K10 -- What Hasn't Been Said

AMD has released a few hints about K10 (Barcelona). This has left people like Ed Stoligo and George Ou wailing and waving their arms trying to figure out what the numbers mean. The shrieking from Mr. Ou has been particularly shrill as he desperately tries to rationalize his belief that Intel will stay in the lead. Unfortunately, what George doesn't seem to realize is that these numbers are correct but not based on K10.

AMD has estimated benchmarks on its website. These include two graphs; let's look at SPECfp_rate2006.


What is most interesting is the quad core Opteron numbers. One would think that this would refer to K10 (since it is quad core) but it doesn't. The QC Opteron score is obtained by taking the Opteron 2222 SE score and doubling it (twice as many cores) and then adjusting it from 3.0Ghz to 2.6Ghz. Xeon 5355 (Clovertown) scales very poorly at only 75%. This is why the theoretical QC Opteron has such a lead. The lead is in fact a 50% higher SPECfp_rate2006 score at the same clock. Yes, this is the same 50% number that George desperately describes as "inflated" and which makes Ed so uncomfortable that he cuts it in half. Sorry, but this number is real; you just have to know where it came from. Back in January, AMD knew that the initial clock for K10 would be 2.5Ghz and they figured that Intel would have a 2.66Ghz Xeon. So, if QC Opteron is 50% faster at the same clock then this drops to 40% for a 2.5Ghz QC Opteron versus a 2.66Ghz Xeon. The SPECint_rate2006 score is 20% at the same clock. These are the numbers that AMD executives have mentioned in interviews. None of these are K10.

This is not complicated at all if you just look at it with a little common sense, but this seems to be what George and Ed have so much trouble with. Back in January, the estimated clock for Xeon Clovertown was 2.66Ghz but recently Intel announced 3.0Ghz. So, AMD changed their statement from 40% faster than the fastest to 50% faster at the same clock. This is neither inflation nor reduction; it is exactly the same number. The numbers for fp_rate are pretty good because they were all done with SUSE Linux. Intel's scores are even using its own compiler. The int_rate numbers are not as good because the Intel scores use Windows while AMD's still use Linux. We'll assume for the moment that these are accurate but the Intel numbers could actually higher. Let's adjust this based on what we know now.

fp
1.5X faster at the same clock, K10 at 2.5Ghz and Xeon at 3.0Ghz.
1.5X * 2.5Ghz /3.0Ghz = 1.25 (25% faster)

integer
1.2X faster at the same clock, K10 at 2.5Ghz and Xeon at 3.0Ghz.
1.2X * 2.5Ghz /3.0Ghz = 1 (0% faster)

This sounds very much like what Ed was saying and we might be tempted to give him some credit except that he never realized that the scores are for projected QC K8 rather than K10. I also don't feel like giving George any credit either even though he does make some correct calculations in his article. It is particularly annoying that George makes sure to get in the phrases "Barcelona hype" and "outrageous promises" at the beginning of his article and only gets around to saying "AMD's Barcelona might end up with a slight lead" down toward the bottom. I notice too that George makes sure to pump up Intel's numbers for higher clock and FSB speed but he too fails miserably to notice that AMD's numbers are for K8 and not K10.

So, let's continue where Ed and George got lost. K10 should be faster than K8. There are a lot of changes but the tricky part is to figure out what would allow four cores to not bog down and what would actually increase IPC. We should assume that the split memory bus, enhanced bus scheduling, increased cache, shared cache, and changes to load/store logic simply allow four cores to not get bogged down. Afterall, we are trying to stuff two more cores on the same bus. However, we know that K8 is not currently using all of its bandwidth and these things should give it a bit more room. Unfortunately, Clovertown cannot use enhanced scheduling because the two dies are separate. We also know that even the Revision F K8's can use as high as DDR2-800 and (from what AMD has said) this will probably be bumped to DDR2-1033 with K10. In other words, while Clovertown will be at 1333Mhz, K8 is already at the equivalent of 1600Mhz and should go to 2066Mhz when Penryn is at 1600Mhz. However, K10 has more improvements than this. We can probably assume that cache bus doubling, prefetch doubling, increased fast path instructions, improved branch prediction, and dedicated stack hardware will actually increase the speed per core. We also know based on what Intel has said that Penryn is 10% faster in integer at the same clock.

fp
1.5X faster at the same clock, K10 at 2.5Ghz and Xeon at 3.0Ghz. We allow 10% for Xeon with higher FSB speed and 10% for K10. The two 10% cancel so no change.
1.5X * 2.5Ghz *1.1 /(1.1 * 3.0Ghz) = 1.25 (25% faster)

1.5X faster at the same clock, K10 at 2.6Ghz and Penryn at 3.2Ghz. We allow 30% for Penryn with SSE4 and 10% for K10.
1.5X * 2.6Ghz * 1.1 / (1.3 * 3.2Ghz) = 1.03 (3% faster)

integer
1.2X faster at the same clock, K10 at 2.5Ghz and Xeon at 3.0Ghz. We allow 5% for Xeon with higher FSB speed and 10% for K10.
1.2X * 2.5Ghz * 1.1 / (1.05 * 3.0Ghz) = 1.048 (5% faster)

1.2X faster at the same clock, K10 at 2.6Ghz and Xeon at 3.2Ghz. We allow 10% for Penryn and 10% for K10.
1.2X * 2.6Ghz * 1.1 / (1.1 * 3.2Ghz) = 0.975 (3% slower)

If K10 gets the same 10% boost in speed then we come up with some interesting numbers. Basically, in integer K10 at 2.5Ghz would be perhaps 5% faster than Xeon at 3.0Ghz and K10 at 2.6Ghz would be 3% slower than Penryn at 3.2Ghz. In fp, K10 at 2.5Ghz would beat Xeon at 3.0Ghz by a pretty fair margin. However, at 2.6Ghz K10 is still a tiny 3% faster than Penryn at 3.2Ghz even if Penryn uses SSE4. So, taking us all the way up to about mid 2008, it looks like we have mostly a tie although it looks like Opteron may have a temporary lead in fp. These numbers are because of reduced scaling with quad core Clovertown but dual core would be better. On the other hand, K10 will clock to 2.9Ghz with dual core. This should put K10 dual core just a little bit more ahead but not much. I assume Intel will bump clock speeds again around mid 2008 but then AMD also moves to 45nm. This is probably about the best educated guess of speeds we can come up with before AMD releases some real benchmarks.