Benchmarking Conroe: First Look at Core 2 Extreme
Loyd Case
31/5/06
Intel's new Core 2 CPUs are probably the most anticipated processors yet to ship since AMD's first Athlon 64. In many ways, Intel's new desktop CPUs, set to launch in July, are the most important CPU products for the company since the original P6 processor in 1995. Intel has been promising lower power utilization and greater performance.
We've described the Core 2 architecture back in March. The Core 2 microarchitecture gives us some clues as to why performance might be better. Conroe is a four-wide architecture, so can issue four instructions per clock, as opposed to the three-wide used in NetBurst and Athlon 64 architectures. The Core 2 will also contain a full 128-bit wide SSE (Streaming SIMD (single instruction, multiple data) Extensions) engine that can execute one SIMD instruction per clock. The Extreme and higher-end mainstream desktop CPUs will offer 4MB of shared L2 cache. Finally, the use of micro-ops and macro-ops fusion, which can combine certain types of instructions as they come into the pipeline, enhances performance.
But the architecture has a much shorter pipeline than the current Pentium D's—14 versus 31 stages. That suggests that overall clock frequencies will be lower. Can the new features overcome the clock rate disparities? We'll take a look at that in a moment. However, let's take a look at what Intel plans to ship this year. Read more on ExtremeTech.com.
Intel's CPU Roadmap
We met with Intel on May 31st, expecting to obtain a few details about what the company would be shipping in July. We got that, plus some hands-on time with a pair of systems. First, though, let's look at the roadmap.
Intel's David Tuhy, Product Marketing Director for Intel's Desktop Division, wasn't willing to divulge all the model numbers. He did say that Intel would simultaneously launch the Extreme and mainstream CPUs, avoiding the staggered launch that has generated criticism in the past. The Extreme part would be clocked at 2.93GHz, while the top mainstream part will run at 2.66GHz. Note that the only thing that differentiates the high-end part from the top mainstream CPU is the clock speed and the fact that the Extreme variant would be multiplier unlocked. All will ship with 1066MHz (266MHz actual FSB clock) front-side bus support. The 2.66GHz CPU and the 2.93 EE version will have 4MB of shared L2 cache. Some of the lower-end models will only have 2MB of shared cache.
Later in the year, Intel plans to ship a 3.2GHz Extreme CPU. Intel is also planning a quad-core CPU for enthusiast desktops sometime in Q1 of 2007.
Prior to the launch of the Core 2 CPUs, Intel will be announcing the 965P chipset. The new CPUs will run on motherboards with 965P chipsets, as well as those with 975X chipsets. Intel motherboards with 975X core logic that support Conroe have been shipping since April of this year. Intel will also launch integrated graphics versions of the 965 closer to the CPU launch date, some time in July.
The Core 2 Extreme 2.93GHz part will be rated at 75W thermal design power, according to Tuhy, with the mainstream CPUs at Intel's previously announced 65W goal.
Hands On Time with Core 2 Extreme
Intel also had two systems set up, both with Core 2 Extreme CPUs. Please note that these were systems configured and built by Intel. We did have full freedom to look at the BIOS settings, as well as run the tests we performed on our own, with no guidance from the Intel staff present. Here's how the systems were configured:
CPU Intel Core 2 Extreme @ 2.93GHz Intel Core 2 Extreme @ 3.2GHz
Memory 2GB DDR2/667 @ CAS 4-4-4-12 2GB DDR2/"733" @ CAS 5-5-5-15
Storage 2 x Hitachi SATA, RAID 0 2 x Seagate SATA, RAID 0
Cooling Stock Intel air cooler Intel Sealed Liquid Cooling
Graphics ATI CrossFire (2 x X1900) ATI CrossFire (2 x X1900)
We were able to run PCMark05 and POV-Ray 3.7, beta 13, and captured the results of CPU and memory tests when running PCMark05.
We need to note several caveats before we discuss the benchmark results.
* The stock system, running at 2.93GHz, ran with a fairly low latency DDR2/667 settings (4-4-4-12)
* The hard drives weren't the same. That shouldn't have any effect on the tests we actually ran, but it's worth noting.
* Since this was a CrossFire graphics setup running on Intel, we had little to compare it with. For example, we did run the F.E.A.R. benchmark, but we haven't run CrossFire on AMD's FX-62 yet, or even on earlier Intel CPUs.
* Our previous POV-Ray 3.7 results were run using beta 12. There's likely little change to the actual rendering engine between beta 12 and beta 13, but we need to report this. Note that in both cases, we were using the SSE-enabled version of POV-Ray.
The 3.2GHz system was overclocked using the Intel BIOS's 10% overclock configuration. This isn't quite the same as simply increasing the multiplier by one notch. This means that the system clock was increased 10% from 266MHz to 293MHz, but the CPU was still running with a multiplier of 11x. This yielded an effective front-side bus speed of 1168MHz, a memory speed of 733MHz (DDR) and a CPU core clock of 3.21GHz. Simply increasing the multiplier by one notch, which is how a 3.2GHz product would ship, would run with a lower memory speed, a lower front-side bus speed, and an actual CPU clock of 3.20GHz.
We added the results from these benchmarks to those we collected from our Athlon 64 FX-62 preview as well as those from our tests of a 3.73GHz Pentium EE 965.
With these thoughts in mind, let's take a look at what we found.
PCMark05 Tests
PCMark05 consists of a series of synthetic benchmark suites, each designed to test an individual subsystem, such as memory, processor, and hard drive. The test auto-detects which CPU you're using and loads dynamic libraries optimized for the processor under test for each function. So an Athlon 64 would run code tweaked to run best on its architecture, while a P4 running the same test would run different code optimized for that processor. Many of the tests are small enough to fit into the large L2 caches of modern processors, so those with higher clock speeds have an advantage. It's an idealized view of performance. In the real world, application optimizations can vary widely.
PCMark05 Results
PCMark05 CPU Multithreaded Test 1
PCMark05 CPU Multithreaded Test 2
PCMark05 Memory Read
PCMark05 Memory Write
PCMark Memory Latency
The overall PCMark scores are higher for the Core 2 Extreme CPUs. That's not a big surprise. What's interesting is the result of the CPU multitasking tests.
We've had some concern that the lack of Hyper-Threading may adversely impact heavy multitasking environment. The Pentium Extreme Edition 965 does support Hyper-Threading, yet the results of the PCMark Multithreading Test 2 shows the Core 2 processors easily outpacing the 3.73GHz 965 EE. Note that PCMark tests that comprise the four-task test likely all use SSE, and Conroe's single-cycle SSE engine shines here.
The memory test results are also interesting, and present a mixed picture. The Athlon 64 FX-62 outperforms Core 2 in the 8MB read tests, which are reads using data blocks larger than the cache size. And the 965 Extreme Edition beats Core 2 in the 192KB and 4KB tests.
The story changes for block writes, with Core 2 winning on the 4KB and 192KB tests, but losing on the 8MB write tests. Surprisingly, though, latency seems better.
POV-Ray 3.7 Beta
The 3.7 version of the open source rendering engine POV-Ray supports multithreading and SSE instructions. We ran the test with SSE turned on and the "all CPUs" option of the POV-Ray benchmark. Note that the Core 2 Extreme tests were performed with beta 13, while the other benchmarks were run with beta 12.
POV-Ray 3.7
We're reminded of the old Maxell tape ad, with the guy sitting in the armchair and his hair being blown back. We expected improved results, given the new SSE engine, but POV-Ray has proved sensitive to raw clock frequencies in the past. The improvement shown here is quite startling.
It's Preliminary
Given that we were running these benchmarks on systems at an Intel facility, using Intel-built systems, we're not quite ready to trot out the champagne quite yet. But what we've seen so far jibes with other reports we've seen. The good news is that these were on production-level CPUs, not early steppings.
Certainly Intel's David Tuhy exuded confidence in his new progeny. Intel is expecting systems to ship in volume the day of the official launch, which has only been announced as being in July. The good news for the homebrew PC builders is that retail boxed processors may not be far behind. Tuhy noted that Intel really wants to see boxed retail products in the white-box channel within a couple of weeks after the launch.
It is too early to pass final muster, and we weren't able to run our full benchmark suite. Until we can get our hands on motherboards and CPUs in our own labs, using our own gear, we can't give Core 2 our stamp of approval. But we're certainly encouraged. All this assumes Intel can really deliver both in volume and on time. If all this comes to pass, then the company may well have a winner on its hands.
Source: http://www.extremetech.com/article2/0,1697,1970192,00.asp
|
