Talk:Bulldozer (microarchitecture)

Update of roadmaps
My recent edit removes some of the previous information from earlier roadmaps that has been superseded, and adds links to the most recent roadmap from Analyst Day 2009. The Bulldozer architecture diagram the last paragraph links to can be found under Chekib Akrout's presentation, though I can't link to it directly as it doesn't open unless you save the PDF first then open it locally. Please don't add the Falcon and Sandtiger references back in, the former is cancelled or renamed and has been pushed back to 2011 while the latter has not been used since 2007's Analyst Day (indeed, M-SPACE hasn't either and I'm inclined to remove that too) and an unofficial comment by AMD's John Fruehe suggests neither is in use since the roadmap changed after the 2007 day. 86.155.150.0 (talk) 15:15, 14 November 2009 (UTC)

Sony "ORBIS"
Seems like this may be the new processor (4 of them anyways) for the Playstation 4. Some specs on the console specific hardware would be useful. — Preceding unsigned comment added by 173.87.170.38 (talk) 19:41, 24 January 2013 (UTC)

Partitioned Multi-threading needs to be mentioned instead of using AMD's "module" marketspeak
This needs to be mentioned to clarify any confusion over the fact that each "module" is in fact a partition of the *core's* execution resources -- each "module" shares the front-end for the core in question as well as the FPU although both have separate L1 data caches. Look here for more information: --Jdbtwo (talk) 21:09, 21 January 2010 (UTC)


 * Each Bulldozer "module", while appearing as a hardware thread to the user, is in fact not a core where the definition of "core" is a self contained microprocessor that can stand on its own via some external interconnect -- clearly AMD's "modules" are not "cores."
 * A Bulldozer module (with two 128-bit FMAC/Integer Cores) is a single unit and won't be sold as a dual-core (like many may think), but as a "single core". So it will compete with single cores. A 16Core processor will feature 16Modules NOT 8 Modules!!! Look here: http://www.anandtech.com/show/2881 —Preceding unsigned comment added by 84.75.175.86 (talk) 18:25, 5 July 2010 (UTC)


 * Both of these preceding comments seem to contradict page two of the Anandtech article of the author's correspondence with AMD, wherein each module contains two integer cores, and that AMD is marketing these processors based on the number of integer cores, with each core appearing as a separate thread to the user. So then, a "16-core" processor would feature only 8 modules. In addition, the PDF linked by Jdbtwo is excessively long to be used as stand-alone evidence of this claim and should be quoted, or at least a reference is needed as to where to find the relevant information within the document. Furthermore, the document was published in 2005 and has the potential to contain some outdated theory or terminology as technology advances. It is cause for a severe case of TL;DR and hardly gets the point across any better than no reference at all. That said, having read the document in it's entirety, I do not see how it supports Jdbtwo's claim that each "module" would be a partition of a core's execution resources, when other sources would claim it to be exactly the opposite.


 * I wouldn't trust Anandtech for a scientific viewpoint of CPU architecture given their overt Intel bias and especially because of the fact that they cater to a general tech savvy audience, not to computer scientists or electrical or computer engineers. Also, a company can market a product as whatever they want, within limits -- just because they call it something doesn't mean it is is that something -- would you base a course an a particular computer architecture just on material from the company's marketing department, without actually exploring the real technical details of the architecture? Such a thing would be foolish. Furthermore, just because a source is excessively long doesn't disqualify it from being valid -- how many Wikipedia articles reference whole books? And I can give you a specific reference within the document. Also, just because a document is not current doesn't mean it's invalid, as long as its facts stand on their own and support the material from which it is referenced I don't see a problem. Theoretical CPU architecture, especially that relating to superscalar and multithreading techniques hasn't changed much since 1995 in general -- the basic concepts are the same. Superscalar processors today generally use Tomasulo's algorithm and that was developed in 1967 so I hardly see your point. You say you've read the document in it's entirety, but it seems you didn't particularly comprehend its relevance to the Bulldozer architecture. When I claimed that a module was a core I was misinformed as there was very little information on Bulldozer available, but now I do agree that a module consists of two cores. Partitioned Multithreading needs to be mentioned because *the only difference* between AMD's module design and full blown Partitioned Multithreading is that in Partitioned Multithreading *instructions from two or more active threads can be mixed and distributed across both partitioned execution pipelines* whereas in AMD's module design *only instructions from one thread are allowed in each of the two execution pipelines.* *This is the distinguishing quality that gives each execution pipeline the status of a full blown core.* Jdbtwo (talk) 11:48, 6 May 2012 (UTC)




 * It's also important to note that *Intel* coined the word "core" with respect to x86 processors, so what is a "core" and what is not a "core" is based on an arbitrary marketing definition and shouldn't be used for reference. Before that time, multi-"core" CPU's were referred to as having multiple execution units or having multiple processors on one chip.--Jdbtwo (talk) 05:09, 1 January 2014 (UTC)

Completely new design?
Completely new design? What does that mean? Some clarification is needed. I'm sure they didn't make a whole new architecture. Maybe the microarch is different? 76.180.19.184 (talk) 05:06, 15 February 2010 (UTC)
 * it's a completely new design, it's impossible to implement some of these features on the old pre-intel agreement AMD64 (that architecture was designed to avoid stepping on intel's patents, this one uses all of intel's current and latest IP)Markthemac (talk) 22:41, 7 December 2010 (UTC)


 * No, I don't think it is new design, as you see here, the initial AMD K8 microarchitecture is much similar with this "Bulldozer". After more than ten years, AMD redesign their abandoned processor. Not new at all! — Preceding unsigned comment added by 119.53.103.186 (talk) 02:28, 28 June 2011 (UTC)
 * Hammer is K8, and the article you linked is from 2001. You need to get your architectures straight. Bulldozer is, in fact, a complete redesign. — Preceding unsigned comment added by 173.166.78.9 (talk) 17:02, 14 July 2011 (UTC)


 * Comparing to say Hammer is K8, I would love to say it is the 64-bit K7 or K7x, there is only architecture(ISA) largely changed, but only slight changes for Micro Architecture, I mean the core. Comparing to the initial K8 architecture, the Bulldozer does not realize the feature to combine the two cores in one module together to give outside world a one. For this only a bit fatter thread, sacrificing frequency(energy) to increase performance, burdening the underlying OS to make itself fixes, ..., the only benefits is that the production cost might be saved. Janagewen (talk) 13:14, 3 September 2014 (UTC)


 * Even though AMD Bulldozer Microarchitecture looks so similar as it mentioned in that 2001's article, but it might have different motivation from what initial K8 possess. In essence, each module of AMD Bulldozer processor is a four instruction issued core, providing four symmetric execution units for the actual computing. It could simply be the successor of K7 design without introducing a second thread, but for the AMD64 architecture, those four independent execution units could hardly be fully utilized, and also said that the third execution unit in each core of K7 and Hammer are seldom used, making resources wasted on multi-core configurations, simply coupled. IA-32 and AMD64 after all are not EPIC architecture, programmers or assemblers could not or hardly assign their codes onto the specific actual execution units explicitly, all the jobs are hid out of ISA leaving implemented by the actual micro architecture. Sophisticated design would improve the performance of the same program onto the same execution unit configuration, but increase the number of semiconductors and the power consumption. Bulldozer simplifies this situation by dividing one thicker thread into two, leaving the new microarchitecture at front-end time-sharing, mid-end equally divided to two, and back-end restored to one. At a time, four instructions within a thread coming into the four issued decoder, decoded, buffered, and eventually executed by specific two of four execution units. Comparing with K7 and Hammer, more instructions go to the fewer number of execution units, this would slow down the performance obviously, so for balancing this situation, each pipeline stage is further refined to contain more instructions on the fly, increasing the frequency at the same semiconductor process and avoid one pipeline's stall blocking the whole processor. From this point, Bulldozer is much like a successor to K7, or evolved from it rather than a new design from the scratch, even though many details on implementations happen changed, for example exclusive cache changes to inclusive. Each module is actually a widened yesterday's core, only equipping with one FPU. But compare with AMD 10.5h processor, if only two execution units are actually utilized within a single core rather the three all, the performance of each integer thread within a module could be weighted as a core, then a module at this point could be counted as a dual-core processor. But one should aware that the resources they both share restrict their independences and might slow down the performance comparing with the traditional dual core configuration. For Intel Hyper-Threading enabled processor, even though both threads have no self-owned execution units, but they share all the actual execution resources rather than the fixed partitioned one. So from this point, each module within a Bulldozer processor could only be counted as a core with two threads, in other words threads in Bulldozer are much rougher than in Intel HT-enabled processors.  Janagewen (talk) 15:04, 14 November 2014 (UTC)

Pictures
Aren't we allowed to use some picture from AMD, for example the block scheme of Bulldozer? --92.106.117.56 (talk) 21:51, 24 November 2010 (UTC)


 * yes of course if it's public it's allowed on wikipedia (unless it's clear marketing crap) Markthemac (talk) 22:42, 7 December 2010 (UTC)

Orochi
The link for Orochi goes to an unrelated page and the disambiguation page has nothing on chip dies. This may confuse some people. —Preceding unsigned comment added by 24.155.65.76 (talk) 06:07, 7 May 2011 (UTC)

SSSE 3 support
Some actual software(e.g. Gsdx plugin SSSE 3 verison for PCSX2 with specific games support - work only on SSSE3 processor) hardcoded with SSSE3 instructions (not work without it). If AMD support it, with SSE4.1, SSE4.2 support they are be equal with Intel and all previously non-working on AMD sofware will be work! Any information about it? —Preceding unsigned comment added by 95.78.220.13 (talk) 10:55, 24 May 2011 (UTC)

8150 Clock speed correct?
I was just noticing on this page it shows the 8150 at 3.2 ghz with 4.2 turbo, and a TDP of 125w, yet with a little google searching i have found most results to show that the 8150 has a base clock of 4.2 with a 4.7 turbo, and 140w. Can anyone clarify if this is correct? here are some links

http://www.overclock.net/rumors-unconfirmed-articles/1033921-wccf-amd-bulldozer-fx-8150p-black.html

http://techdeville.com/2011/06/05/new-amd-bulldozer-fx-8150p-black-edition-details-leaked-turbo-boost-can-go-upto-4-7ghz/

http://techx64.com/leaked-details-of-amd-bulldozer-fx-8150p-black-edition-processor-with-turbo-boost-upto-4-7ghz/


 * All of these speeds are speculative. AMD has not announced any of them. I'm not sure they belong in the article. If they do, they certainly should not be stated as fact. — Preceding unsigned comment added by 173.166.78.9 (talk) 17:07, 14 July 2011 (UTC)

FX-3150 has 3 cores?
How can the FX-3150 have 3 cores? A Bulldozer processor consists of modules, and each module consists of 2 cores (as defined in this article). 8 cores=4 modules, 6 cores=3 modules, 4 cores=2 modules, 3 cores=... 1.5 modules? How is that possible? If AMD plans to disable part of a module, or if the 3 cores are indicative of something else, a footnote should be added. Otherwise, the source of this information may be erroneous.

-> It's stated wrong in the article. Clusters/Modules are mixed up with cores. One (Cluster) Module contains 2 cores in the current Bulldozer architecture. — Preceding unsigned comment added by 195.212.29.188 (talk) 11:22, 29 January 2013 (UTC)

Next Gen Bulldozer at 28nm?
I think the entry for Next Gen bulldozer being on a 28nm node is incorrect. GF's 28nm node is a bulk production node which is intended for GPUs and small APU lines (think the follow on to brazos) 28nm bulk is just going to be too leaky for a big CPU. From what I gather the next CPU production node from GF is 22nm FD-SOI. Can somebody confirm this?

"October 12 Bulldozer "FX" Release" NPV conflict.
I have just tagged the section named above with an NPV section header. While I do agree that Wikipedia should provide some information about the reception of the Bulldozer micro-architecture (and I am a bit surprised that it has taken this long to get added), I believe that the way it has been done right now has NPV issues. I wouldn't know where to begin on this issue though, so hopefully someone can take a look and see if we can cover the reception in a more neutral manner. -XJDHDR (talk) 18:49, 14 October 2011 (UTC)


 * I just had a go at making it more neutral, having read many reviews myself. Please add references if you can, I don't have the time - I just hate to see such blatant bias. - Anon


 * Well thank you and well done. The section looks a lot better now. So much so, in fact, that I'm now considering removing the NPV header. Unfortunately, it is now 10pm here so I will have to do this tomorrow. As for references, the original author of this section used this forum link as one of their references. Of course, forum posts can't be used as reliable sources but it links to a number of articles and reviews about the FX-8150 which should be adequate. I know that Anandtech also wrote a good review of the CPU here. -XJDHDR (talk) 20:21, 14 October 2011 (UTC)


 * I would suggest adgjusting the first paragraph a bit, the 8150 actually ties the 2600k at some tests (such as video transcoding), and is considerably faster than the 2500k (~20%). Other general cases the current statement is more or less the case. example using Handbrake and Espresso.  --Xeridea (talk) 03:35, 22 October 2011 (UTC)


 * I completely agree. There are some |linux benchmarks that show very good multi-thread scaling for BD. The "FX Release" section should probably be renamed to "Reception" and really the whole section should be rewritten. To me, the section is one-sided and doesn't meet NPV. In additional, Tom's Hardware has a reputation for a pro-intel bias, so I would prefer to use a more neutral news source such as HardOCP, Anandtech, or Guru3D. WinampLlama (talk) 06:42, 27 October 2011 (UTC)


 * I also agree. Many of the benchmarks used in the initial testing of the FX-8150 were either Intel optimized, single-threaded / not highly multi-threaded, synthetic benchmarks known to be biased toward Intel or benchmarks, synthetic or otherwise, that had no relevance. ( eg. the app tested is hardly ever used or the benchmark uses x87 code etc. ) Here are some more useful benchmarks : linked to by AMDzone here, a review by PCinpact ( in French ) here , an OpenBenchmarking.org thread scaling comparison here , a inpai.com.cn review with a comparison of an 8150 with an i5 at the same frequency here , and a Phoronix thread scaling review here --Jdbtwo (talk) 10:26, 17 December 2011 (UTC)

Redirect from FX series?
I tried searching for this article using "AMD FX series" but the first result was a list of FX processors (which is mysteriously lacking the Athlon 64/X2 FX CPUs). Because the FX series includes both the Athlon 64/X2 era and Bulldozer FX series era, what's the best why to include this? WinampLlama (talk) 06:51, 27 October 2011 (UTC)


 * I think it warrants a separate article. Every other series of processors from AMD and Intel have an article outlining the series as a whole (e.g. Intel Core), as well as another article giving the detailed specifications of every processor in that series (e.g. List of Intel Core i7 microprocessors), and yet another article explaining the characteristics of that line's microarchitecture(s) (e.g. Sandy Bridge, Nehalem, etc).  Thus, the "Processors" section of this article should probably be spun off into its own article entitled "AMD FX", and the List of AMD FX microprocessors should contain the Athlon 64 processors as well as the Bulldozer-based models. Digital infinity (talk) 19:51, 5 March 2012 (UTC)

Couldn't find cited info in footnote link
I couldn't find the cited information in a footnote link under "Process technology and clock frequency." The link I am referring to is "The chip operates at 0.775 to 1.425 V, achieving clock frequencies of 3 GHz or more[14]." I searched several times in the PDF document that opens at: "http://isscc.org/doc/2011/isscc2011.advanceprogrambooklet_abstracts.pdf0" when "14" is clicked on for 0.775 and 1.425 and found nothing.

I am interested in finding specifications on the normal range of Vcore voltages when using AMD's Cool and Quiet Mode. — Preceding unsigned comment added by 64.139.96.91 (talk) 22:27, 14 March 2012 (UTC)

Split the section "2nd Generation" into a new article
Like these articles: Sandy Bridge and Ivy Bridge (microarchitecture), Athought they were the same microarchitecture eventually. And I think the title of the new article should be "Piledriver (microarchitecture)", after all it had a official codename. Do you think so? Porsche 911GT2 (talk) 10:45, 11 July 2012 (UTC)


 * I don't think this section is ready to be split at the moment. Although referenced, it reads like a pile of random notes. If at least the notability of the section had been established, then I may have thought different. Op47 (talk) 21:04, 27 July 2012 (UTC)


 * I have removed the tags. If you want these articles then it would be better to create the articles from scratch rather than reuse what is written here. Op47 (talk) 14:26, 9 September 2012 (UTC)

ECC RAM support
The table in the article shows ECC memory support for all CPUs, but in the cited sources (CPU-World, Xbit-Labs) there is no mention of ECC support. — Preceding unsigned comment added by 81.203.49.172 (talk) 16:00, 31 October 2012 (UTC)
 * Unbuffered ECC Memory is supported by all AMD desktop CPUs by means of built-in memory controller. The tiny bit left is BIOS support, which is present only in selected brands - such as ASUS, Biostar, MSI, Gigabyte. Asrock does not support it. If unsure lookup in manual, on first pages in "Memory" table there should be mention of "unbuffered DDR3", but not "non-ECC unbuffered". Later case means no ECC support. If still unsure ask Level 2 or 3 Mainboard Support. ECC brings 5% less memory bandwidth, but MUCH MORE fault tolerant to bit rot and bit flip system, which makes sense for modern systems. 93.129.44.178 (talk) 13:42, 13 December 2012 (UTC)

Rewrite
The article is completely Windows-centric. It needs complete rewrite!93.129.44.178 (talk) 13:46, 13 December 2012 (UTC)

Richland APU
Creating a discussion about AMD's comparison of Richland and previous generation APU power since there might be dispute to what they are referencing. The info is from a slide of an official AMD talk given in January 2013 where they state that Richland will have 20-40% the performance of previous generation APUs. Is the previous generation they are comparing to Llano or Trinity? I say since AMD calls both Trinity and Richland 2nd gen APUs in their official Road Map slides, they are referencing Llano. This is also taking into account the leaked specs of the fastest Richland APU being only marginally faster than the fastest Trinity APU that in my estimation would not yield anywhere near even a 20% increase of performance. But I would like to hear from others if they agree or disagree with my conclusion.--Castaa (talk) 05:16, 2 March 2013 (UTC)
 * honestly, now that I tink bout it, a 20% increase from trinity seems a bit out of proportion... I still want to hear from some others tho  and thanks for bringing this here, I prefer to discuss content on the article's talk page rather than my user talk. Aunva6 (talk) 05:41, 2 March 2013 (UTC)

External links modified
Hello fellow Wikipedians,

I have just added archive links to 2 one external links on Bulldozer (microarchitecture). Please take a moment to review my edit. If necessary, add after the link to keep me from modifying it. Alternatively, you can add to keep me off the page altogether. I made the following changes:
 * Added archive https://web.archive.org/20101210224414/http://www.extremetech.com:80/article2/0,2845,2368186,00.asp to http://www.extremetech.com/article2/0,2845,2368186,00.asp
 * Added archive https://web.archive.org/20110625081444/http://blogs.amd.com:80/work/2010/08/02/what-is-bulldozer/ to http://blogs.amd.com/work/2010/08/02/what-is-bulldozer/

When you have finished reviewing my changes, please set the checked parameter below to true to let others know.

Cheers.—cyberbot II  Talk to my owner :Online 10:10, 21 January 2016 (UTC)

If the MMU in each Bulldozer Module Could Be Mentioned?
MMU plays an important role in today's computing OS, which is used to translated the linear/virtual address into the physical form addressed by the processor cache and memory controller! As is known that the single MMU within Intel Netburst Microarchitecture is shared by two threads, which could further slow down the system, if two of them are from different processes, in need of two different sets of paging tables swapping in and out. AMD Bulldozer module has two independent L1 cache for each integer core, in other words, it might be possibly that each integer core is equipped with an independent MMU component, one does not need to wait the other for accessing the memory sub-system! From this very standpoint, one module is indeed equipped with two cores rather than two threads. — Preceding unsigned comment added by 119.53.111.243 (talk) 15:24, 23 October 2019 (UTC)

FX-8120 TDP
The FX-8120 came in a 95 W and 125 W variants, and a few other chips are like that too. The chip table is a bit misleading, showing all 8120's as 125 W. Gigabyte Motherboard w 95 W 8120 Support CPU-World 95 W 8120 CPU-World 125 W 8120 SVMLegacy (talk) 21:31, 23 April 2020 (UTC)
 * List seems riddled with errors, also non-existing B3 revisions are listed. --Denniss (talk) 09:32, 24 April 2020 (UTC)