For the longest time, it has been difficult to get a true apples-to-apples comparison of the latest-and-greatest ARM (ARMH) processor cores against Intel's (INTC -2.27%) low-power processors. While the world has had to endure legions of poorly-written mobile benchmarks, the release NVIDIA's (NVDA 3.48%) Jetson TK1 development board on Linux has allowed Linux site Phoronix to do some benchmarking with meatier general-purpose CPU tests. ARM's Cortex A15 and Intel's Silvermont finally face off!

The results are pretty interesting
The Phoronix Test Suite includes a variety of workloads from ray tracing to code compilation. With that in mind, let's take a look at the results (bolded implies the superior score and some tests are in milliseconds so lower is better):

 

NVIDIA Jetson TK1

Intel Celeron J1900 (4 Cores @ 2GHz)

SciMark (Composite)

351.17

363.85

SciMark (Monte Carlo)

232.09

178.45

SciMark (Fast Fourier Transform)

42.21

56.48

SciMark (Sparse Matrix Multiply)

376.77

516.92

SciMark (Dense LU Matrix Factorization)

623.76

434.89

Jacobi Successive Over-Relaxation

481.03

632.48

VP8 libvpx Encoding

0.98

7.51

X264 Encoding

32.83

34.20

Timed Apache Compilation

138.09

125.01

Timed ImageMagick Compilation

345.56

298.81

C-Ray

100.90

131.90

OpenSSL RSA 4096-bit Performance

54.13

102.87

NGINX Benchmark

5931.98

7137.58

Source: Phoronix.

Interestingly enough, the quad 2GHz Silvermont tends to beat the quad ARM Cortex A15r3 (albeit at an unknown clock speed, although 2.3GHz would be a reasonable assumption given NVIDIA's earlier announcements on Tegra K1). While the VP8 libvpx results look like an outlier (perhaps this was hardware accelerated on the Intel but not the Tegra), the overall picture here is that in computational workloads, the ARM and the Intel trade off depending on the workload.

For the compilation and encryption workloads, the Intel seems to have a distinct advantage likely due to superior cache/memory performance (which is what such workloads stress) while the ARM seems to like floating-point intensive applications, as seen in the ray-tracing test. This makes sense given that the Silvermont floating point subsystem isn't particularly aggressive while the ARM designs tends to be more aggressive there.

Worlds collide
The unfortunate part of this testing is that we didn't get any power consumption numbers as this is a key metric in determining whether a mobile-oriented CPU core is successful or not. The rated "TDP" of the Celeron J1900 is 10W and from NVIDIA's recent statements, the K1 is also roughly a 10W part at full load. Of course, these numbers include the graphics processor, so in pure CPU workloads, the maximum power draw should be considerably less.

At any rate, the more interesting observation here is that ARM is scaling up and providing levels of performance roughly comparable to a low-end PC and Intel is pushing down into the mobile realm. In terms of CPUs, the next big face-off will be among the following in 2015:

  • NVIDIA's Project Denver
  • ARM's Cortex A57 (successor to A15)
  • Intel's Airmont and Goldmont
  • AMD's Puma+

The worlds of mobile devices and PCs have already begun to merge and the distinction will become even less clear as the next several years progress. The question from an investment perspective is whether the ARM camp will be more successful in moving up the value chain into PCs and servers or whether Intel will be more successful in pushing down into smartphones and tablets.