Rumour: Bulldozer 50% Faster than Core i7 and Phenom II.

[ShadowVVL]

I thought badaboom sucked, but it has a vary reasonable price.

how good is physX in games vs not using physX?

 

[nanaki333]

I thought badaboom sucked, but it has a vary reasonable price.

how good is physX in games vs not using physX?

the side by sides i've seen, it's nothing you're going to notice while actually playing the game. i saw VERY little difference.

 

[JFAMD]

JF how important is FLOPS to a super computer?

For some it matters, for others it does not. It is all about the workload. Things that are heavy in integer and low on floating point are not really impacted by FLOPs.

 

[Zstream]

I thought badaboom sucked, but it has a vary reasonable price.

how good is physX in games vs not using physX?

It does suck!!

 

[zod96]

When exactly is Bulldozer due out to retail stores?

 

[hamunaptra]

You have obviously not used dual cards in the last few years. I have no issue with my SLI 460's. And now that the ATI scaling issues are resolved with thr 6xxx series CF is now a viable option. And if you need more performance than a single card can provide(3d, eyefinity, 120hz display) you have no choice but to go multiple GPU.

I have and while the speed increase was there, the smoothness was pretty much non existent. And, unless it was churning out very high fps...it felt like a single card despite it saying it was putting out higher frames.
Its absolutely horrible, I can tell microstutter from a mile away. Ive had 2 SLI setups.
Theres just nothing that compares to a single card / gpu still.

 

[JFAMD]

When exactly is Bulldozer due out to retail stores?

Client products launch in Q2

 

[ShadowVVL]

Ive been meaning to ask Will BD be in 4, 6 and 8 core or just 8 and up?

 

[drizek]

4 and 8 for desktops, 12 for servers.

 

[Castiel]

I'm going to be very interested if the 4 module BD can compete with the SB dual core's.

 

[busydude]

I'm going to be very interested if the 4 module BD can compete with the SB dual core's.

If that is the case, then consider it a massive failure.

 

[Elixer]

Those hacked drivers are actually cracked drivers. nVidia drivers contain DRM that prevents SLI from working unless the BIOS contains encrypted keys which you must buy from them. All boards with two slots are capable of SLI, but all SLI "capable" boards are ones who paid nvidia their extortion money for those keys. The cracked drivers merely remove the DRM which would otherwise disable SLI.
Another thing the DRM does is disable GPU physX if an AMD GPU is detected, that includes an AMD IGP.

The only reason they have not been sued for anti competitive practices is because this is self sabotaging behavior, plus damages get to accrue so it will hurt more when AMD finally does complain to the US government.

AMD cannot prevent a board maker from paying nvidia for those keys and including them in the board, so you could potentially see SLI capable AMD boards.

Sad part about this is, if all board makers would put their foot down, they could very easily make nvidia drop their silly scheme.

 

[Vdubchaos]

I will believe when I see.

I'm not pulling a trigger on a rig for few more months anyways, so it will be worth waiting to see what not only Gaming industry offers but what AMD does as well.

Also hope SSD and 120hz monitor prices will go down/settle as well.

Patience is the key.

 

[Ajay]

I'm going to be very interested if the 2 module BD can compete with the SB dual core's (w/HT).

FIFY.

 

[podspi]

Actually, I'd be very surprised (and disappointed) if the 2 module BD couldn't beat a SB dual core in highly multi-threaded code. How it will do in single-threaded code is another, and far more interesting question.


There are a lot of people who are expecting zero (or negative) IPC gains, and zero increases in clock speed. I'm not sure why, since everything we've heard of so far is that IPC has increased, and that BD is optimized for higher clocks on a smaller manufacturing process (K10.5 was not optimized for higher clocks, and it can easily hit 4ghz). This suggests that at the least, AMD will be competitive again in multithreaded code (as it was -- very barely -- against Nehalem).

I (like many) doubt BD will hit SB-level IPC, but I think ~ Nehalem-level IPC isn't too optimistic (the design has been out for years, AMD has known the IPC level), and if they can do that + higher clock speeds, they should be able to hold their own against SB and possibly IB.

 

[Skurge]

Why does that myth persist?
(I suspect because people confuse graphics with physics):

http://www.youtube.com/watch?v=XBc3AR-Dl10

I didn't know trees were that wobbly and that branches weren't affected by gravity or that pressure waves from explosions didn't affect leaves. And crysis has a very high physics setting that does a lot better than that.

Either way. There are a lot better physics engines than physx. I couldn't care less if they ran on the GPU or the CPU as long as they perform well.

 

[taltamir]

I didn't know trees were that wobbly and that branches weren't affected by gravity or that pressure waves from explosions didn't affect leaves. And crysis has a very high physics setting that does a lot better than that.

Either way. There are a lot better physics engines than physx. I couldn't care less if they ran on the GPU or the CPU as long as they perform well.

ha, isn't that the truth. The trees wobble as if under hurricane winds. but there is no such wind or the debris it would induce, and they don't all wobble at the same direction...
then the tree get neatly cut in half at any point where he shoots 4 bullets.

 

[cusideabelincoln]

ha, isn't that the truth. The trees wobble as if under hurricane winds. but there is no such wind or the debris it would induce, and they don't all wobble at the same direction...
then the tree get neatly cut in half at any point where he shoots 4 bullets.

In Crysis the trees break where you shoot them, and if you continue to shoot them you can cut off even more pieces of the same tree.

That is just simply a horrible video showcasing the tech.

 

[JFAMD]

4 and 8 for desktops, 12 for servers.

Server will be 8, 12 and 16 core.

If you want 12-core server parts, you can get them today.

 

[Markfw]

Server will be 8, 12 and 16 core.

If you want 12-core server parts, you can get them today.

I want that 12-core, but on a motherboard with overclocking abilities !

 

[taltamir]

In Crysis the trees break where you shoot them, and if you continue to shoot them you can cut off even more pieces of the same tree.

That is just simply a horrible video showcasing the tech.

i was talking about GRAW

 

[HW2050Plus]

Actually, I'd be very surprised (and disappointed) if the 2 module BD couldn't beat a SB dual core in highly multi-threaded code. How it will do in single-threaded code is another, and far more interesting question.

There are a lot of people who are expecting zero (or negative) IPC gains, and zero increases in clock speed. I'm not sure why, since everything we've heard of so far is that IPC has increased, and that BD is optimized for higher clocks on a smaller manufacturing process (K10.5 was not optimized for higher clocks, and it can easily hit 4ghz). This suggests that at the least, AMD will be competitive again in multithreaded code (as it was -- very barely -- against Nehalem).

I (like many) doubt BD will hit SB-level IPC, but I think ~ Nehalem-level IPC isn't too optimistic (the design has been out for years, AMD has known the IPC level), and if they can do that + higher clock speeds, they should be able to hold their own against SB and possibly IB.

I do not think that single thread performance is that important. This is because if an application does not need performance there will be no more than one thread started. If it needs plenty performance then it will start many threads. Therefore I think it doesn't matter that much, except you are having quite old software you do not want to update.

Then regarding IPC it is also a bit complicated. Especially for Bulldozer, as it splits a "MegaCore" aka Module in two almost independent cores (front-/backend shared). So it is interesting if you ask for IPC for the module which is definatly higher than everything else (including Sandy Bridge and even possibly successors) on the market. Or you look at IPC for a BD core where it is lower than a Sandy Bridge/Nehalem/Conroe core if hyper threading is deactivated but of course higher if hyper threading is activated on them.

As you see the IPC question is not very suited, as you get no simple answer.

Let's take two 8-thread capable parts, Bulldozer and Sandy Bridge.
With Bulldozer you get 8 fast cores (to OS).
With Sandy Bridge you get 4 very fast cores but only if you do not have more that 4 threads and you get 4 slow cores. If you have more than 4 threads the very fast cores drop in performance.

If you run only 1 thread in the system (amongst some more really minor threads of course) than it will be definately faster on Sandy Bridge if it is integer only and likly faster on Bulldozer if it heavily uses floating point.

Now to make all this even more complex: IPC is defined by instructions per clock. As Bulldozer cores will be clocked very high (~40% higher clock than Phenom) they need longer pipelines. These longer pipelines will take more time if they stall. That both together reduces even the IPC but increases performance. Weired not?

And to make the confusion complete, Bulldozer has changed number of superscalarity and abilities. IPC for integer drops since float IPC increases (per half module). However to make it complete: The IPC/pipeline heavily increases for integer as well as for float.

So question about IPC is just wrong. It is wrong because of multi-core (a two core CPU with high IPC is beaten by a 4 core with lower IPC), because of Hyper Threading (where IPC is dependend on if it is measured at which core and under what condition) and in future because of "module technology" where it is even unclear to which IPC relates to (to core or module).

You could only resolve this if you define IPC as instruction per clock for whole CPU. But then it is meaningless either since adding two more cores increases IPC.

There is one totally dominant parameter for CPU today and even more in the future and that is number of cores.

Why?

You compare two CPUs with different clocks? Hardly to get any 10% difference.
You compare two CPUs with different "IPC"? Hardly to get any 10% difference.
You compare two CPUs with different core count? Easy to get 50%-100% difference.

With a new process technology, e.g. with the step from 45 nm to 32 nm you could get the following:
10% more clock and 10% more IPC/core or
Double the core count.

The above is just the difference between what Intel and AMD did. Intel's Sandy Bridge takes the first option and delivers only slightly performance to previous generation. AMD follows the second path and delivers ~80% of performance gain.

 

[Castiel]

FIFY.

It takes an 8 core BD to beat the 950 so i was correct with the 4 module beating a 2 core SB

 

[drizek]

Server will be 8, 12 and 16 core.

If you want 12-core server parts, you can get them today.

Not sure why I said 12, I meant 16.

Are you saying that there will be a 12 core bulldozer or are you talking about Magny Cours?

 

[Castiel]

I do not think that single thread performance is that important.

:Facepalm: