Hacker News Reader

Comment by zelos | Hacker News Reader

Have the meltdown/spectre patches really had a roughly 30% effect on compilation performance on Intel CPUs? The i7-8700K went from 22 compiles/day to 14.

https://www.anandtech.com/show/11859/the-anandtech-coffee-la...

https://www.anandtech.com/show/12625/amd-second-generation-r...

The fact that the i7-8700K shows up as slower than the i7-8700 now is a bit weird too.

View on HN

> The fact that the i7-8700K shows up as slower than the i7-8700 now is a bit weird too.

It's impossible. The 8700K and the 8700 are the same chip, just that the 8700K has an unlocked multiplier and a higher default clock[0] - better binning. The 8700 will never be faster in any application.

[0]: http://www.cpu-world.com/Compare_CPUs/Intel_CM8068403358220,...

It is improbable, but there are many reasons why a specific 8700 might outperform a specific 8700k. There is a factor luck of the draw involved in the specific chips they test with.

Can you name some reasons? I thought concepts like luck of the draw in this context normally translate to silicon lottery and are meant to describe overclocking potential. But there is no reason why a chip sold as one of those cpus should not be able to reach the specified turbo clock, and in that case the 8700 just has to be slower in such a test.

The only thing I can think of is maybe the 8700 has better binning and is able to hit its turbo speed while putting out less heat and so is able to maintain its turbo speeds for longer than the 8700k.

I think it is more likely to just be testing margin of error

One reason might be that manufacturer can sell chips that would qualify as 8700k as 8700 simply to meet the 8700 demand if they have more good chips than they can sell at the higher price point.

I remember this happening for chips from the first step revision of Nehalem.

When the demand for the i7 920 was way up, there were a bunch of better chips sold as such.

Savvy overclockers knew which batch numbers they were and sought the particular runs.

Unlocked multipliers don't mean you've chosen the optimal multiplier. And something binned for higher clock rates doesn't necessarily mean it runs cooler at the same clock rate as another part. As others have mentioned, thermal management is a real issue here. A throughput benchmark doesn't care about brief bursts of speed, but rather sustainable rate.

Finally, even with identical parts, the assembly can affect performance. The same heat-sink installed slightly more effectively in terms of the mechanical interface (no contamination, just the right amount of heat-sink compound, no air bubbles) will allow the chip to run at higher power without any thermal throttling.

I honestly would have thought that the binning/testing done on those processors would make sure that the difference always go into the same direction, as in more performance for the higher specced part. Because I assume that in a benchmark just as this the cooling is sufficient to keep the turbo clock indefinitely, meaning that it makes no difference if the 8700 without limitations would beat the 8700K. But what you and the other say sounds convincing, maybe scenarios like this can happen in practice more easily than I thought.

It's a common real-world issue that someone buys the "faster" CPU option in a laptop and finds it runs slower than their coworker's because the thermal design and/or quality of assembly causes it to overheat and throttle.

With many systems, turbo is only a temporary burst benefit for a lightly loaded system. Many server or HPC systems would intentionally disable turbo to have consistent sustained throughput.

Definitely. But this is not a laptop with cooling issues, in this case we are talking about a professional benchmark setup. If cooling was an issue during this, it would just add an other negative point to the bunch of irregularities in this review.

There'd be something like a 2% error margin in their timings. They can only report what they measure. It's actually a good sign such discrepancies show up in their data; it means it is likely real data.

Reply on news.ycombinator.com