Product: AMD Phenom 9900 CPU
Availability: Hopefully before 1 April
New AMD CPUs have been a bit of a rare beastie in the test labs over the last year, and there haven't been too many tears shed over that: since late 2006, Intel has literally run AMD into the ground performance-wise, except in the four-socket and HPC markets - which are a very niche exception to the rule.
But, just before Christmas, an AMD CPU test platform turned up. It wasn't a fully-fledged Spider system, but an Asys M3A32-MVP Deluxe 790FX chipset and an unnamed "AMD Engineering Sample" CPU.
At the same time, Asustek dropped off two spare HD3870 TOP cards.
Add the OCZ 2 x 2 GB Platinum DDR2-800 CL5-4-4-10 RAM, plus an MGE 500W PSU. Here are our first impressions.
The AMD engineering sample in question is actually a Phenom 9900 2.6GHz CPU which has not yet been released, and, as it should appear in the high-end AMD desktop line some three months from now, it was a good fit to the high-end, quad-GPU capable Asus AM2+ socket mobo, and the twin HD3870 factory-tuned GPUs.
We used a CoolerMaster Hyper 212 air cooler, a reasonably good mid-to-high end offering proven to handle well above 150 watts under load, plus an extra fan to blow air over the heat-piped chipset/VRM portion just in case.
Take a look at the erroneous "ultra fast FSB" in CPU-Z... sure Intel can't beat this!
Windows XP SP2 ran with a brand new AMD Catalyst 7.12 driver, 3DMark06, Lightsmark, Everest 4.20 and Sandra XII SP1. The 3DMark and Lightsmark were run in both standard and Crossfire modes.
The board came with 0101 BIOS version, which worked fine with changing the multiplier and FSB settings of same Phenom 9900 processor. Upping the multiplier from 13 to 14 achieved seemingly stable 2.8GHz operation without any voltage change, but further upping to 15 for the magic 3GHz didn't boot, even if the voltage was pushed all the way from 1.30 to 1.35 volts. You can see the CPU-Z 1.42 screenshot for the 2.8GHz run.
Hoping to get better results, we updated the mobo BIOS to the newer 0701 version. However, that disallowed any overclocking or even voltage changes.
With the newest version, the 0801, the overclocking results remained the same: full stability at the default voltage for 2.8GHz, partial stability (i.e. Cinebench 10 and Povray failing) at FSB206 (2.88GHz) at default voltage, and full stability at 1.325 volts, but there was no way of booting Windoze at 3GHz, even at 1.36 volts.
In summary, you get the one-notch 7% overclock at default voltage, and some 11% overclock stability with a mild-voltage push - but anything more seems to hit the design limits. However, with a bit more work, 3GHz will probably be achievable.
Now, what can be compared with the Phenom 9900? In terms of price, Intel's Core 2 Quad Q6700 is the closest match. But then, the Q6700 has been out there for nearly a year while the Phenom will only be on the shelves in three months time.
This is a 45 nm-based Q9450, a $300++ class 2.66GHz FSB1333 Yorkfield which can, in most cases, comfortably do 3.5GHz FSB1750 on many X38/X48s and NV 780i/790i mobos at that time with regular air cooling. I believe that is a fairer comparison. You can see some early benchmark comparisons between these two in the screenshots below - it doesn't look good for AMD except in the Sandra memory bandwidth scores.
Finally, if you want to compare AMD's best vs Intel's best, then it is still the Phenom 9900 in March/April 2008 vs Intel QX9770 or a newer version of that. Keep in mind the current stepping overclocks beautifully till about 4.2GHz FSB1700 (air/water) or 4.5GHz FSB1800 (Asetek Vapochill) for everyday use with mild voltage adjustments - and that a new major Intel stepping, supposed to be out by March, should push these scores about 10% higher while keeping the voltages same.
Here are a few comparative benchmark figures: and there is just no point comparing, whether in default or overclocked modes. Intel wins everywhere, by up to twice in some cases. Have fun!
Phenom 2600 MHz default results:
... and 3DMarks with one & two GPUs, compare it with the Intel single-GPU result:
Lightmark on Q9450:
and Phenom 9900:
Lightsmark doesn't benefit from Crossfire, but does depend greatly on the CPU for effects like, say, radiosity. Intel wins here by a wide margin.
Keep in mind the power figures - the 130W TDP of Phenom 9900 will be some 50% higher than the rated Q9450 TDP, for lower overall performance. While the AMD North Bridge does save some watts compared to the X38, it still can't make up for this huge "carbon emission" gulf.
The Phenom lacks venom, but it's still a decent CPU. If AMD delivers Phenom 9900 as it is our in our tests and brings down the power consumption by about 10% by March, it will be a good mid-range CPU, priced around $300-400, for a decent desktop PC.
The problem is, these mainstream CPUs don't help push up the average selling price.
At the high end, the K10 needs some microarchitectural changes, especially the throughput improvement, to fight the Core 2. Seeing its performance compared with what Intel has, even a megahurtz push beyond 3GHz won't help the current K10 much. More on this in part 2, where overclocked results for both top Intel and AMD CPUs are listed.
The 45 nm round, which will hopefully appear in 2008, should better incorporate some good extra work on the core IPC, on top of doubling the L2 caches and tripling the L3 cache. Otherwise, whether Intel has any minor 45 nm mainstream process quirks or not, they will solve those quickly, and probably won't have any effect on the Nehalem launch round in 3Q 2008. For AMD, waiting for the high end 'Bloomfield' Nehalem to strike with only the current K10 as a shield, and cheap Yorkfields & Wolfdales biting from below - well, it's tantamount to a suicide. µ
Good: AMD native quad core finally works at decent clock
Bad: Power consumption, seeming FP letdown vs Intel
Ugly: Overall performance and availability, future prospect vs Q9450
Sign up for INQbot – a weekly roundup of the best from the INQ