Intel Announces Devil's Canyon Core I7-4790K: 4GHz Base Clock, 4.4GHz Turbo 157
MojoKid (1002251) writes "Last year, Intel launched two new processor families based on the Haswell and Ivy Bridge-E based Core i7 architecture. Both chips were just incremental updates over their predecessors. Haswell may have delivered impressive gains in mobile, but it failed to impress on the desktop where it was only slightly faster than the chip it replaced. Enthusiasts weren't terribly excited about either core but Intel is hoping its new Devil's Canyon CPU, which launches today, will change that. The new chip is the Core i7-4790K and it packs several new features that should appeal to the enthusiast and overclocking markets. First, Intel has changed the thermal interface material from the paste it used in the last generation over to a new Next Generation Polymer Thermal Interface Material, or as Intel calls it, "NGPTIM." Moving Haswell's voltage regulator on-die proved to be a significant problem for overclockers since it caused dramatic heat buildup that was only exacerbated by higher clock speeds. Overclockers reported that removing Haswell's lid could boost clock speeds by several hundred MHz. The other tweak to the Haswell core is a great many additional capacitors, which have been integrated to smooth power delivery at higher currents. This new chip gives Haswell a nice lift. If the overclocking headroom delivers on top of that, enthusiasts might be able to hit 4.7-4.8GHz on standard cooling."
Re:More useful metrics? (Score:4, Informative)
Because it doesn't have the wow-wee factor compared to raw clock speed numbers.
Re:More useful metrics? (Score:5, Informative)
Why don't we ever read about more useful metrics, such as the amount of (floating-point) operations per second per $ of a given CPU?
Because the target market for this thing doesn't consider that a useful metric, and never has.
For some years now (at least back to the P4 era, if memory serves), Intel has always offered the mad-crazy-overclocker-must-go-faster-edition CPU at the top of their (desktop, sorry Xeon buyers!) price list, usually ~$1,000. This part is always an astonishingly poor value, unless what you want is the fastest x86 money can buy. Most of them go to gamer e-peen setups, they may sell some to compute customers who have some pathologically hard-to-parallelize problem and thus need the fastest single threaded performance they can get, rather than more cores with lower performance per thread but far lower cost.
If you are actually shopping for CPUs, you probably want something like CPUboss [cpuboss.com], or CPUbenchmark [cpubenchmark.net] which allows you to do fairly easy comparisons of performance/price (albeit for performance as measured by one or more general benchmarks, if your workload is somewhat atypical, your mileage may vary).
Re:More useful metrics? (Score:4, Informative)
Why don't we ever read about more useful metrics, such as the amount of (floating-point) operations per second per $ of a given CPU?
Because most people don't care about these things anymore. Take this from TFS:
Haswell may have delivered impressive gains in mobile, but it failed to impress on the desktop where it was only slightly faster than the chip it replaced.
In reality, Haswell had double the FLOPs thanks to the new FMA instructions, near double the integer throughput thanks to AVX2, and a significant boost to multithreaded code thanks to TSX.
In practice, people saw maybe a 10% speedup in what they actually do. A flops/$ metric would significantly inflate the actual value people would see from these CPUs.
The thing is, these measurements are either synthetic (who has code consisting of nothing but FMA?), hard and uncommon to use (Integer SIMD is rare and AVX2 has a confusing idea of "lanes" that splits some 256-bit ops into two 128-bit ones), or not on all CPUs (TSX is disabled on their unlocked K line for some reason).