Thursday, January 13th 2011

Bulldozer 50% Faster than Core i7 and Phenom II

Here, take some salt. AMD reportedly gave out performance figures in a presentation to its partners, performance figures seen by DonanimHaber. It is reported that an 8-core processor based on the "Bulldozer" high-performance CPU architecture is pitched by its makers to have 50% higher performance than existing processors such as the Core i7 950 (4 cores, 8 threads), and Phenom II X6 1100T (6 cores). Very little is known about the processor, including at what clock speed the processor was running at, much less what other components were driving the test machine.

Taking this information into account, the said Bulldozer based processor should synthetically even outperform Core i7 980X six-core, Intel's fastest desktop processor in the market. Built from ground-up, the Bulldozer architecture focuses on greater inter-core communication and reconfigured ALU/FPU to achieve higher instructions per clock cycle (IPC) compared to the previous generation K10.5, on which its current Phenom II series processors are based. The processor is backed by new 9-series core logic, and a new AM3+ socket. AMD is expected to unveil this platform a little later this year.Source: DonanimHaber
Add your own comment

424 Comments on Bulldozer 50% Faster than Core i7 and Phenom II

#1
laszlo
hope is true but i really don't need it..even 4 cores are enough if i upgrade
Posted on Reply
#2
WhiteLotus
This can't be true. 50% is a staggering increase on something that already delivers mind-blowing performance. I call bullshit. Real numbers between 20-25% increase.
Posted on Reply
#3
Googoo24
by: PirateBoy
Here's to seeing what it will be like on release then. :toast:
Also, to clarify, Fusion and the Intel variation of APU seem to be entirely different. Fusion is significantly broader in scope (apparently). But that's better suited for another discussion.
Posted on Reply
#4
kirtar
by: TheLaughingMan

Server 16 core (8 modules) is slated to be the first Bulldozer on a market. For now it is expected to be accompanied by a 12 core (6 module) version on release day.

Desktop Bulldozers will be released soon after that and expected to be a 4 core, 6 core, and 8 core versions with a clock around 3.4 Ghz.

While I can vouch for the server chips coming out first and core counts, the clock speed and number of chips released at first are best guess.
Actually I remember reading that the client launch was before the server launch. IIRC JF-AMD mentioned on [H] or something like that (Q2 client, Q3 server).

Honestly, I'd just ignore the "numbers" and wait until actual reviews are out. Now as long as it'll run -bigadv fairly well I'll be very happy.
Posted on Reply
#5
cdawall
where the hell are my stars
I can see faster like mussels said more cores is faster in some apps and a complete redesign means that int/fpu should be up in theory remember k7 was better than the initial k8 chips in certain apps on release this will happen when "k11" is released when comparing to k10.5

My worries
  • quadfather fiasco will fx be released on am3+ or server socket
  • 6 core chips will be 16% faster clock for clock (estimated) still not faster than i7 much less sandybridge
  • will the coldbug be back with the redesign
All that being said I want a 990fx based asus board (crosshair v extreme or m5a series 980gx) If 8 core is only faster than a 950 I see amd releasing it just over 950 prices if it beats a i7 6 core well I still don't see $1000 price tags maybe $500-600 amd still needs to bring its name out of the dirt unlike intel. Whoever said board will be expensive is oblivious to amd it appears am3+ is an all inclusive chipset 980G will still be midrange 940G low end Amd still has to sell dual and quad core chips cheap to be competitive with i3 and i5.
Posted on Reply
#6
JF-AMD
AMD Rep (Server)
OK, first off, let me start by saying that we don't comment on speculation. If people want to speculate on this, have at it. This is not an AMD article and I have no idea who this guy is.

We are in the middle of quiet period so you would never see AMD making a performance statement. I have no idea about the validity of the article because, amongst other things, I don't speak turkish.

To date the only performance statement we have made is around the server throughput of Bulldozer.

But, to address some of the comments.

by: HalfAHertz
No it is an octa core - every core in the module has a dedicated SP but they share a "fat" FP which can either do 1 FP calc for each or an advanced 256-bit calc (which we probably won't se for a while because nobody will have ready code...)
Actually every core has its own dedicated FMAC for floating point. If you want to do 256-bit floating point with AVX you can merge the 2 FMACs. Intel handles 256-bit AVX by merging their 128-bit FPU with the SSE functions (and you have to recompile all of your code to make SSE into AVX-128.

by: DriedFrogPills
I think it's more likely to be scenario 2 as I remember an article on anandtech midway through last year, that stated each core is one half of a bulldozer module. What we really need is for JF-AMD to clarify the module versus core thing
Modules are an architectural facet of the design, we will not market modules. The cores are cores. They are not "half cores" as some suggest. If you look at what defines a core, it is the integer execution pipeline. When the system boots up it will see all of the cores as integer cores, the OS will see them and the application will see them. All arguments have fallen flat on their face.

by: Googoo24
Uh..What? The CPU' are rumored to come out the 2nd and 3rd quarter of this year.
AMD has already said in public Q2 for client, Q3 for server, so you don't need to say rumored.

by: Googoo24
You guys are also forgetting there is supposed to be a 12 core Bulldozer as well.
Yes, for server. Server will have 8, 12 and 16-core models.

Client will have 8-core and something below that (I don't know the models, I am in server.)

by: TheLaughingMan
Server 16 core (8 modules) is slated to be the first Bulldozer on a market. For now it is expected to be accompanied by a 12 core (6 module) version on release day.

Desktop Bulldozers will be released soon after that and expected to be a 4 core, 6 core, and 8 core versions with a clock around 3.4 Ghz.

While I can vouch for the server chips coming out first and core counts, the clock speed and number of chips released at first are best guess.
Not sure how you can vouch for the server chips coming out first because I know the launch dates and server chips launch in Q3 with client launching in Q2.
Posted on Reply
#7
cdawall
where the hell are my stars
by: JF-AMD
OK, first off, let me start by saying that we don't comment on speculation. If people want to speculate on this, have at it. This is not an AMD article and I have no idea who this guy is.

We are in the middle of quiet period so you would never see AMD making a performance statement. I have no idea about the validity of the article because, amongst other things, I don't speak turkish.

To date the only performance statement we have made is around the server throughput of Bulldozer.

But, to address some of the comments.



Actually every core has its own dedicated FMAC for floating point. If you want to do 256-bit floating point with AVX you can merge the 2 FMACs. Intel handles 256-bit AVX by merging their 128-bit FPU with the SSE functions (and you have to recompile all of your code to make SSE into AVX-128.



Modules are an architectural facet of the design, we will not market modules. The cores are cores. They are not "half cores" as some suggest. If you look at what defines a core, it is the integer execution pipeline. When the system boots up it will see all of the cores as integer cores, the OS will see them and the application will see them. All arguments have fallen flat on their face.



AMD has already said in public Q2 for client, Q3 for server, so you don't need to say rumored.



Yes, for server. Server will have 8, 12 and 16-core models.

Client will have 8-core and something below that (I don't know the models, I am in server.)



Not sure how you can vouch for the server chips coming out first because I know the launch dates and server chips launch in Q3 with client launching in Q2.
Server 12 and 16 core chips will be multidie in a single package like current 8 and 12 core chips correct? Ala the q6600
Posted on Reply
#8
JF-AMD
AMD Rep (Server)
correct
Posted on Reply
#9
cdawall
where the hell are my stars
by: JF-AMD
correct
Connected via the ht link or internal cache or what?
Posted on Reply
#10
JF-AMD
AMD Rep (Server)
via HT
Posted on Reply
#11
bear jesus
by: JF-AMD
via HT
I don't know if you can say or if it really will make a difference to performance but will there be a speed increase for the inter chip HT link over the current generation?
Posted on Reply
#12
cdawall
where the hell are my stars
by: JF-AMD
via HT
How many ht links will each chip have will there still be a quad chip ready like the old 8 series and dual chip readdy 2 series and any plans of an am3+ 1 series right now?
Posted on Reply
#13
JF-AMD
AMD Rep (Server)
by: bear jesus
I don't know if you can say or if it really will make a difference to performance but will there be a speed increase for the inter chip HT link over the current generation?
Not saying at this point.

by: cdawall
How many ht links will each chip have will there still be a quad chip ready like the old 8 series and dual chip readdy 2 series and any plans of an am3+ 1 series right now?
Each processor has 4 HT links. The procesoors are capable of being unsed in 1P, 2P or 4P configurations, all with the same processor. There is no longer a seperate 4P only SKU. As such, they are all priced the same - which customers really love.

As for 1P, most of those platforms will be C32, we will not share infrastructure with client.
Posted on Reply
#14
wahdangun
by: JF-AMD
Not saying at this point.



Each processor has 4 HT links. The procesoors are capable of being unsed in 1P, 2P or 4P configurations, all with the same processor. There is no longer a seperate 4P only SKU. As such, they are all priced the same - which customers really love.

As for 1P, most of those platforms will be C32, we will not share infrastructure with client.
wow if that was true then AMD really have a winner on this segment. btw why they release desktop first ? usually AMD release server first and then a desktop part a couple of month later ??
Posted on Reply
#15
DanTheMan
I think AMD realizes that they better start pulling all the stops out for the desktop models to compete on a somewhat scale or they will lose all respect and hardcore fan base. I love AMD for performance / price but the more I look at SB and i7 the more I like. And considering I want my new machine this year I will be looking at the reviews really close. AMD needs to get it right this time and from the looks of it - it may not steal Intel's thunder but it will make a loud boom - just enough to keep the CPU market competive and open for business a while longer. All in all it's a win win for the consumer, even if BD does not hit the highest benchmarks it will keep the prices from hitting a high that we will never return from. Maybe from the looks of it it might be Q2 - new CPU and Q3 - new 7000 GPU's - sweet!
Posted on Reply
#16
cdawall
where the hell are my stars
by: JF-AMD
Not saying at this point.



Each processor has 4 HT links. The procesoors are capable of being unsed in 1P, 2P or 4P configurations, all with the same processor. There is no longer a seperate 4P only SKU. As such, they are all priced the same - which customers really love.

As for 1P, most of those platforms will be C32, we will not share infrastructure with client.
last questions in the 12/16 core chips will the ht betweem core have to go thru the chipset or is it all internal.

And the dual dual channel or quad channel memory bus the rumor mill is playing with is that a pair of 128bit mem controllers with the ability to run in ganged/unganged mode similar to the current chips? And will each die of a 12/16 be able to access its own individual memory ie one of the dual channels of a quad channel setup
Posted on Reply
#17
JF-AMD
AMD Rep (Server)
by: wahdangun
wow if that was true then AMD really have a winner on this segment. btw why they release desktop first ? usually AMD release server first and then a desktop part a couple of month later ??
There is no set schedule of who goes first, it alternates based on a bunch of different factors. Desktop got the lead slot this time around.
Posted on Reply
#18
mechtech
Does this mean office will open 1.5 times faster than with my 955BE? Or would an SSD be better over my mech hdd?? ;)
Posted on Reply
#19
JF-AMD
AMD Rep (Server)
by: cdawall
last questions in the 12/16 core chips will the ht betweem core have to go thru the chipset or is it all internal.

And the dual dual channel or quad channel memory bus the rumor mill is playing with is that a pair of 128bit mem controllers with the ability to run in ganged/unganged mode similar to the current chips? And will each die of a 12/16 be able to access its own individual memory ie one of the dual channels of a quad channel setup
All HT happens between processors, not through the chipset.

Quad channel on servers. The 128-bit memory controllers can be unganged to allow simultaneous read/write on the different channels.

The memory channel structure is identical to what we do on the AMD Opteron 6100 series today.
Posted on Reply
#20
wahdangun
by: cdawall
last questions in the 12/16 core chips will the ht betweem core have to go thru the chipset or is it all internal.

And the dual dual channel or quad channel memory bus the rumor mill is playing with is that a pair of 128bit mem controllers with the ability to run in ganged/unganged mode similar to the current chips? And will each die of a 12/16 be able to access its own individual memory ie one of the dual channels of a quad channel setup
it will be dual channel ram but with higher frequency than core i7 so it will have higher bandwidth

i think the HT was internal on CPU
Posted on Reply
#21
kirtar
by: wahdangun
it will be dual channel ram but with higher frequency than core i7 so it will have higher bandwidth

i think the HT was internal on CPU
You're mixing client and server information
Posted on Reply
#22
Magikherbs
by: Yukikaze
This really depends on what they call a core:
1. A real core.
2. One half of their SMT arrangement.

In the case of 1, I agree. In the case of 2, their octa-core processor is not a "true" octa-core. According to what I know about bulldozer every pair of cores is a hybrid between Intel's SMT approach (HyperThreading) and a true pair of separate cores. It is getting hard to define this architecture by the number of cores in the way previous generations could be, but on strict terms, this is a 4-core processor with AMD's flavor of SMT.

In case 1, we're talking about nothing special. In the case of 2, we're talking about some serious processing power.
Is this what you mean ?

http://www.youtube.com/watch?v=VIs1CxuUrpc
Posted on Reply
#23
JF-AMD
AMD Rep (Server)
by: Yukikaze
This really depends on what they call a core:
1. A real core.
2. One half of their SMT arrangement.

In the case of 1, I agree. In the case of 2, their octa-core processor is not a "true" octa-core. According to what I know about bulldozer every pair of cores is a hybrid between Intel's SMT approach (HyperThreading) and a true pair of separate cores. It is getting hard to define this architecture by the number of cores in the way previous generations could be, but on strict terms, this is a 4-core processor with AMD's flavor of SMT.

In case 1, we're talking about nothing special. In the case of 2, we're talking about some serious processing power.
Can you please give us a complete description of why a bulldozer core is not a "true" core? Please be thorough so that I only have to answer this once.
Posted on Reply
#24
PirateBoy
Gathered from Xtremesystems -

"AMD chose to go with performance per watt over performance per clock, which is perfectly fine, but it does mean that I don't think we'll see such a big gain.

There's a lot of speculation and wishful thinking going around. Personally I hope BD earns its FX title and isn't just another chip they threw more cores on.

It all comes down to perception, and that can VERY easily be controlled by carefully chosen benchmarks. Thuban wins enough tests that a full review could determine AMD wins 100% of the tests and blows Intel's much more expensive processors out of the water. Something has to change, it's not good from an enthusiast perspective. Maybe it's great from a shareholder or employee perspective though.

The marketing of "FX" is obviously working, folks are already sold on the FX moniker alone without knowing much about the product performance because of past associations.

My understanding has been that BE AMD chips are like the 'K' series of Intel chips and the FX's were the Extreme Edition equivalents. I believe all parties should leave all their chips unlocked.

Typically when companies are silent, it means their product isn't living up to par. Even with the most optimistic projections, its already well understood that bulldozer won't match up with SB or necessarily be even close in terms of IPC. AMD seems to have gone with a high throughput design, which is fine, but that means you have to have realistic expectations.

K10 and 10.5 derivatives have been around since June 23, 1999.

They don't compare it to the high end i7s. Saying 50% over i7 is just wrong."

I want my cores so I can snap em' off and eat em lol
Posted on Reply
#25
Googoo24
by: PirateBoy
Snip
Yes, you think it will be a big flop. We understand.
Posted on Reply
Add your own comment