[citation][nom]brett1042002[/nom]^ Fixed. Same old AMD song and dance.[/citation]
You might have had a good point if we didn't know what Bulldozer's problems were. AMD knows what needs to be done and really, the biggest problems with it are not the hardest things to fix.
[citation][nom]Chesteracorgi[/nom]Nice speculation, and your source for this is??[/citation]
Here are three of the biggest reasons why Bulldozer failed. Try to keep up, I'm going to get a little technical here and I've noticed how some people on the forums can't understand the complexities of this.
Bulldozer based FXs had several problems (no surprise there). It's biggest problems aren't even architectural (not because of Bulldozer), believe it or not. I'll gladly explain what they were.
1. The Bulldozer FX die (there is only one right now, all FX CPUs use the exact same die) was designed using a poor method. All CPUs (excluding Bulldozer based FXs) are either entirely or partially hand designed. Most modern processors have partially computer generated designs for the least performance critical parts of the processor to increase the speed at which it is designed and the ease of designing it. Performance critical parts are usually hand-designed and optimized. The Bulldozer die was completely computer designed. Computers tend to know the fastest way to design a processor, not the way to design the fastest processor, so it's no surprise that it isn't very good.
On the other hand, Intel seems to do mostly computer generated designs, but they could simply have better tools for doing so. Intel having some of the best fabs in the industry probably also plays a good part in this.
According to former engineers at AMD, this decreases performance by about 20% while increasing die size and power usage by about 20%. The former engineer's claims have been confirmed by others, so it seems to be legitimate.
2. Now, this next one is an easy one to prove. FX's cache is very slow. That is also part of why it has so much of it... In an attempt to make up for it's low performance. FX's cache has huge latencies compared to Nehalem and Sandy/Ivy Bridge. I did not check a comparison that also included Yorkfield or at least Kentsfield and I do not know how much better they are than Bulldozer based FXs, if at all. Phenom II also has high latencies, but it is still often better than Bulldozer based FXs.
Also, Bulldozer has a lot less L1 cache than Phenom II does. How this effects performance is hard to say because for L1 a small, fast cache usually beats a large, slow cache.
3. FX has poor memory controllers. They can handle high speed memory, but they are inefficient. The FX memory controller has about 25% less bandwidth and significantly higher latency than the Sandy Bridge controller has with the same number of channels at the same frequency.
Those three problems are FX's biggest non-architectural problems (that I'm aware of). I could also mention some obvious architectural problems with Bulldozer if you like. However, fixing just these three non-architectural problems should increase performance greatly, decrease power usage significantly (although probably not as great as the performance increase), and also decrease the cost for AMD to manufacture the dies (smaller dies are cheaper because there are more dies per wafer and smaller dies have less area for a defect, so more CPUs can pass binning).
Piledriver based CPUs could have all of these, some of these, or even none of these (it's improvements could be strictly architectural) possible fixes implemented in order to improve upon AMD's previous processors.
Other reasons for improvements could be because AMD is finally implementing some of the RAM tech that they license for their CPUs such as T-RAM and ZRAM (Thyristor memory and Zero-capacitor RAM). T-RAM is as fast as SRAM, but as dense as DRAM. ZRAM is a little slower than SRAM and T-RAM, but is twice as dense as DRAM. Using T-RAM to replace SRAM would greatly reduce the amount of transistors needed for the caches without decreasing performance (in fact, the smaller size means that more of it can be closer to the components of the CPU that access it, so it's latency could have improved slightly). AMD has had licenses for T-RAM and ZRAM for years, so they might have a working chip with either of them by now.
Also, every single bit of news about Piledriver and Trinity points to large improvements, so it's more than mere speculation based on what can be done by AMD. They don't all agree on how much better Trinity and Piledriver are than Llano and FX/Bulldozer, but they all agree on great improvements.
With Bulldozer, we just had something that AMD had had been working on for a decade and little to no information on how it would perform, no info on it's serious problems, until it was out. Oh sure, there eventually was that crap about each core in a Bulldozer CPU not being a true core but only a significant fraction of one or some stupid *** about it being hardware multi-threading. However, looking at FX now, just fixing the non-architectural problems would bring Bulldozer up to modern performance levels, so it really isn't a bad architecture. Whether or not fixing Bulldozer's architectural problems will sky-rocket it's performance or not is unknown, but we know the problems and AMD knows what needs to be fixed.
It remains to be seen whether or not AMD fixes them and if so what they fixed. If you haven't heard any news about Piledriver, well, that's fairly odd considering how many times it's popped up even here at Tom's in some of their news articles.
My sources for AMDs cache problem are Anand's review of Bulldozer, the memory problem is from both Anand's review and Tom's reviews of Bulldozer, and the poor design method, that's from Softpedia news and another Tom's article. A quick Google search shows many other sites with the same news. The same goes for Piledriver and Trinity news.
Here is a particularly insightful read from [H], if you're interested (it involve the first problem that I cited):
http://hardforum.com/showpost.php?p=1037482638&postcount=88
And the anand article:
http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/2
http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/6
Sorry, but I lost my link to the Tom's article. Regardless, Anand is a good site and should be enough evidence for the cache and memory controller problems.
Facebook
Twitter
Instagram