The facts about AMD’s ACP power rating
One issue that I’m frequently asked about is AMD’s alternative energy efficiency rating called Average CPU Power (ACP). AMD created this new standard on its own last year and despite controversy, they’ve managed to get the media to accept this new definition of energy efficiency. AMD has told the press that AMD’s ACP rating is equivalent to Intel’s TDP rating and I know this because this is what AMD spokesperson John Taylor told me when I was Editor at Large at ZDNet. Here’s an excerpt of what he emailed me on September 10th 2007:
“We believe that for the datacenter customer, AMD ACP is a more useful metric when configuring/budgeting. As Intel TDP is the only available metric for Intel, it is the most comparable to AMD ACP. The trick is: how does intel formulate its TDP? My sense is it is defined very closely to AMD ACP, which reinforces a “yes” answer to your question. But i’ d’ prefer Intel confirm that.”
Well I’ve not only spoke to Intel which emphatically denies this, I know for a fact that this comparison is not correct and I’ve also gotten David Kanter (one of the leading microprocessor analysts in the world) to corroborate this. I asked David Kanter to review AMD’s claim that Intel’s TDP rating was “most comparable” to AMD ACP and he explained:
“It’s pretty hard to justify the comparison between the two from a technical perspective. AMD’s ACP is defined in a very different way from Intel’s TDP, according to my understanding.”
We can easily verify that AMD’s ACP rating is not comparable to Intel’s TDP rating by looking at the actual performance of the newest AMD Shanghai based servers versus Intel’s current servers. When we look at the official published SPECpower measurements between an AMD “Shanghai” 2384 2.7 GHz system with an ACP rating of 75 watts and an Intel L5430 with a TDP rating of 50 watts with comparable components in the rest of the server, we would expect a power difference of roughly 50 watts (25W per processor) if AMD’s claim that AMD’s ACP was most comparable to Intel’s TDP rating. But according the official SPECpower benchmarks which is optimized for low power consumption, the Intel L5430 server peaks at 161 watts while the AMD 2384 based server peaks at 264 watts.
That’s more than a 100 watt delta and when we account for the fact that the Intel server has an additional North Bridge memory controller to deal with, the actual difference between the CPUs is even greater than 100 watts. We can negate the fact that the AMD server has two more memory DIMMs which consume an additional 8.4 watts of power because AMD uses hard drives that use 6 watts less power than the Intel system. This strongly suggests that an AMD TDP rating of 95 watts is far more likely to explain the 100 watt more power consumption than the Intel system with 50 watt TDP processors, so calling the “Shanghai” 2384 processor a 75 watt part simply doesn’t reflect the actual efficiency of the chip.
Based on what the experts say and what the evidence suggests, AMD claiming that their ACP rating is comparable to Intel’s TDP rating is simply incorrect and the ACP rating is effectively overstating the energy efficiency AMD processors. This is the equivalent of a car company that came up with its own Miles Per Gallon (MPG) fuel efficiency rating to inflate its actual MPG rating.
Now to be clear, TDP was never meant to be an energy efficiency rating but it is in effect exactly that when it’s used in the context of advertising and press releases. TDP is actually an engineering metric for server manufactures to figure out what kind of cooling they have to design in to their system to accommodate a CPU. A far more appropriate measure that people should be looking at for servers is the SPECpower rating. But if the media insists on quoting CPU wattage ratings, they should use a consistent measurement and the only one that is comparable and accepted by the entire industry is TDP.
i have an idea just get two heat sinks that are identicle with the same case and fans and stuff
and overclock both of the processors till they overheat and fry. use a laser temp gague to see what the temp was when the CPU fries
B&C is crazy, please don’t listen to him.
Great article George!
I think you should decide wether it is possible to compare actual power consumption with TDP/ACP ratings or not. You can’t use measured power consumption to verify that the two metrics aren’t comparable and afterwards state that they have nothing to do with actual power consumption anyway.
The reason for AMD inventing ACP is quite obvious by the way: Intel does not have to account for any northbridge power dissipation in their CPU TDP, since they have that functionality in a separate chip. That has of course been a disadvantage for AMD when the media compared TDP figures.
North bridge or not, you have not explained why an AMD Shanghai system with supposedly 75 watt CPUs and no north bridge are consuming 100 watts more than an Intel Harpertown system with 50 watt TDP CPUs and an additional north bridge to deal with.
The point is that we ideally shouldn’t be using TDPs to measure efficiency; we should be using something like SPECpower. But if people insist on comparing CPU wattage ratings, then they should be consistent and use something that is universally accepted like TDP.
Why so complicated? Why did you use processors with such different TDP/ACP figures when it was your intention to verify wether the two metrics are comparable? This also leaves you with the problem of a significant gap between peak ssj_ops values and overall system load in the end. There are numerous entries for Intel’s E54xx series, which are 80W TDP parts and achieve about the same maximum ssj_ops throughput as AMD’s system. Interestingly, those results would suggest that a 75W ACP is comparable to an 80W TDP.
We’ll see how peak power consumption will compare to Intel’s L54xx (50W TDP) parts when there will be results for "Shanghai" HE Opterons (55W ACP).
"Why so complicated? Why did you use processors with such different TDP/ACP figures when it was your intention to verify wether the two metrics are comparable?"
Are you playing ignorant or do you really not know the answer to that question? AMD doesn’t produce a chip with comparable ACP so there’s no way to make that comparison.
It’s a simple concept really. If AMD is claiming a 50W difference on two Shanghai processors claiming that their power rating is comparable to Intel’s CPU power rating, then we better see a less than 50W gap in power consumption due to Intel’s extra North Bridge controller on the motherboard. But if we see a 100 watt delta, then AMD’s claim that the power ratings are comparable are demonstrably false.
"Are you playing ignorant or do you really not know the answer to that question? AMD doesn’t produce a chip with comparable ACP so there’s no way to make that comparison."
The Opteron 2384’s ACP is 75W, while all Intel E54xx Harpertown quad-cores have a TDP of 80W. There are numerous results for systems using this CPU. Your "best vs. best" comparison policy can’t be applicable here, since you’re only using the benchmark disclosure to extract peak power consumption.
Simple math doesn’t work here, since you have different chipsets, memory configurations, harddrives, power supplies and overall instruction throughput. All you have is the total power consumption and the rest is guesswork. Using systems which achieve the same throughput and have similar ACP/TDP ratings eliminates at least some major unknowns from the equasion.
"North bridge or not, you have not explained why an AMD Shanghai system with supposedly 75 watt CPUs and no north bridge are consuming 100 watts more than an Intel Harpertown system with 50 watt TDP CPUs and an additional north bridge to deal with."
It’s not so surprising when you think about it: The Intel setup you chose for reference is tuned for lowest power consumption and not for high throughput. You can easily see that when comparing it to other submissions using the same CPU (most of them unfortunately use FBDIMM, but that can’t be responsible for a difference of up to 80W).
If you consider that there is only a single submission which attempts to achieve the highest possible overall score for a Shanghai system, using the current top performer (since there are no HE-parts available yet), it can’t be very surprising to find a few results within the numerous Harpertown submissions, which beat the Opteron setup in peak power consumption by a large margin. Lowest peak power consumption is after all not the immediate goal for this benchmark and I have hope you’ll understand what that means for your "best vs. best" comparison policy…
First of all, all those setups, AMD or Intel, are designed for highest performance per watt.
Second,
http://www.spec.org/power_ssj2008/results/res2008q1/power_ssj2008-20080212-00036.html
E5440 with 80W TDP still takes less power at 254 watts. The link above shows a system with the identical 4 DDR2 DIMMs and it’s 10 watt lower than the AMD Shanghai system with 75W ACP. If ACP was comparable to TDP, then the Intel system should at least be 30 watt higher than the AMD system when you factor in the North Bridge, yet it comes in 10 watts lower.
"If AMD is claiming a 50W difference on two Shanghai processors claiming that their power rating is comparable to Intel’s CPU power rating, then we better see a less than 50W gap in power consumption due to Intel’s extra North Bridge controller on the motherboard."
That’s flawed logic. If AMD would try to account for Intel’s extra northbridge for ACP calculation, it would be stupid to reapply that delta again to the expected values.
In fact I could even come up with a similar set of calculations and assumptions about the way ACP is meant to be applied which would prove the comparison to be valid, even with your choice of results. It is after all an increase of 63% in peak power consumption for a setup using CPUs of 50% higher ACP and stressing non-CPU components with a 20% higher load (peak throughput). But again, I think the choice of reference setups is the real problem here.
I declare George the winner against Thorn by TKO in the 4th round.
I’ll say this as simply as possible and end it right here.
- AMD claims that Intel’s current chipsets consumes more power than AMD-based chipsets.
- AMD claims that Intel’s 80W-TDP CPU consumes more power than AMD’s 75W-ACP CPUs.
Yet when it’s all said and done, Intel’s overall system power consumption is lower than AMD despite having a CPU rated at higher power and despite having a chipset rated at higher power.
Since it’s an accepted fact that Intel’s current chipsets (which include a memory controller) do consume more power than AMD chipsets, the only thing that explains this system-level power discrepancy is that AMD’s CPU power rating "ACP" is grossly inflated.
Welcome back, George!
Thanks George, you have a very valid point here. I always thought that ACP metric was a clever move on AMD’s part to make their parts look more power efficient.
That’s what AMD has done since Core 2 released. I still remember AMD executives claimed that Barcelna would beat Kentsfield by 40%, and simulated 2.6Ghz CPU delayed for one whole year. They just reduce the CPU frequency and claim they have a power efficient CPU while Intel never do that (we all know E2160 eats only 19W while TDP is 65W). Seems AMD is still on the wrong way.
Well, it is simple.
AMD had lost power advantage since the Harpertowns came out. To compensate, the AMD realized that its memory controller has room to fudge numbers. The memory controller on the Opterons can control up to 16 ranks of memory per Socket. For a 2S system, you could load up to 16 DIMMs of Dual rank memory. However, you don’t have to fully load up the memory ranks. You can easily load 2 single ranks on each processor, effectively using only 4 ranks total vs 32 total ranks that’s supported. That gives you the 20W difference between AMD’s ACP(75W) and AMD’s TDP(95W)
I think that comparing a Nforce 3600 chipset based Shanghai system to a 5100 Intel San Clemente is unfair for AMD because the Intel system only gives you 12 ranks of total memory. (when 12 ranks are used, 5100 converts into 48GB mode and uses a little more power than if 8 ranks are used in normal mode) Most san clemente systems on the spec2008 power benchmark site use 4 dimms of dual rank 4GB ram to minimize power usage(by the way, those are 2Gbit chips, so power is lower too compared to 1Gbit chips)
All in all, I think the L5410, L5420s on San clemente is a very potent web frontend box, and AMD Shanghai’s are pretty good backend database boxes, or memcache boxes. Although with the release of Dell R900 bios v1.1.9, the 7300 intel chipset now supports a whooping 128 total ranks of memory, giving you a 32 dimm*8GB quad ranks= 256GB total for very little money. I would now claim here that even with the disadvantage for FB-DIMMs, the R900 is the best Database box money can buy right now(4GB Quad rank memory is too cheap to ignore. Paying 2200 dollars for 32 dimms of 4GB quad ranks is too good, that Oracle and MySQL people would now have access to 128GB machines for less than 5000 dollars.
Your argument is good until Nehalem-EP comes out, which is very soon.
Yes you’re right that San Clemente only supports 6 dual-rank DIMMs (officially from Intel). Tyan however will certify and support 8 dual-rank DIMMs which for most people is enough. If you need more RAM and you’re not willing to wait for Nehalem-EP, then Harpertown with FBDIMM or Shanghai with DDR2 unbuffered are your only options, and Shanghai comes out with better performance/power. If you need less than 24 or 32 GB of RAM, San Clemente and Harpertown are your best deal.
Nehalem-EP is very good. However, DDR3 memory right now is at least 4 times the price of equivalent DDR2 Registered or FB-DIMM. Memory density is also lower unless you are willing to shell out 200 dollars per DIMM to MetaRam for their rank-fudging technology. Give it some time, and let the Server OS Kernels to catch up to Nehalem’s architecture. As far as I know, opensolaris 2008.11 is the only non-Microsoft server OS to officially support full Nehalem features.
San clemente also has a weakness. Dual channel memory vs quad channel on Harpertown. On DB workloads, you can see quiet a bit of difference. On single threaded apps, they can’t use the quad channel fully. Also, FB-DIMM gives you protection on the address lines as well as data compared to only data on the registered DIMMs.
16GB or 32GB of ram is last year’s hardware. Right now, Dell R900 with 128GB is a potent DB platform. Within a year, when DDR3 market settles down(ie. Price parity per GB compared to DDR2), hopefully the Nehalem platform is mature too then, with possibly a shrink to 32nm Sandy Bridge. Then is the time to buy Nehalem. It is kind of stupid to only give 2S Nehalem 24GB of ram(12x 2GB) for a server considering the awesome bandwidth.
Fit everything in ram. Max the ram sockets. Then max the hot swappable SAS bays with SSDs. That’s the way to do it nowadays. You are dead if you have to seek on rotational hard drive. Those are web 2.0 ways.
You still get a lot of performance with or without full optimizations, so it’s a bit of FUD to suggest that only optimized OSes will receive benefits. The enhanced CPU core and insane memory performance of the Nehalem platform helps all existing applications without a recompilation.
Hype by another hype, that’s AMD now. Some people are just easy to forget. Remember AMD claimed Barcelona will beat Kentsfield by 40% margin?