining how someone could also change the CPUID brand string to read “3.73 GHz”. In my opinion the CPU is not a fake. In fact the SL7HY S-spec is quite close to other Prescott CPUs from the same era (SL7Z4, SL7E6, SL7K9 etc.).
That still leaves some questions unanswered, and the true answers might never be known:
- Why does the lid say Pentium 4 but the CPU brand string says Xeon? I really don’t know. It makes no sense.
- Why is the S-spec not listed anywhere? Probably the SL7HY CPU is an OEM-only “off the books” model, never available through normal channels and therefore never fully documented.
Note that the OEM model theory is not at all outlandish. There are other known “OEM only” S-specs from the same era, for example SL7QB or SL7Q8, and there have been OEM-only Intel CPUs before and after.
Update: It is maddeningly difficult to establish the timeline of which features were available when with Prescott CPUs. The Prescott line-up was, to put it bluntly, a hot mess. There were 32-bit CPUs, there were 64-bit CPUs, there were 32-bit CPUs with NX, 64-bit CPUs with NX, all with a mix of other random features like SpeedStep, enhanced halt states, thermal monitor, etc.
It is well documented what the S-specs were. It’s less well documented which S-spec was sold under what name, and it is even less well documented when a given S-spec became available. Intel’s ark is a mess and does not for example list the model “F” Prescotts which were the first to support AMD64 aka EM64T. Wikipedia does list those, with dead links for reference.
This problem is, of course, not exactly new. Back in 2004-5, people complained how hard it was to figure out which Prescott was what. According to this article, things were so bad that some Pentium 4 CPUs sold under the 540/550/560 processor number designation supported EM64T and others did not. Review articles showed readers how to figure out what a given box contains.
It is virtually certain that Prescotts with 1MB cache and all of HT, EM64T, and NX (same as the mysterious SL7HY) were available in late 2004. The Prescott specification updates (Intel document 302352-xxx) from 2004-2005 would provide good hints what happened when, but they are unfortunately not available… or at least weren’t until recently.
Late Update: Digging through the Prescott specification updates, there are several noteworthy data points:
Spec update 302352-006 is dated September 24, 2004 and it is the first to mention E0-step (0F41h) Prescott models, some of which had EM64T. But that’s not all–some D0-stepping Prescotts (SL7L9/SL7L8/SL7LA) had EM64T too! It is not clear from that spec update which processors might have supported the NX bit, but there are several errata related to NX.
Spec update 302352-008 is from November 2004 and it is the first to list the 3.8 GHz E0-step Prescott.
Spec update 302352-013 is from February 22, 2005 and it is the first to list N0-step (0F43h) Prescott-2M models, including the 3.73 GHz Prescott P4 EE. This spec update also clearly shows which models include C1E (enhanced halt states), TM2 (Thermal Monitor 2), NX, and Enhanced SpeedStep. With the exception of Enhanced SpeedStep, these features are not new in the N0 stepping and existed in some E0 models already.
Spec update 302352-022 is from October 2005 and it is the first to list G1-step (0F49h) Prescotts with 1M cache. Oddity: Errata lists show separate entries for Socket 478 and LGA775 G1-step processors, but no G1-step Socket 478 S-spec is listed.
The next spec update (302352-023) is from November 14, 2005 and it is the first to list R0-step (0F4Ah) Prescott-2Ms with VT-x support (3.6 and 3.8 GHz models).
Spec update 302352-026 from February 2006 lists several new R0-step models without VT-x (e.g. SL8PZ, SL8PY). Those appear to be the final Prescotts, with Smithfield and Presler CPUs now holding the line until the arrival of Core 2 processors.
Later Update: That Celeron D with S-spec SL8S4 is very, very fishy. There is no question that it was listed on Intel’s processorfinder site and claimed to have 0F42h CPUID and G0 stepping, and it’s still listed on ark. But it’s not listed in the Celeron D specification update, and I can’t find any evidence that anyone ever saw a S-spec SL8S4 Celeron. On the one hand, the fact that no photos or CPU-Z screenshots exist doesn’t prove that the S-spec never existed, on the other hand ark is known to list processors that never shipped (like the 4.0 GHz Prescott). Unless someone provably has or had a SL8S4 Celeron, its existence has to be treated as questionable.
The first Celeron D 3xx Sequence spec update which lists Celeron D 310 is Intel document number 302354-011 from August 17, 2005 and it only lists one S-spec (SL8RZ) with 0F41h signature and E0 stepping. Given that the SL8S4 S-spec was apparently listed on Intel’s web site in June 2005, that is quite odd. The situation is the opposite of the SL7HY which clearly exists but is not listed anywhere.
Even Later Update: Digging into the provenance of the 0F42h microcode update was worthwhile. It is included in the BIOS of ASUS P4P800-E Deluxe. The supported processor list for said board includes “Celeron D 310 revG0”, presumably the elusive SL8S4 S-spec, as being supported starting with BIOS version 1008. And sure enough, version 1008 from August 8, 2005 has the 0F42h microcode update, while version 1007 from May 25, 2005 does not.
Now, the P4P800-E is a Socket 478 board, and the SL8S4 Celeron is a Socket 478 CPU. So if that microcode update was intended for the SL8S4, it would actually make sense to have that in a S478 board. What’s more, the same BIOS also contains a microcode update for the G1-step (0F49h) Prescott, and both the 0F42h and 0F49h updates are dated April 21, 2005, and both are revision 3. That suggests the 0F42h and 0F49h processors are extremely close, and the G0/G1 stepping nomenclature makes sense.
What makes much less sense is that the same 0F42h microcode applies to the SL7HY Xeon which is significantly older and has rather different feature set (although that doesn’t mean it can’t be more or less the same silicon). None of this makes a lot of sense.
BIOS Update: After some fortuitous searches it’s now clear why CPU-Z calls the SL7HY processor a G0-stepping Nocona. Here MSI references a “G0 Stepping Nocona Xeon” in a BIOS update for a Socket 604 board. And here is a possible S604 Nocona engineering sample with QFFO S-spec but unknown CPUID signature. The best bit is here, claiming that a Dell Precision 670 (again Socket 604) BIOS included a Nocona G0-step microcode update for a processor with 0F42 CPUID signature. The Dell BIOS appears to store the microcode updates in a custom format, with a header which contains the CPUID signature and a four-letter identifier. For the 0F42 CPUID, the string is 0GCN, or NCG0 when reversed, which almost certainly means Nocona G0 stepping. There are similar strings like NCE0 (Nocona E0), PXA0 (Paxville A0), or IRN0 (Irwindale N0). What this implies is that the ‘Nocona G0’ designation for the 0F42 CPUID almost certainly came from Intel.
Curiously, Wikipedia claims the existence of a released G0-step Nocona Xeon with SL8RW S-spec, clearly based on data from Intel’s ark. But CPU-World disagrees and claims that the SL8RW has 0F49 CPUID signature, making it the G1 stepping. Intel’s actual Specification Update (302402-024) agrees with CPU-World (i.e. SL8RW has a 0F49 signature and G1 stepping), contradicting Intel’s own ark. Now who do you believe, Intel or Intel?