Official: AMD re-introduces FX Brand for high-end Processors

[dorion]

So it seems the Modules allow separate multipliers, identical voltages. That's good, I wish my Phenom had separate power planes.

It seems like the power being supplied to a module and the core's clock speed might not drop down at the same time.

The process by which the cores idle seems to be twitchy, only small bumps are causing the cores to rocket up to 3.6GHz(which is good), and the time that the core stays in a higher clock looks to be substantial(which is also good).

I'm making these assumptions from the two graphs and comparing Bulldozer to my Phenom 965.

 

[podspi]

I think voltage is done on a per module basis, NOT per core, so that is why we are seeing some idling cores @ 1.4V. Also, at least on my Thuban, cores are constantly jumping up even if they don't hit 100%. I think the reason for this is not that they aren't hitting 100%, but because of the way CPU utilization is calculated -- they're hitting 100% for fractions of a second but the logging system just isn't quick enough.

Also, IIRC the temperature in Overdrive is the average. Can anyone check this? (I will later if not).


I don't have a more modern Intel CPU to test on, but my C2D increases its clockspeed WELL before it even approaches 25% utilization. I don't think cores above idle are a problem, since the screenshots pretty clearly show SOME utilization is going on. If it was done any other way we'd get the 'turbo' lag like we do in automobiles Who wants that?

 

[drizek]

So it seems the Modules allow separate multipliers, identical voltages. That's good, I wish my Phenom had separate power planes.

It seems like the power being supplied to a module and the core's clock speed might not drop down at the same time.

The process by which the cores idle seems to be twitchy, only small bumps are causing the cores to rocket up to 3.6GHz(which is good), and the time that the core stays in a higher clock looks to be substantial(which is also good).

I'm making these assumptions from the two graphs and comparing Bulldozer to my Phenom 965.

The other interpretation is that 3.6GHz is actually just a low, intermediate multiplier, and they will only reach the max multiplier of 5.0GHz when the processor is at 100% load.

 

[Idontcare]

Is it better now?

Not really, I still have more questions than answers after seeing these screenshots.

 

[drizek]

#1 Question: When are you going to fire the guy who prepared this presentation?

I am running iOS 5 on my iPad right now, and it is pretty bad. Horribly buggy, it locks up all the time, dozens of buttons that do absolutely nothing. Many of the new functions require a voodoo dance of button mashing in exactly the right order to get them to work, etc.

But you sure as hell wouldn't have known that watching the presentation yesterday.

Apple still has 4 months to work out the bugs (and they will), AMD was supposed to be shipping this thing any day now. It's just embarrassing.

 

[dorion]

Also, IIRC the temperature in Overdrive is the average. Can anyone check this? (I will later if not).

Could be AMD still hasn't given us separate thermometers for each core/module.

The other interpretation is that 3.6GHz is actually just a low, intermediate multiplier, and they will only reach the max multiplier of 5.0GHz when the processor is at 100% load.

We can dream.

 

[RobertPters77]

I'm excited. Can't wait to see what BD brings to the table.

 

[radaja]

This. Launch. Is. Worthless. Without. BENCHMARKS! (well, not worthless, but incredibly frustrating! I AM HYPED AND WANT TO SEE PERFORMANCE)

This thread is useless without benchmarks :biggrin:

as for the fx coming back. its a good thing! it makes it easier for the non-tech people to identify what is the fastest procs amd has.

well theres Truecypt 7.0,dont know why they let this slip?

 

[podspi]

That screenshot really doesn't look that strange. Here is a picture from my Thuban:





I swear, if I close everything they all go down to 972mhz, and the voltage hits 1.2V.

Edit: BTW, I've done benchmarks with TC on and off, and the scaling is remarkably close to the 17% increase in clock speed I get from TC (3.4ghz --> 4ghz) in singlethread benchmarks. Even though it doesn't look like it is working correctly, it definitely is (I didn't believe it either at first).

 

[Black96ws6]

More details slipping out from E3. Looks like there's a few things listed (probably marketing speak) that BD has that SB doesn't. Kind of hard to make out though and whether it's a real benefit or not?:

 

[Black96ws6]

Another:

 

[notty22]

I thought AMD was coining a new sniglet 'mega-tasking' , I googled it , it turns out they used that term before.
http://www.hexus.net/content/item.php?item=7505

AMD has coined a new term to describe this - megatasking. In this article we put that claim to the test, pitting Quad FX against Core 2 Quad and the dual-core competition.

Recently saw a new Sharp TV commercial, LOL at the new term they were using Viewmongous !

 

[podspi]

I feel like a CPU is an odd thing to release at E3. I've heard rumors that BD has some special gaming sauce, but I can't imagine what it could be.

 

[Khato]

It shows each integer core can operate at different frequencies.. even if they are from the same module.

Which leads to the more important observation of... what frequency is the FPU operating at? Heh, that could well explain some of the wildly different claims with respect to the delay being due to frequency issues... The integer cores could be doing great, while the FPU can't clock anywhere near high enough.

It certainly implies an interesting design decision, as well as far more design headaches... Especially if they allowed for the possibility of the FPU and integer cores all running at different frequencies. Ick.

 

[notty22]

I feel like a CPU is an odd thing to release at E3. I've heard rumors that BD has some special gaming sauce, but I can't imagine what it could be.

Well AMD likes to promote the whole package to the gamer. Like the "dragon".
The 990fx offers 40 pci-e lane bandwith, so 3-way SLI/crossfire might /could bench higher than a current SB system , with or without nf200 bridge chip, by virtue of more potential bandwidth. Lot of maybes, but hopefully will lead to plenty of good benchmark reviews !

edit: http://sites.amd.com/us/promo/processors/Pages/fx-processor.aspx

 

[PCboy]

I'm going to buy it no matter what. AMD needs to get back in the game. :thumbsup:

 

[Mopetar]

More details slipping out from E3. Looks like there's a few things listed (probably marketing speak) that BD has that SB doesn't. Kind of hard to make out though and whether it's a real benefit or not?:

It's just things we've known about:

Execute Two 128-bit AVX in 1 cycle
Execute 128-bit AVX and SEE in 1 cycle
Execute FMA operations (A = B + C * D)
Supports XOP

Sure it's a benefit if you want to use those instruction sets or have a workload with a lot of 128-bit floating point calculations.

 

[Idontcare]

Apple still has 4 months to work out the bugs (and they will), AMD was supposed to be shipping this thing any day now. It's just embarrassing.

This is what I am surprised by as well.

The more I see of BD the more I come to the conclusion AMD was nowhere close to being on track to launch the thing today.

I think they made the right call, sure it would have been better to have the chip out sooner instead of later but if the platform infrastructure is too beta'ish then both they and their customers benefit by delaying the launch so they can tend to the details that need tending to.

I bet the mobo BIOS writers breathed a sigh of relief as well, getting another 2 months reprieve to get the chipset and cpu bios parameters all the more bug-free.

 

[bunnyfubbles]

I feel like a CPU is an odd thing to release at E3. I've heard rumors that BD has some special gaming sauce, but I can't imagine what it could be.

well that's what AMD would want in order to win over the ever important game enthusiasts, it certainly helped them during the Athlon XP and Athlon 64 days, but I think Bulldozer is going to be more of a workhorse, I really don't see it challenging the i5 2500K for the top spot of ideal gaming CPU unless AMD has some absolutely amazing trick up their sleeve (and they really haven't given us a reason to believe so)

I think Fusion has a better chance of making an impact for gaming, albeit only for the most budget oriented builds where CPU power isn't as relevant anyway, something even Llano's older CPU cores will be more than "good enough" for.

off the top of my head if AMD really wanted to try and push Bulldozer as an enthusiast grade gaming product I'd wager they'd need to tie it to some sort of app or feature like recording/encoding gameplay video on the fly (something like FRAPS) and then editing/encoding for posting to video sharing sites.

Then again, we're still pretty much completely stuck in the land of conjecture with what little we know about the architecture and how its performance will pan out.

 

[Vesku]

I bet the mobo BIOS writers breathed a sigh of relief as well, getting another 2 months reprieve to get the chipset and cpu bios parameters all the more bug-free.

That's what I'm starting to think was the major factor for delay, platform issues. There were some news items 2-3 months ago about chipset/BIOS problems. Seems they need more time to make more motherboards launch ready.

 

[Riek]

did people here forget that programs are not sheduled at the same cpu continiously? that you have thread hopping over the cores? That idle cores cannot be seen by the naked eye on a multithreaded OS? Cores are turboboosted and shut down continously. The variation in frequency is probably due to that.
i'm not even sure if it is possible to read out lower idle state voltages if there is some load hopping through the cores.. wouldn't be suprised if the vcore measuring is fast enough for that..

the only meaningfull thing you can get from this is that turbo and idle is working, that the current vcorem is 1.4V and that turbo isn't limited to module level (otherwise you would see 4cores close to eachother in frequency and not 3).

 

[Madcatatlas]

So the big thing is an announcement of a brand name?

Sounds like a nV event. Haha.

Haha! well put!

 

[Drakula]

More details slipping out from E3. Looks like there's a few things listed (probably marketing speak) that BD has that SB doesn't. Kind of hard to make out though and whether it's a real benefit or not?:

64 FLOPS per cycle (128-bit FP) for AMD vs 32 FLOPS per cycle for Intel
64 FLOPS per cycle (128-bit AVX) for AMD vs 32 FLOPS per cycle for Intel

From E3:

So it is using the revised power plane thing from the original Phenom? Or did Phenom II had that, I lost track of it.

 

[CPUarchitect]

"native eight-core processor"

It's interesting how they insist on calling it an actual 8-core processor, even though each pair of cores share a front-end and floating-point cluster, and has fewer integer ALUs than before. There's little reason to assume this is any better than Hyper-Threading (sharing everything). Peak sustainable throughput per cluster is the same as one Intel core. Effective performance will depend highly on the rest of the architecture though, which may or may not help justify calling it an 8-core. Still, I doubt they can exceed the ~30% performance advantage of Hyper-Threading, unless single-threaded performance is actually compromised to begin with.

Separate L1 data caches per core could be an interesting trade-off. It doesn't suffer from port contention or bank conflicts like Sandy Bridge.

 

[podspi]

It's interesting how they insist on calling it an actual 8-core processor, even though each pair of cores share a front-end and floating-point cluster, and has fewer integer ALUs than before. There's little reason to assume this is any better than Hyper-Threading (sharing everything). Peak sustainable throughput per cluster is the same as one Intel core. Effective performance will depend highly on the rest of the architecture though, which may or may not help justify calling it an 8-core. Still, I doubt they can exceed the ~30% performance advantage of Hyper-Threading, unless single-threaded performance is actually compromised to begin with.

Separate L1 data caches per core could be an interesting trade-off. It doesn't suffer from port contention or bank conflicts like Sandy Bridge.

Peak sustainable throughput may be the same -- but I believe the idea is that average throughput is going to be much higher.

Of course we don't know the performance of BD yet -- but AMD claims an 80% increase in throughput for two cores vs a hypothetical single-core BD.