AMD Rolls Out "Warsaw" Opteron CPUs With 12 and 16 Cores

[childofthekorn]

AMD's Bulldozer and Piledriver based processors have also been proven to draw well beyond their 125w TDP..... The TDP is a rating of the amount of heat that has to be dissipated, presented in watts, and has nothing to do with power consumption at all.

FYI the TDP rating you're mentioning is heat dissipation, not power draw.

 

[Guest]

To bad the 20 core Opteron plans were scrapped after the fiasco of the Bulldozer.

 

[crisan_tiberiu]

AMD's Bulldozer and Piledriver based processors have also been proven to draw well beyond their 125w TDP..... The TDP is a rating of the amount of heat that has to be dissipated, presented in watts, and has nothing to do with power consumption at all.

FYI the TDP rating you're mentioning is heat dissipation, not power draw.

So, for a CPU to disipate 95W, the power consumtion sould be higher than the thermal TDP, otherwise a CPU could be a very efficient heating source for the home ^^. Imagine a heating radiator, you feed it with 95W and it disipates 95 ^^ 100% randament, pro 😛

 

[Guest]

Ok ...great with the new forum I can't even see if I posted. Annoying.

 

[RaduZ]

Piledriver seems to be a lot more efficient when clocked down at these speeds. Kaveri is nearly as fast as Haswell, clock for clock, but it also consumes nearly 2x the power with near 95watt draw during load with 4 "cores". These Warsaw chips have 16 cores, or 4x as many cores, and nearly the same power draw. Warsaw is clocked about 30% slower, but it more than makes up for it with 4x the units.

I could see this competing and makes me interested.

Kaveri isn't anywhere close to Haswell clock for clock. In single thread, fixed clock tests, Kaveri's Piledriver cores get a hell of a kicking from Haswell's cores.

When only one virtual core is getting used, modern OSes will sleep the other cores, which causes Intel's CPUs to disable hyperthreading, which allows all of the resources of a cpu core to be dedicated to the single thread.

When it came to the multi-core benchmarks, Kaveri was basically tied with Haswell in most benchmarks and only slightly behind in a few others. Almost any game that uses more than one virtual cpu worth of computing will benefit since HT will remain enabled, which puts Kaveri on par with Haswell.

Since we're talking about server CPUs, my guess is this is most of the time. I would like to see some benchmarks comparing the two to see if my theory is even close, like within 20%.

Hyper threading can be resumed as just being the ability of one physical core dealing with two threads at once for better usage of the pipeline. It barely have any impact in single threaded scenarios, that are where the IPC is measured.

What Haswell (and both current and older AMD and Intel as well) do to improve the single threaded performance when only one core is being used is to boost the clock using the thermal headroom. Basically, a dynamic over clock. Naturally you must know it.

Kaveri tied a dual core Haswell at multi threaded benchmarks. Even though Kaveri's Piledriver modules shares a FPU, they still have the other core physical parts. Speaking clearly, four Piledriver modules (cores) tied two Haswell cores. When benchmarked against a comparable priced Haswell quad core, Kaveri is left behind by a large margin.

That's why I said that Haswell has the upper hand in clock per clock performance. There are countless tests to prove that and it has nothing to do with hyper threading or turbo boost, since those tests are made with fixed clock and only one core (or module) is used on both processors.

-

Also, don't mistake TDP for power consumption. TDP is the measure of how much heat a processor dissipates, in watts. A TDP of 95W does not means that the processor draws just 95W from the wall.

While it's true that it's amazing to pack 16 cores under a 95W TDP, you should notice that AMD do this using the relief from the GCN inside Kaveri, that combines in a 95W TDP package four piledrive modules, 512 GCN "cores" and things such as the memory controller.

Using lower frequencies (and thus lower voltage), AMD uses the thermal room provided by the omission of the GPU to pack so many cores in a 95W TDP chip.

What's great about Warsaw is it's price. I truly don't expect to see 16 piledrivers modules at low clock beating upcoming Xeons that packs 12 higher clocked Haswell cores in most scenarios. The offering for just US$500 is what makes Warsaw interesting.

Sorry for being a natzi but Kaveri has Steamroller cores which do not share the same FPU thus increasing the single core performance. The problem is that the new 28nm fabrication process is not yet as refined as the old one and cancels out the improvement you get from the new core design.

 

[InvalidError]

Sorry for being a natzi but Kaveri has Steamroller cores which do not share the same FPU thus increasing the single core performance.

Steamroller still does share the FPU pipelines between cores on the same module. What is no longer being shared is the instruction decoders.

 

[jacobian]

Haswell is an APU, same as Sandy Bridge, Ivy Bridge, Clarkdale and Lynnfield. It is unfair to compare these processors "clock for clock" though because the architectures are completely different. Bulldozer, Piledriver and Steamroller are designed for heavily threaded environments. Sandy Bridge, Ivy Bridge and Haswell are general purpose processors that perform best on light-threaded tasks.

I don't see how you can say that the Intel Sandy Bridge and so on are designed for lighter-threaded tasks relative to AMD Piledriver considering that Intel's Core i7 gives you hardware support for 8 threads and Ivy Bridge Xeon gives you 12 cores and 24 logical threads. Yes, the execution cores are shared, but so are they in AMD's Piledriver, Steamroller, and so on. I personally like to think of AMD's core more like "threads". They're not real dedicated cores. Each Piledriver module contains two logical "cores" which share pretty much everything except for dedicated integer cores. This sounds awfully like something that's somewhere midway between Intel's hyperthreading and two real cores with dedicated hardware to all stages of processing. The type of processing where AMD can have real advantage in multi-threaded applications is where you need many integer cores, such as desktop productivity applications or servers. In this case, the Piledriver FX 8350 may actually beat Core i7 and Core i5, but then probably not by a big margin since Intel's individual cores are a bit more productive.

 

[InvalidError]

I personally like to think of AMD's core more like "threads". They're not real dedicated cores. Each Piledriver module contains two logical "cores" which share pretty much everything except for dedicated integer cores.

AMD's "integer cores" include quite a fair bit of stuff by their own right... they are pretty much a whole CPU apart from shared instruction decoders and FPU on Piledriver and Steamroller gave each int-core its own instruction decoder, bringing the cores that much closer to fully independent. Looking at most benchmarks, the shared FPU does not appear to hurt most productivity, gaming and other applications much.

In a real SMT implementation, there is almost no hard-wired resource partitioning between threads aside from context management. AMD's cores are almost entirely dedicated to a single thread - you could rip a Piledriver or Steamroller int-core out of an AMD module and make a still functional single-threaded int-only CPU by slapping on a DRAM controller and miscellaneous other components that are shared between cores/modules on full-sized CPUs.

 

[Draven35]

Yes, and it would be nearly useless without an FPU. You'd apparently be surprised how much is handled by the FPU these days.

 

[InvalidError]

Yes, and it would be nearly useless without an FPU. You'd apparently be surprised how much is handled by the FPU these days.

Most microcontrollers today do not have FPUs and that does not prevent them from getting used to run tons of mission-critical stuff. The significance of a lack of an FPU is highly dependent on what you want to do... tons of things either do not require floating point, are perfectly fine performance-wise with software emulation or can make-do with fixed-point math.

If AMD gets their wish with HSA/HUMA, most remotely intensive floating point stuff will be handled by the IGP anyway.

 

[jacobian]

I personally like to think of AMD's core more like "threads". They're not real dedicated cores. Each Piledriver module contains two logical "cores" which share pretty much everything except for dedicated integer cores.

AMD's "integer cores" include quite a fair bit of stuff by their own right... they are pretty much a whole CPU apart from shared instruction decoders and FPU on Piledriver and Steamroller gave each int-core its own instruction decoder, bringing the cores that much closer to fully independent. Looking at most benchmarks, the shared FPU does not appear to hurt most productivity, gaming and other applications much.

In a real SMT implementation, there is almost no hard-wired resource partitioning between threads aside from context management. AMD's cores are almost entirely dedicated to a single thread - you could rip a Piledriver or Steamroller int-core out of an AMD module and make a still functional single-threaded int-only CPU by slapping on a DRAM controller and miscellaneous other components that are shared between cores/modules on full-sized CPUs.

That's why I say that AMD's module is somewhere midway between a hyperthreading with half as many cores and a true full-fledged dual core. For integer stuff, AMD's cores indeed must perform like hard real cores while for floating point operations they can be a bit handicapped. It doesn't hurt AMD FX chips in gaming performance, since after all, Intel doesn't have any six or eight core desktop CPUs to compare them with. All benchmarks of AMD are usually against Intel's quad-core chips, but if the comparison was made among the chips with equal number of cores like FX-4XXX vs Core i5, then AMD would lose, although it's hard to say if this has to do with shared FPU or simply much slower instructions per cycle performance in AMD CPUs..

 

[InvalidError]

It doesn't hurt AMD FX chips in gaming performance, since after all, Intel doesn't have any six or eight core desktop CPUs to compare them with.

Intel does have 6-core CPUs in the form of the i7-39xx/49xx and you can get 8-12 cores in the Xeon lines. But you get relatively little benefit from that for gaming since most games hardly make significant use of more than two threads with only rare exceptions scaling beyond four.

Most of Intel's gaming benefits come from Intel's much stronger INT performance since you still need tons of INT/non-FP instructions to move data in/out of FP registers or re-order data before shoving it through the FPU.

 

[Draven35]

Yes, and it would be nearly useless without an FPU. You'd apparently be surprised how much is handled by the FPU these days.

Most microcontrollers today do not have FPUs and that does not prevent them from getting used to run tons of mission-critical stuff. The significance of a lack of an FPU is highly dependent on what you want to do... tons of things either do not require floating point, are perfectly fine performance-wise with software emulation or can make-do with fixed-point math.

If AMD gets their wish with HSA/HUMA, most remotely intensive floating point stuff will be handled by the IGP anyway.

Most microcontrollers? like what- an arduino? because the microcontroller used in your tablet has an FPU. Most 'mission-critical stuff' that doesn't need an FPU- doesn't need that much CPU in the first place and was taken care of for a decade and a half by, largely, the M68K series, intel's i960, etc.

 

[InvalidError]

Most microcontrollers? like what- an arduino? because the microcontroller used in your tablet has an FPU.

That is not a microcontroller... Arduino is a whole development platform based on an ARM SoC and ARM itself is a fully-fledged CPU design.

Microcontrollers are minimalistic CPUs used for all sorts of mostly self-contained applications where low parts count, small size, ultra-low-power (well below 1W... some microcontrollers use less than 0.001W) - microcontroller usually have integrated firmware flash memory, integrated EEPROM to save user data between power-ups, have only a few KB worth of internal SRAM to operate on. In many simple applications, microcontrollers do not need any support components beyond their power supply.

Your PC probably has 20+ microcontrollers embedded in its various sub-components... the motherboard alone may have 5+: the BIOS recovery system may have a microcontroller to flash the system BIOS directly from USB drive when the main BIOS has failed, the digital VRM is most likely managed by a microcontroller too, the system monitoring chip may have one, the soft-power may also be handled by another, another may be emulating a system ((E)E)PROM with secure update features to store system certificates, serial numbers, encryption keys, etc.

When I say microcontroller, I mean stuff like Atmel, Microchip, microblaze/picoblaze (Xilinx' soft-core) and countless other names you most likely have never heard of but are likely present in just about any device you buy.

 

[Draven35]

Only one Arduino model (the Due) has an ARM processor on it, the rest are based on various versions of the Altmel AVR. Which is a microcontroller.

 

[poordirtfarmer]

I’m not hung up on single threaded performance. I’m into video, and better video s/w is not only heavy on the arithmetic, but takes advantage of lots of cores & taps the GPU too. I’d love to see a review using a pair of the 6338s in a rig doing editing, rendering and encoding of an hour or two of AVCHD. The SUPERMICRO MBD-H8DG6-F-O E-ATX m/b would seem to be a nice board to try them on.Hint, hint.

 

[Draven35]

As we've said before, AMD seems to have no interest in us reviewing Opterons in a workstation configuration. That *should* tell you something.