Financial Mathematics Text

Friday, July 11, 2014

Is AMD a better value than Intel?

Bruce Greenwald, in his book Value Investing: From Graham to Buffett and Beyond, devoted an entire chapter to Intel and its competitive advantages. Of particular note was its economies of scale which permitted it to spend less on research & development than AMD as a percentage of sales.

That appears to still be true today but less so:

I'm not sure how much one can infer from this. The technology market is very different today than it was in the past. Tablets have entered the arena while the PC market has declined. Some video game consoles now sport AMD processors and graphics cards. Intel entered the security arena with McAfee (which they now renamed Intel Security.)

Today I won't be inferring anything from this at all. What I'll be focusing on is the PC CPU space (and a brief look at video cards). I'll be doing so in a very restricted way but I think it may be enlightening nonetheless. (Whether or not it will be informative for an investment thesis is another matter.)

This research partly came about as a result of troubleshooting my machine's various problems. So this may be useful information if you're looking to buy/upgrade a PC.


So there's a lot of tools online (many of which can be downloaded for free) that run some software and see how your system performs and then calculates a number based on that performance. Here are a few examples of programs I've played with (there are many more):

Geeks3D has a variety of tests but of interest is FurMark. This stresses your GPU (gets it working at full capacity) which is useful to diagnose a failing GPU or see if an overclocking is stable. They have a list of some comparative results of graphics cards so you can see how well yours performs versus others.

3DMark has a variety of tools. There are limited versions which are free and ones that cost money with extra features. I played around with their basic version which can be used on Windows, Android and Apple devices. The free version shows your score on a page online which is stored. You can then search results for various tests and see how others did. Also pay attention to the number of cards used. Some use two or more cards (via SLI or crossfire technology) to enhance performance. You can filter for those that only have x # of cards. The advanced search allows you to search by both CPU and GPU so you can see what might happen if you, say, upgraded your graphics card.

Unigine has a number of benchmarks. I played around with their Valley benchmark. As far as score comparison goes, as far as I can tell, the only place to find them is on forums such as this thread.

There are two main points to make here.

(1) Results vary quite a bit. So when you see a score, that's probably an average figure. Each time you test your own system you'll get different results. And different systems will get different results.

(2) These tests only tell you how well a system performed at one particular task. It doesn't tell you how well these devices perform at other tasks.

The benchmark data I'll be looking is Passmark. They have nice tables set up with data on CPUs, video cards, hard drives, RAM and other things.

CPU Benchmarking

So let's start off with CPU benchmarking. Here's Passmark's High End CPU Table:

I took the liberty of removing results which had no price or where the price was an asterix (which seems to be mostly ones that are no longer available for purchase.)

The first 23 in the list are all Intel processors. Only 10 out of the top 100 are AMD. So as far as the best processors go, Intel appears to be a winner (at least according to this one metric.)

But I'm going to jumble up that list a bit (see below.)

AMD and Value

People often claim, however, that AMD is the better value. How can we measure that?

Passmark reports a "CPU Value" metric. This metric takes the Passmark CPU score and divides it by the price. There is a problem with this as it makes two assumptions which may turn out to be false. Here are the implicit assumptions:

(1) The relationship between the Passmark CPU score and price is a linear relationship.

(2) That linear relatonship has an intercept of zero.

Let's see if these assumptions hold.

Here's a chart of the above list plotting the Passmark score versus the price:

High End CPU Passmark Benchmark versus Price

The fit isn't too bad but clearly the intercept is much higher than zero. The logarithm has a slightly better correlation. The regression is somewhat driven by those outliers in the high price range. And most of those are actually the Intel Xeon series which are servers.

Now we might expect server processors to perform differently than a standard PC processor. Sometimes "torturing" the data can be useful (with the caveat that we must be careful since we can often torture the data to say anything we want.)

So what I did (and Passmark made it convenient to do so) was selected the most popular (as far as I can tell) sockets. For Intel these were the LGA 1150 & 1155. For AMD these were the AM3/AM3+ and FM2/FM2+. I believe this gets rid of the servers so we'll be primarily looking at processors going in standard PCs.

Here's what that looks like:

Passmark CPU Benchmark versus Price

There's obviously a much better linear fit here. The intercept is still not zero. As such, Passmark's "CPU Value" metric is misleading.

You can actually run the regression on each socket and get a different line. So take the line for what it's worth.

There's obviously some major outliers especially with the AM3/AM3+ socket processors. So either some of those are grossly mispriced or there's more to the story than Passmark's benchmark score tells us.

Here's this list. I've sorted them by % deviation from the line. The idea being those that are much higher than the line are a good value and those well below the line are overpriced.

If you sort by score, AMD only has 3 of the top 20. So even with this refined list, AMD loses out. Intel, on average, makes better processors.
But based on "value" (which I'm operationally defining as % deviation from the line), AMD has 13 processors, including taking the top 7 spots.

So while Intel may, on average, produce better processors, AMD still seems the better bang for your buck, at least according to this one Passmark metric.

GPU Performance

I'll keep this section brief because the charts look ugly. I suppose I could spend a ton of time torturing the data to make them look prettier but it would probably be a waste since torturing data too much is uninformative.

Passmark has GPU benchmarks here. The high end cards are listed here.

The chart looks ugly:

Passmark GPU Benchmark versus Price

Not only is the intercept not zero here, but it doesn't look at all linear. (You can draw a best-fit line but what would be the point?)

The problem, I think, has to do with the fact that GPUs are used for a variety of purposes such as gaming, video editing, mining Bitcoins, etc. The result is that one GPU may be optimized for one task so it may perform well at that task but not some other task. The Passmark Benchmark is just one score. It may not tell the entire story.

If someone wants to torture the data some to see if they can get anything, feel free. I couldn't come up with anything and I suspect torturing it more would just make it tell me what I want it to say.

As far as purchasing a GPU, you might just ask what you're interested in doing with it and then seeing how it performs at that task. For example, NVIDIA lists FPS for a variety of games for their grahpics cards. See the GTX 660 as an example.

I'm not convinced there's a good way (at least as far as Passmark's benchmark is concerned) of finding the "right" tradeoff between performance and price.

Summary Conclusions

1) Intel and AMD are quite different companies than when Greenwald initially wrote his book.

2) Benchmarks offer a way to compare devices at least in one aspect. The caveats here are that there will be some variability between results and devices may perform better in other ways.

3) Intel still, on average, makes better processors than AMD (at least by Passmark's CPU benchmark.)

4) AMD still, on average, offers better value processors than Intel (at least by Passmark's CPU benchmark.) You may be able to use that Deviation % as a means of picking up a good value processor (assuming Passmark's CPU benchmark is all that matters.)

5) Passmark's "value" metric is misleading as it assumes that the relationship between the score is linear with a zero intercept. In the case of CPUs, the intercept is not zero. In the case of GPUs, the relationship (if there is one) does not appear to be linear. It probably shouldn't be used at all.

6) In spite of fixing some of my computer problems, I still have a few that I can't figure out. I'm thinking I should rebuild it. (I'm also thinking I should never buy a ready-made PC again.)

No comments:

Post a Comment

Some common OpenID URLs (no change to URL required):