Monday, December 04, 2006

My Issue with Anandtech | or | why I'm not buying Intel's Kentsfield numbers

By this point in time if you read any news sites at all that have anything related to technology, you probably have heard of Intel's Kentsfield, their Quad-core processor. You probably have also heard amazing performance numbers for it, and probably have been told that it is the highest performing processor you can get.

Excuse me while I call foul on these reports.

Fact is, I'm not buying Kentsfield performance numbers, not by a long shot.

Now, I have several issues with the numbers and performances being bandied about by Intel and Intel sponsored reviewers, and I'm going to start with Intel's Core2 processor. Core2 is why I'm fairly certain that once Kentsfield processors arrive in masse, people are going to find that it is a lemon.

Now, Kentsfield is built on Intel's Core2 processor design, better known as Conroe. It basically fits 4 of the Conroe cores onto one piece of Silicon. Let's review Core2 shall we? Now, if you listen to a lot of review sites out there, say Anandtech for one, you would believe that Intel's Core2 processor is the ultimate processor you can buy.

Not.. really. A lot of reviewers called foul on when Conroe launched. Many noted that the initial benchmarks were designed to fully utilize the massive on board cache of Conroe. Benchmarks showing realistic data sets were far and few between, and real-life use was a mystery.

Reality began to set in when HardOCP took the Core2 processor to the benchmark table. HardOCP strives to avoid using canned benchmarks, and instead tries to record actual game usage. Thier... results... were surprising to many of the Conroe backers. While Core2 was faster than Athlon64, it wasn't by a blow out. Rather, it was only 2-5% faster. The picture HardOCP painted got worse. HardOCP was using the 2.66ghz Core2 and the 2.93ghz Core2 on the tests, and the performance delta between the two was as small as the delta to Athlon64. 2-5%

So, lets do some quick math here, starting with the mhz speeds of 2660 and 2930.

2930 - 2660 = 270

2930 / 270 ~ 11

The 2.93ghz Core2 has about an 11% clock rate advantage over the 2.66ghz version, but the performance wasn't even close for 10% delta.

Now, I personally know that overclocking an Athlon64 2800+ Socket 754 on Biostar's T-Force by 250mhz nets a significant performance advantage. Even overclocking it by just 100mhz nets a larger than 5% performance increase.

I also know that dropping memory timings on the Athlon64 3200+ on Socket 754 from 3, 3, 3, 8 to 2, 2, 2, 5 netted a larger than 5% performance boost. In many cases using ultra-tight memory timings can net anywhere from 15-20% more performance out of an Athlon64 design.

Yet, as Core2 began to propagate out into the market, users have found that dropping memory timings didn't help. Overclocking does not help either. Once users begin to rise above the 2.6ghz mark, Core2 just does not scale with performance.

Many users agreed early on that the problem was Intel's connection technology. Intel is still using the Front Side Bus design to connect their chips to the rest of the system...

and that's why I think the Kentsfield / QuadCore performance numbers are manure. Boosting the FSB connection speed is only a band-aid on an overall larger problem. Core2 is saturated as is trying to move information in and out on Dual Core Systems. Does Intel really expect us to believe that QuadCore, built on the same architecture, can really be so powerful?

Core2, once it got out into the open, has proven to be a worthy competitor to the Athlon64. But the Athlon64 owner who pays attention to their choice of memory and their memory timings are not going to be outrun by Core2.

Intel Apologists no doubt can take the stance that, well, even if the performance numbers are hyped, it will be the most powerful processor period because of the massive caches. It still will be more powerful, so nyah.

The... counter argument... is that power isn't everything. Consumers also have to figure cost. Back when the AthlonX2 4800+ had just come out I ran a post called Rig Recommendations on a Planetside forum. Since then, I've passed the Rig Recommendations article around and it now sits on Mepisguides. One of the particular points I brought up in Rig Recommendations, and points I continued to bring up on forums, was that gamers had to factor in the cost of their processor compared to the power output.

I pointed out that each price point of an Intel Processor, the AMD processor of similar power output was cheaper. If you compared prices directly, the AMD processor was always much more powerful than the Intel Processor. To quote myself:
The Intel 3.8ghz 32bit costs $590

The Intel 3.6ghz 64bit also costs $590

The socket 754 Athlon64 3700+ that outruns them both costs $254
Now, being honest here, AMD hasn't exactly held the "Ultimate" performance crown that often. Remember the AthlonXp? Remember how the Intel 3.2ghz Pentium4 processor ran rings around the AthlonXp 3200 processor?

Fact is, while AMD hasn't always held the performance crown, they've been very consistent at holding the price crown. I still remember explaining to somebody in BurgerKing back in 2002 why he wanted an AMD processor and not an Intel design, and I explained it to him on price. Yes, if he had the money to get to get the most powerful system available, Intel was going to be his choice. But if was stuck on a budget, Intel wasn't even a consideration.

The same holds true today. If you want to build the most powerful system you can for a majority of applications, you are going to get a Core2 system, and now a Kentsfield. But, if you actually have been placed on a budget, or say a fixed incoming, AMD's processors deliver much better bang per buck.

Now, bringing it back to Anandtech. I used to read Anand's site on a daily basis a little over a year ago I think. Anandtech tended to be pretty well balanced. Till, around after Core2 launched, I began to get the feeling that all the motherboard and CPU articles were selling Intel Products. I didn't see any investigations into the FSB issues with Core2. I haven't seen any parallels on how Conroe's Saturation of the FSB mimics the Barton Saturation of the FSB. I haven't seen anything about how Intel will need to move to a Direct Connect Like Architecture in order for newer processors to scale up in performance.

Instead, I read the article about AMD's 4x4 platform and I'm greeted by expressions about how Core2 dominates peformance. Em. No. It doesn't.

Over time I've watched as Anandtech has slowely splattered egg on Anand's face with... questionable content... on the site.

I really wonder, when users actually start getting hands on with Kentsfield, how big is the backlash going to be? How many people are going to realize that they've been mislead?

Scoring high in PCMark is one thing. Being able to run City of Heroes for 8 hours during a Hamidon Raid, or hosting a Web server under *nix for years on end are quite different. And I personally think that Kentsfield will be a lemon once it hits real work.

No comments: