Procesory AMD 15h "Bulldozer&Bobcat"- info, testy, spekulace

[flanker]

hlavně mě budou zajímat i rozměry čipu...32nm proces, velikost pouzdra, je hypotewticky možné v případě potřeby vydat něco jako 8-pouzdro na velikosti socketu AM3??

[flanker]

Aaaaa, J.F. se už vyjádřil..." Server will be 8 cores, 8 threads and 16 cores 16 threads.

Client will be 8 cores, 8 threads."

[flanker]

[AMD Bulldozer blog

Alexpro August 10, 2010

Will it support SSE5?How about the SSE4, will it be fully compatible now? How about AES-NI?
Reply

*
John Fruehe August 10, 2010

We support SSE4 today already. We will support AVX and AES-NI. SSE5 is not being supported by anyone that I am aware of.


#

Will the inter-core latency and L3 cache be reduced sharply?
Reply

*
John Fruehe August 10, 2010

We are not disclosing any of that detail until launch, but customers should be very happy with the performance.
Reply

#

nyní citace JF na Xtremesystems
"Each interlagos processor can run either 8 256-bit executions per cycle or 16 128-bit executions per cycle.

In newer AVX-updated code, we will have the same number or executions as Intel. In older non-AVX code we will have 16 128-bit executions to their 8 128-bit executions. "

[flanker]

spuštěno hlasování

[HEAD]

Hlasoval sem pro moznost cislo 1.

[flanker]

pochopitelně je myšlen průměrný výkon včetně multithredingu (video, audio, 3D, render, archivace atd). Já osobně se domnívám, že výkon v některých apliakcích (render a video) bude skutečně velmi vysoký, jinde bude cca srovnatelný se současnou K10.5, v průměru tedy vidím výkon 8-jádrového Zambezzi někde na úrovni SB se 4 jádry a 4 virtuálními vlákny. Samozřejmě, specifický výkon bude někde i vyšší než má SB 4C/8T.

[flanker]

podívejme se na další zajímavé odpovědi, konference se nám pomalu blíží...
Hi, thanks for the blog, I heard that there was some problem (although I don’t understand the technicalities) that Bulldozer was using FMA4 while Intel was sticking with FMA3 and that the two were not compatible.

Has that been resolved? If not how will it affect Bulldozer performance assuming that most developers will stick with Intels standard simply because of their larger market share.
Reply

*
John Fruehe August 10, 2010

I think you have it wrong, we are going to write a blog about this in the future. All of the major OS vendors will be supporting all of the features of our processors, this is a manufactred issue, not a real concern.


Anant Mishra August 10, 2010

IPC increase in bulldozer architecture is very less if its only giving 50% more performance with 33% increase in core…

It totally cannot beat i7 with this kind of performance…
Reply

*
John Fruehe August 10, 2010

I can’t speak to i7 performance, but what you are looking at (IPC) is only one aspect of performance. IPC, clock speed and pipeline depth are just a few of the variables that can impact performance. Your statement is akin to saying that because one monitor is 1024×768 and another is 1920 by 1600 that one will have a clearer picture. But it really depends on the capability of the video card on each PC as well.

There are too many people trying to hold up IPC as the only determinant of performance and they will all miss the big picture.


Anant Mishra August 11, 2010

@Clock Speed.. Not getting your point, its not like BD will be hitting 4.5GHz in stock config..

I just mean to say IPC is gng to be lesser than what I expected, after-all BD was supposed to kick the livin crap out of intel!

Also would the BD be crazy costly at launch(like $1000 or more, not asking you to give a no. though) or they will cost such that normal humans could buy them too??
Reply

*
John Fruehe August 11, 2010

There is much misinformation floating around about Bulldozer.

Single threaded performance will increase significantly over the existing Magny Cours and Lisbon products today. The amount of the increase will not be disclosed until launch, but the numbers you see floating around on the web are based on faulty assumptions about the architecture. That is all I can say at this point.


David Halliday August 12, 2010

As a maker of Simulation software FP performance is the #1 thing for us. We recently just replaced our aging socket F machines with X5600 intel boxes and found a large improvement.

I would love to switch back to AMD for the next generation but its all about FP and performance on massive sparse linear system.

Also I would concur that Hyper Threading completely trashes our performance even on X5600 CPU’s and we disable in BIOS. Access to the large linear system (much bigger than cache) always flushes the data and causes massive stalls.

An interesting feature would be to add cache coloring so a cache could be reserved for instructions and some local data but not large global data such as our matrices.
Reply

*
John Fruehe August 13, 2010

The reason that you saw a big increase in FP is that you want from 64-bit (2 32-bit issues per cycle) to 128-bit (4 32-bit issues per cycle). When we went to quad core we went to 128-bit FP as well so we have had it for a while.

We will definitely have some great FP performance, keep watching the blog for more info in the future on our innovative new Flex FP.


Alfredo August 13, 2010

on the Quad Core AMD CPU there is 4 x 128 bit
Math Co-processor per Core not 4 x 32 bit also AMD CPU our very Overclock able example a 3200+ Athlon CPU Running 2.2 Ghz can be overclock to beat Pentuim 4 at 3.2 Ghz in this case the Ghz os not the Power is how the CPU runs inturnally that Counts, Becuase the AMD run 1 and 1/2 time it posted Speed and when Overclocked surpasing Intel Overall Speed!
Reply

*
John Fruehe August 13, 2010

A 128-bit FP can issue 4 32-bit instructions per cycle. 4 X 32 = 128. There is not a 128-bit instruction, but there is the ability to run multiple instructions per cycle based on the FP width.

Poper August 15, 2010

Speaking about performance. 12.8% core improvement is clock vs clock?

12core 2.3GHz Magny Cours – 100%
8module/16core 2.3GHz Interlagos – 150%
8module/16core 3.2GHz Interlagos – 210%
16module/32core 2.3GHz Interlagos – 300%


Reply

*
John Fruehe August 16, 2010

Everyone doing the “12.8% performance” estimates are wrong, REALLY wrong. Don’t listen to them.


Anant Mishra August 17, 2010

There is an 8C/16T Bulldozer and a 4C/8T BD
now check the architecture diagram released by AMD…

there is 1 decoder but 2xINT schedulers and 1 FP scheduler.. basically a module is a dual core CPU but the 2 cores have the FP common..

so what will happen is when any single threaded application is executed.. the single thread will be divided into two threads in the module both threads calculated and again joined into 1..

also the INT pipeline width is doubled in BD core that will again inc IPC..

ur not taking into account CMT dude..
Reply

*
John Fruehe August 17, 2010

No, this is wrong.

Server will be 16C/16T and 8C/8T. Client will be 8C/8T.

A single threaded application will run on one core.

[THANATOS]

To info vyzerá zaujímavo. Možno AMD znova vypustilo Fake info ako pri rade ATI 4000.
Rád by som videl funkciu kde jeden modul vie rozdeliť jedno vlákno medzi dve, ak by sa to podarilo verím, že by to malo obrovský dopad na výkon pri jedno-vláknových aplikáciách.
Zvedavý som, ako už Flanker spomínal, že verzia do socketu am3r bude osekaná aký výkonový dopad to bude mať.

[THANATOS]

John Fruehe odpovedá na otázku aký bol výkonový nárast medzi Penryn (45nm) na Westmere (32nm)

"From xeon 5500 to xeon 5600, the performance gain was ~33% (int ~24%, FP ~42%). That was for adding 50% more cores, so if you do the math on the cores to uplift ratio, you are getting ~66% scaling.

From Opteron 6100 to Bulldozer you will get ~50% more performance from 33% more cores. That is essentially ~150% scaling. "

Obviously 100% would be perfect scaling (add 1 core, get 100% more performance) in a control mode. We are getting more than perfect scaling, which says that there is more than “just more cores” that is providing the uplift.

uživateľ MA odpovedal na otázku ohladne predpokladaného nárastu výkonu v jednovláknovom režime

"Most likely single threaded workloads will fare much better, they will have all the shared resources in a module dedicated to them, besides the spacious 2MB L2 cache."

a John Fruehe odpovedal týmto komentárom

"Bing, bing, bing, we have a winner!

People making single threaded assumptions based on heavily threaded results are not considering the architecture and what it is capable of."

tieto odpovede sa dajú nájsť na tejto stránke http://blogs.amd.com/work/category/bulldozer/

Článok je The parallel Universe


Vďaka týmto odpovediam som riskol a zvolil "Výkon bude obrovský, srovnatelný s arch. Sandy Bridge do LGA2011 8c/16T", samozrejme ak AMD dodáže dodaŤ desktop CPU s 8 a viac jadrami. Beriem to tak, že aj 8 jadier bude schopné konkurovať Intelu LGA2011 tam, kde nepôjde HT správne. Vyzerá to nádejne, bude to určite zaujímavý súboj.

[yuri.cs]

Trochu poopravim
Xeon 5600 jest jadro Gulftown/Westmere-EP (32nm); Xeon 5500 ma jadro Gainestown/Nehalem-EP (45nm); Penryn (45nm) byla jeste C2D architekrura bez QPI, Turba, atd.

Je sice hezke, ze vykon Bulldozeru papirove(!) naskaluje pridavanim jader o 150% a Intel o mrzskych 66%. Tohle nam ovsem nic nerekne o realnem narustu vykonu, netusime jak srovnani probihalo...

Prechod na 32nm u Intelu pridal vykon na jadro jen minimalne, ovsem uz 45nm Nehalemy vitezi nad AMD.
Prechod u AMD od Istanbul (K10.5+) k marchitekture Bullzozer by mel prinest obrovsky narust vykonu na jadro. Tady se tudiz vylihne tech 150%, jako zaklad u techto 150ti procent je totiz pomala K10.5+.


//Hlasoval jsem pro 2. moznost: +-Bloomfield. AMD je totiz svazana nevyhodu kompileru od Intelu... i kdyz tohle ma moznost BD zmenit, kdo vi...

[THANATOS]

Vďaka za opravu
nárast výkonu AMD 150% a Intel 66% ako píšeš napísal aj nejaký uživateľ v komentári a John Fruehe to komentoval tým, že scaling a performance nie je to isté. Tú časť som nechápal ako to presne myslel . A ešte k tomu 150% nárastu písal, že to nikde nespomínal. Ak si to ešte neprezrel prezri si tie komentáre, možno si všimneš niečo, čo som ja prehliadol.

Jedna vec by ma zaujímala. Koľko modulov prirovnávame k Intelu, 2 alebo 4?
1. možnosť to nebude, až taký pesimista nie som aj 2. možnosť je málo pravdepodobná (pridané nové inštrukcie, ekvivalent turboboost), skôr 3. a 4. to bude, ale to je len môj názor.
K tomu kompiléru, bude to hrať určitú úlohu ale možno to AMD nejako obíde.

To Flanker
Chýbaš do diskusie, je nám smutno .

[yuri.cs]

Prvni Bulldozer march preview aka sesumirovany clanek od Dresdenboye je online

planet3dnow.de /24str xD/

=================

+ 150% skalovani vyjde jednoduse z: + 33% pocet jader a + 50% vykon, coz jsou hodnoty udane JF na jeho blogu. Je to ovsem zavadejici...

Ten pojem modul bych uz neresil. Jsou to proste 2 obycejna jadra, ktera si sdileji FPU, tot vse. OS vidi 1 modul jako 2 klasicka vlakna.

Porovnavani s intelem, viz JF:

SB on the client side is 2C and 4C
SB on the server side is 6C and 8C

BD on the client side is 6C and 8C
BD on the server side is 12C and 16C

Z toho vyplyva, ze HT od Intelu opravdu neberou v AMD jako "plnohodnotne jadro". Naopak jejich implementaci sdilene FPU berou jako naprosto regulerni.

//Nemuzu najit tu zpravu, ale myslim ze by se mel objevit i 10C/20T(!) SandyBridge, ktery by zrejme nemel protivnika u AMD.

[flanker]

jj, SB do LGA 2011 vypadá, že skutečně vyjde do desktopu! Ale zpět k Bulldozeru. Stále bude notně brát že 8 jader bude 8 jader (protože 1 modul AMD obsahuje 2 fyzická jádra a ty mají svou vlastní L2 cache o kterou se dělí). Každopádně, neobávám se toho, že by Zambezzi nestíhal na Nehalemy 4c/8t (to již totiž stíhá AMD Thuban). Neumím si však představit nový skok v jednovláknových operacích, každopádně me to mile potěší (jsem zastánce hlavně multithreadingu jako uživatel)

Jak vidíte možnost osadit v případě "nouze" třeba 5-6 modulů do pouzdra AM3+ (velikostně stejné jako AM3 či AM2+ a AM2), bylo by možné přidat další 2 moudly na 32nm + 4MB L3 cache? Nebo plošně je to již na nový socket? Thuban například je na 45nm téměř na limitu pouzdra.

[yuri.cs]

Jak vidíte možnost osadit v případě "nouze" třeba 5-6 modulů do pouzdra AM3+

Me hlavne nejde do hlavy pouziti kompatibilni patice. Uz Nehalemy maji 3x 64b pametove kanaly, SB na LGA 2011 ma mit dokonce 4x 64b. Do patice zalozene na AM3 se 3 a vic pametovych kanalu asi nenarve. Ze by plne postacovaly soucasne 2x 64b i pro budouci 8C, nebo dokonce 16C procesory?

[flanker]

nejspíš pro 8C ano, ono srovnání Lynfieldů a Bloomfieldů, výkonově je tam minimální rozdíl. Samozřejmě v teoretickém testu propustnosti bude Sandybridge bodovat. Ale serverové Bulldozeru řešení ho má. Každopádně AM3+ bude dualchannel. To je potvrzené. Ale domnívám se, že AM3+ bude poslední kompaktibilní paticí a další rok se dočkáme něčeho nového. SPíš mi šlo o to, zda by bylo možné narvat do pouzdra die takové velikosti, které by obsahovala 32nm 5-6 pouzder (10-12 tijádro) s L3 chache zvětšenou o 2-4 MB. Teoreticky 1 modul Bulldozeru je jako 1.5 jádra Dnebu.

[flanker]

A core is an integer execution unit.

Each bulldozer core will only execute one thread per cycle. But we will have ~2X the number of cores as the compeition. As they ty to jam two threads on one core we will be able to execute that same number of threads on dedicated cores, allowing for better scalability.

The processors will continue to have FP units tightly coupled. If you take the FPU out of the core you would add a lot of unnecessary complexity and kill the performance.

GPGPUs (GPUs used for computing) are great for large amounts of parallel FP instructions (specialized technical work) but not great for most of the standard FP commands that are coming doing in small chunks. A large truck is great when you have to move a large amount of data (saving you several serial trips back and forth) but they are not good if you just want to drive up the street to get a cup of coffee. Even with GPUs integrated into CPUs you will probably still continue to see FPUs, at least in the near term, because for small, one-off non-parallel FP tasks, it is a more efficient way to handle the traffic.

[HEAD]

Vykon ve hrach odhadujete jakej?Co to tak ctu, tak amd chce zvedat vykon hlavne zvysenim poctu jader a to je u her na nic...
Vyrovna se ve hrach na stejne frekvenci aspon dnesnim i5/7 a nebo zase jen lehce na P2???

[yuri.cs]

To ti asi zatim nikdo s jistotou nerekne. BD ma vylepsenou FPU, sdilenou L2 cache, ...

[flanker]

navíc hry nejsou prioritou PC platformy (navíc neznám hru, kde by se vyloženě současné quady nechytaly), to vůbec a každopádně upozorňuju, nečekejte cenu tzv AMDčkovskou jak s eněkteří domnívají, ne, AMD highend nové generace bude dražší než Thuban, mainstream se momentálně prý vůbec nechystá. Je to primárně znovu serverový CPU.

[ttxman]

Vykon ve hrach odhadujete jakej?Co to tak ctu, tak amd chce zvedat vykon hlavne zvysenim poctu jader a to je u her na nic...
Vyrovna se ve hrach na stejne frekvenci aspon dnesnim i5/7 a nebo zase jen lehce na P2???

No na to uz vicemene odpovedel flanker prepisem z toho amd blogu.

Single threaded performance will increase significantly over the existing Magny Cours and Lisbon products today. The amount of the increase will not be disclosed until launch, but the numbers you see floating around on the web are based on faulty assumptions about the architecture. That is all I can say at this point.

12.5% vyssi vykon v jednovlaknovym zpracovani je oznaceny jako chybny a zaroven je tam napsany, ze se vykon zvedne vyznamne. Takze bych rek, ze to bude minimalne 20%. Navic to na cem stoji vykon v novych hrach jsou v FP vypocty a podle vseho FPU jednotka je uplne prepracovana a mnohem vykonejsi.

No a nevim jak jeste porad muzes chtit porovnavat vykon frekvence/frekvence. To podle me zkoncilo Athlonama XP. Pokud je cena a spotreba priblizne stejna (plus pro nekoho jeste moznosti v pretaktovani) tak na frekvenci vubec nezalezi. Vnitrni usporadani architektur AMD a Intelu je natolik odlisny ze to proste objektivne porovnat nejde.