facebook rss twitter

Review: AMD Phenom 9900, and the Spider Platform

by Tarinder Sandhu on 7 December 2007, 08:07


Quick Link: HEXUS.net/qakmy

Add to My Vault: x

System setup and lots of notes

System AMD AM2+ Spider system Intel LGA775 system
Processors AMD Phenom 9900 (2.6GHz, 2MiB L2 cache, 2MiB L3 cache, AM2+, quad-core)
Intel Core 2 Quad Q6700 (2.67GHz, 8MiB L2 cache, LGA775, quad-core)
Cooler AKASA AK-876 AKASA AK-965
Motherboard MSI K9A2 Platinum (AMD 790FX + SB600) ASUS P5K Premium WiFi-AP Edition (Intel P35 + ICH9R)
Memory 2GiB (2 x 1GiB) Corsair PC8500 Dominator EPP
Memory timings and speed 4-4-4-12 2T @ 800MHz 5-5-5-15 2T @ 1069MHz
Graphics card(s) 2x Sapphire Radeon HD 3850 256MiB
Disk drive(s) Seagate 160GB SATAII (ST3160812AS)
Optical drive(s) Sony DW-Q30A
BIOS revision VP.0BC (11/08/2007) 0204 (06/26/2007)
Mainboard software Vista pre-installed drivers + AMD AHCI Driver 3.1.1540.11 Intel Inf
Graphics driver CATALYST 7.11 (8.432-071101a-054442C-ATI)
Operating system Windows Vista Business 64-bit
PSU FSP Epsilon 600W Enermax Galaxy DXX 850W
Monitor Dell 2405FPW


Benchmarks Sandra XII Lite Win64 float buffered memory-bandwidth
CPU-Z v1.41
LAME multi-threaded benchmark - 701.5MB file - encoded into 128kbps stereo.
DivX 6.6.1 (existing DV avi source-file, home-theatre profile, 1700Kbps, insane-quality video, 192Kbps, Stereo, 32KHz Audio)
POV-Ray 32-bit 3.7.0 beta 21a - internal benchmark mode
Company of Heroes: Opposing Fronts v2.103 DX9 - 1024x768
Quake 4 v1.30 - OpenGL - 1024x768
Enemy Territory: Quake Wars v1.2 - OpenGL - 1024x768 and 1680x1050


Our system setup page doesn't quite reflect the AMD-supplied box's. The reason for abandoning the NPRP (new product release program) kit lay with the unoptimised nature of the software build and BIOS. Comparing that to our lean-and-mean Intel build would have been rightly unfair, so we grabbed the Spider-forming components out of the NPRP box, installed them on our MSI 790FX chipset-based motherboard, and benchmarked on the same software build - Windows Vista Business 64-bit with a benchmark-friendly setup that removes some of the variances that are present in full-flavour Vista - to see if we could get an apples-to-apples comparison.

The Intel platform was constructed with similar value for money in mind. We've run the Core 2 Quad Q6700 processor - the natural competitor to the AMD Phenom 9900 - on a P35 chipset-based motherboard and with DDR2 memory, instead of X38/DDR3, which pushes platform price through the proverbial roof. It will be interesting to note how the P35's CrossFire-forming x4 lane negatively impacts upon performance when compared to the 790FX's PCIe 2.0 x16 for both cards. That's especially relevant as the Radeon HD 3850's 256MiB framebuffer will be quickly swamped and frequent access to main memory, via the PCIe bus, will need to be made.

We've used the same Corsair memory in both systems but have had to run the AMD box at DDR2-800 speeds. Try as we might on the NPRP box, the Phenom 9900's memory-controller wasn't partial to some DDR2-800-overclocked memory that required 2.1V to operate at DDR2-1066 speeds. We've seen this occurrence before, also on the MSI motherboard, and now begin to question the validity of AMD's DDR2-1066-supporting memory-controller. You'll note that we had no trouble in running the same modules at 1066MHz on the Intel platform, though. We've tried to be as even-handed as possible and run the memories at different latencies, to reflect the operating speed.

There's a huge performance-defining parameter that we've not touched upon thus far. Cool'n'Quiet 2.0, a feature of the Phenom series of CPUs, reduces clockspeeds when the cores aren't thrashed with computational load. Our testing on the NPRP kit, with C'n'Q enabled, highlighted a significant drop-off in performance when compared to the numbers on our own motherboard.

The technical explanation goes something like this: the C'n'Q implementation evaluates the load on each core and throttles it down if it's not judged to be near-100 per cent. The problem we encountered was that single-threaded applications would fool C'n'Q into construing load as non-optimal, resulting in the CPUs' cores throttling down and benchmark performance degrading as a result. We noticed the same sub-optimal performance in multi-threaded DivX and WAV encoding, where the energy-saving C'n'Q wouldn't allocate the full 2.6GHz of computational power to each core. To illustrate this point, we ran a complete set of numbers on the NPRP platform, with C'n'Q enabled, and will detail them on page eight.

Further, we know that AMD's Phenom CPUs suffer from a L3 cache-related problem - Erratum 298 - that causes a ~15 per cent drop in performance when 'patched' with a workaround. We understand the problem to affect a very small minority of shipping Phenom and Barcelona processors, and our benchmarks were carried out without the workaround in place and we noted no stability-related problems.

Coming back to our Spider platform, all the AMD numbers are from C'n'Q disabled, leading to maximum platform performance at the cost of the energy-saving. The same goes for the Intel platform, though.

Head on over here to see how the chipset compares to an X38, with significant emphasis on the southbridge's performance.