Dan Hagen

DailyTech – Report: AMD CEO Dirk Meyer Was Outed Over Tablets, Smartphones.

Awesome. Fire the guy who saved the entire company. Well, I guess now we can say goodbye to economical CPUs, and watch AMD’s prices skyrocket… plus the addition of shitty DRM if an Apple COO is taking over.

The company has been doing awesome with the last generation of chips. They seriously have no need to enter the ultra-portable market, but now firing Dirk is a tremendous mistake that will simply water down AMD’s resources to tablets and smarts while Intel is pummeling them with SB and Nvidia is back with a functional Fermi.

On the morning of December 24th I was excited to hear the UPS man ringing my doorbell with a new CPU cooling fan. After retrieving the package from my porch and cutting it open, my heart sunk as I realized that a large section of the fan was crushed into about seven different pieces. Crushed. I’ve always known that UPS isn’t the greatest shipping company by any means, given the fact that any package delivered by them looks like it was carried not in the truck, but dragged behind the truck. Nonetheless, I have never actually received an item victim to severe carrier damage until now. I didn’t rage though. There’s many reasons why I order from Newegg, and one of them is their customer service. I would simply go straight to my account on the Egg and request an RMA.

I had an interesting set-back though. When I specified that the problem was due to carrier damage, I was told that Newegg did not accept RMAs due to carrier damage and I would have to call customer service over the phone instead. With their offices closed on Christmas Eve and Christmas day, I had to wait until Sunday before I could make the phone call.

After clearly explaining to the customer service rep that it looked like carrier damage, she was swifted in issuing an RMA process for me. Which is funny, because the website clearly stated they don’t do RMAs for carrier damage, but I guess they do when you talk to support one-on-one. Whatever. I trust that customer support knows what they’re doing and I should probably have a new fan early next week.

I actually already have the exact same fan, which I ordered last spring, and it’s running in my computer right now. It’s a Startech 70mm “CPU replacement fan” that works perfectly if you want to keep your stock AMD heatsink but replace the crappy fan. This is one of my few complaints with AMD’s stock coolers. They work fine, but the fans always turn out to sound like shit. This is why I bought the Startech fan last spring, and it’s worked great, but I think I’ve played with it so much that it’s beginning to sound crappy now. It’s the loudest fan in my system and it produces noises that it shouldn’t. I’ve reinserted it about ten times and reapplied new linseed oil and stickers to the back of it about ten times as well, so it’s probably my fault. Once I get a new unadulterated fan I will simply swap it while refraining from doing anything else to it, and then put my current fan into my AM3-socket cooler computer downstairs which still uses the fan that came with the CPU.

Speaking of, I should probably sell that computer anyway, since it’s not getting any use and I could use the money.  Readers, if you need or you know anyone else who needs a new computer, let me know. You know who you are. All two of you.

From this thread.

In a typical AI routine, the computer looks at the current state of the game, and considers all the possible next steps that could happen, over some period of time (or moves, if it’s turn-based). This is sometimes called “state space search”. By considering all permutations it can decide what the best next move should be (i.e., the one that results in the best possible outcome given the opponent’s possible moves).

Some algorithms use a pruning technique to make the problem space easier (e.g., only considering some alternatives). But the process is still the same.

The issue is that you need to search as deeply as possible into the future, but there’s a fan effect in which the possible permutations increases nonlinearly over time. That kind of sucks if the human opponent is sitting there waiting for the computer’s next move.

So the computer budgets a certain amount of time to decide what to do next. A fast processor can look deeper into the problem space (set of possible outcomes) before that time runs out.

—–

It’s quite a simple explanation. If AI makes decisions in real-time, meaning that time constraints are factored in the decision-making process, then having a faster CPU will result in better decisions made by the AI.

As an analogy, consider the lightbulb testing problem, where you are to test how durable a lightbulb is by dropping it out of various stories of an n story building. You want to find out the lowest story (LS) at which the lightbulb breaks, and the measure of how well you perform this task is by how many lightbulbs you break in the course of finding this story LS. So, if the LS is 50, and you drop the lightbulb at 49, that is not considered a broken lightbulb, and every story dropped above 50 results in a broken lightbulb. The ideal broken lightbulb count is therefore 1.

There are many ways to approach this problem. The most obvious one is to start on the first story, and work your way up one story at a time, until the lightbulb breaks; then you have found LS. But, say each lightbulb drop incurs a time penalty, because it takes time to travel to a floor, as well as to drop the lightbulb. If you are constrained by time, you might not be able to use this algorithm, because you might have to try all n stories to discover an LS of n.

So, we can already see how increased processing speed will improve an AI’s performance. If an AI were to solve this problem under time constraints, it can choose which method of finding the LS to minimize the number of broken lightbulbs based on how much time it is given, and its processing speed.

Note that the fastest way to discover LS is by binary search, that is, try n/2, and if it breaks, try n/4, and if it breaks, try n/8 etc., more colloquially known as divide and conquer. This results in “poor performance” (many broken lightbulbs), but is bounded by log2(n), i.e. it will find LS in at most log2(n) attempts.

Alternate “algorithms” exist which trade-off time and performance in different ways. You can bound performance to two broken lightbulbs by starting at story sqrt(n), and going to 2sqrt(n), 3sqrt(n) etc. until the first lightbulb is broken, and then trying the sqrt(n) stories in between the last and second-last attempts one by one until LS is found. This guarantees a performance metric of 2 and a time metric of 2sqrt(n)-1 attempts. An AI which decides what algorithm to use based on time constraints might use a variation of this algorithm to discover LS given an amount of time.

My current computer contains an AMD Phenom II x4 965 3.4Ghz processor, which is basically AMD’s flagship quad-core CPU.

I’m running it at stock speed with the stock cooler, but I applied some Arctic Silver 5 thermal paste for better heat dissipation. As it stands, with regular fan settings and AMD Cool ‘n’ Quiet turned on, the Tcase sensor reports 33 degrees celsius while idle and maxes out at 60 degrees on full load (running Prime95). Core/Tjunction values are 4 or 5 degrees higher than Tcase values for my particular processor, which is normal.

Admittedly my computer case is very cheap. It has weak cooling, with a single rear 92mm exhaust fan. Tomorrow I’m going to install a second fan in the front. It will be a 120mm intake fan. I’m not sure how much it will do in the way of CPU temperatures; if anything, it should definitely improve my video card temperatures by a little bit. I have a factory-overclocked Gigabyte Radeon 5770 that has very good cooling – 3 copper heat pipes and a good fan – but it still gets considerably hot (low 70s) which I again blame on my cheap case.