Intel Shows Off Conroe (Desktop) Processors

[rhashem]

The G5 brought apple close to parity with Intel/AMD. And I think there is considerable doubt that PowerPC was ever in the lead.

The G5's quite legitimately beat the Pentium 4's and Athlon's of the time. Apple's GCC-based benchmarks of the G5 versus the Pentium 4 were perhaps the most legitmate benchmarks they've ever conducted --- no theoretical mastrubation about programs compiled with super-compilers that nobody actually uses.

However, a few months later, the Opteron came out, and once that hit 2.0 GHz, x86 took the lead again.

 

[rhashem]

This actually makes me sad. PPC's big selling point early on was that it was a much simpler architecture that could then scale to obscenely fast clock rates compared to the more complex x86.

The thing is, both the 970 and the modern x86 chips do instruction translation (x86 or PPC -> native, undocumented ISA). That's because PowerPC isn't quite as simple as the CPU implementors wanted it (ie: it has some 3 operand or 2 output instructions). Once you're doing instruction translation anyway, your core is limited by your native ISA, not the legacy ISA exposed to software.

Unfortunately, this hasn't been the case. And it's mostly due to manufacturing issues. If Intel was to take the design for the PPC970MP, and re-jigger it for their manufacturing, they could probably get it to fly at 3.5+GHz.

Heck, IBM could do that, if they rejiggered it to fix the power density issues. The 970 was designed for 3 GHz operation at 90nm (looks like they finally made it too). Going by historical projections, that means 4 GHz at 65nm (the same process as the Core Duo). The problem is that even at that speed, it'd be about as fast as a 2.5 GHz Conroe in integer code. That's the thing really hurting the G5 in a lot of code --- it's integer performance isn't enough to keep up with its FPU performance. Careful coding can get around that limitation, but most code isn't that carefully written (and GCC can't optimize for such a picky processor anyway).

 

[backdraft]

Usual Intel...

http://www.theregister.co.uk/2006/03/10/intel_heat/

 

[DeathChill]

http://www.theregister.co.uk/2006/03/10/intel_heat/

Wait, you're complaining about Intel talking about the advantages of technologies they're now implementing? They never said they invented them, but certainly Intel is in the lead power per watt wise.

Also, you wouldn't think you'd hear complaining about 'lies' on an Apple messageboard. Everything Apple says benchmark or performance wise is always tilted in their favour..

 

[thunng8]

The P-M is considered clock-for-clock equivalent to the G5 because of poor floating-point performance. The Merom design fixes that, so the SPEC scores for a 3 GHz Conroe are 2800/2500, versus 1438/2076 for a 2.5 GHz 970MP. So a 2.5 GHz Conroe would match the 970MP in SPECfp and outdo it by a large margin in SPECint. Considering Woodcrest (the CPU you'll need for a Quad-like machine, because of SMP) will be available up to 3.3 GHz, Apple doesn't even have to ship Intel's fastest parts to outdo the old Quad.

Just some minor corrections:

IBM just submitted spec scores for 970MP in their JS21 blade:
2.5Ghz: 1587/2119
2.7Ghz: 1706/2259

Which is a bit higher than the previous numbers quoted.

Conroe will clock at 2.67GHz at lanuch, not 3Ghz.
Woodcrest is an up to 3GHz part (and that where the 2800/2500 figure comes from). Also mentioned is an 80% specIntrate increase of Woodcrest over Paxville (Netburst Xeon chip). This would leave Woodcrest (dual core) with a specIntRate of approx. 34*1.8~=60.

Conroe XE is rumoured to clock at up to 3.33Ghz although nothing was confirmed at IDF about this part.

But overall, Conroe/Woodcrest/Merom looks very good and will definitely beat current dual core G5s in SPEC benchmarks (especially big advantage in Int).