06 Tesla vs Core i7

Hardware Specifications and Roofline Limits

• Processor Architecture and Capabilities:
  • Intel Core i7-960: Manufactured on a 45 nm process, containing 4 cores and 700 million transistors, running at 3.2 GHz with a 130W power envelope.
  • NVIDIA Tesla GTX 280: Manufactured on a 65 nm process, containing 30 Streaming Multiprocessors (SMs) and 1400 million transistors, running at 1.3 GHz with an identical 130W power envelope.
• Peak Computational Throughput:
  • Single-Precision (SP) Floating-Point (FP): The GTX 280 peaks at 624 GFLOP/s, significantly outpacing the Core i7’s 85.33 GFLOP/s.
  • Double-Precision (DP) FP: The GTX 280 achieves 78 GFLOP/s, while the Core i7 reaches 42.66 GFLOP/s.
• Memory Bandwidth:
  • The GTX 280 delivers 127 GB/s of measured Stream bandwidth, providing $4.4\times$ the bandwidth of the Core i7’s 16.4 GB/s.
• Roofline Ridge Points:
  • The ridge point dictates the arithmetic intensity required to transition from memory-bound to compute-bound execution.
  • The GTX 280 DP ridge point sits at $0.6$ FLOP/byte, while the Core i7 DP ridge point sits at $2.6$ FLOP/byte.
  • The lower ridge point of the GTX 280 ensures that peak computational performance can be achieved at substantially lower arithmetic intensities.

To understand how these theoretical limits dictate actual execution, specific throughput kernels must be evaluated against these memory and compute boundaries.

Workload Evaluation and Performance Limiters

• Memory Bandwidth Constraints:
  • Applications with working sets spanning hundreds of megabytes (e.g., LBM and SAXPY) exceed Core i7 cache capacities, resulting in $5.0\times$ and $5.3\times$ speedups on the GTX 280 due to its raw bandwidth advantage.
  • Workloads processing large sparse matrices (SpMV) are constrained by DP FP limits rather than memory, restricting the GTX 280 advantage to $1.9\times$.
• Compute Bandwidth Constraints:
  • Strictly compute-bound kernels (SGEMM, Conv, FFT, MC, Bilat) scale directly with the raw FLOP/s capabilities defined by the roofline, yielding $2.8\times$ to $5.7\times$ speedups on the GTX 280.
  • The Bilat kernel relies heavily on transcendental functions; the Core i7 spends 66% of its cycles on transcendentals, whereas the GTX 280 provides direct hardware support for these operations, resulting in a $5.7\times$ speedup.
• Cache Utilization and Blocking:
  • Aggressive cache blocking on the Core i7 prevents data-intensive kernels from hitting the memory bandwidth roofline.
  • Because of cache blocking, Ray casting (RC) is only $1.6\times$ faster on the GTX 280, and SGEMM, FFT, and SpMV are shifted into compute-bound states on the Core i7.
  • The Sort kernel executes $0.8\times$ slower on the GTX 280 because the 1-bit split primitive requires significantly more instructions than a scalar sort operating entirely within the Core i7 cache.

Beyond raw compute and cache limits, memory alignment rules and thread coordination mechanisms drastically alter execution efficiency across these architectures.

Structural Dependencies: Memory Addressing and Synchronization

• Gather-Scatter Operations:
  • Multimedia SIMD extensions on the Core i7 require data to be aligned on 16-byte boundaries, severely penalizing scattered data layouts.
  • The GTX 280 implements native gather-scatter addressing, executing non-sequential memory accesses directly.
  • The GTX 280 Address Coalescing Unit and memory controller dynamically batch concurrent thread requests to identical DRAM lines or pages, minimizing gather-scatter latency.
  • The GJK kernel, which is highly dependent on scattered object data, achieves a $15.2\times$ speedup on the GTX 280 specifically due to this native gather-scatter hardware.
• Thread Synchronization and Atomics:
  • Throughput on synchronization-bound kernels (Hist) depends entirely on atomic memory updates.
  • The Core i7 utilizes a dedicated hardware fetch-and-increment instruction, holding the GTX 280 to a narrow $1.7\times$ speedup.
  • Kernels requiring the resolution of independent constraint batches followed by barrier synchronization (Solv) heavily favor the Core i7, which executes them $2.0\times$ faster than the GTX 280.
  • The Core i7 relies on its strict memory consistency model and atomic instructions to maintain order, whereas the GTX 280 lacks this memory consistency model, forcing it to launch synchronization batches inefficiently from the system processor.

The architectural deficiencies exposed by these synchronization and memory addressing workloads drove targeted hardware revisions in subsequent CPU and GPU generations.

Generational Evolution and Successor Architectures

• CPU Enhancements (Intel Xeon Platinum 8180):
  • Resolved the lack of non-sequential memory access by integrating hardware gather instructions (AVX2) and scatter instructions (AVX-512) directly into the SIMD execution units.
  • Achieves a $50\times$ aggregate performance improvement over the legacy Core i7-960.
• GPU Enhancements (NVIDIA P100):
  • Resolved synchronization and caching deficits by adding unified cache hierarchies and fast atomic operations.
  • Improved DP FP performance ratios from $\frac{1}{8}$ the speed of SP FP (on the GTX 280) to $\frac{1}{2}$ the speed of SP FP.
  • Achieves a $15\times$ to $50\times$ performance improvement over the legacy GTX 280.
• Comparative Scaling:
  • Despite CPU enhancements, the modern P100 GPU maintains a stable $2\times$ to $5\times$ throughput advantage over the modern Xeon 8180 across the core throughput workloads.