Not slight more power. 25% more power. Now factor this for 100 chips and ask your Datacenter manager how much would it cost in two years to cool this extra power consumption.
Yes our imaginary system uses slightly more power. Like 5%? Probably less with "real" usage.
But that's only assuming you are correct with your figures (and I don't assume you are correct with your figures).
This post sounds like a post from someone who hasn't considered the numbers at all. A cost saving upfront will always be balanced with ongoing costs of cooling/power/space. But the only numbers you have seen (pure conjecture anyway) point to a situation which could easily be financially viable in the large scale.
Last edited:
Arachnotronic
Lifer
- Mar 10, 2006
- 11,715
- 2,012
- 126
The follow-on to Avoton is Denverton, and I think that Intel will probably keep clocks constant if it's an Airmont-based SoC (and I think it is). What we will probably see though is a significant increase in the # of CPU cores in a similar thermal envelope.
I expect to see this part shipping in late 2015 if this Intel roadmap is to be believed:
8x A57 vs. 4x Broadwell should be interesting.
Clocks will be higher on Broadwell D-1520 at 2.2 base, 2.6 Boost compared with (what I expect) 2.0-2.2 on Seattle. Broadwell of course includes hyperthreading.
Both have a dual channel memory controller but intel's supports 2133 mhz DDR4 while seattle is limited to 1866 mhz DDR4. Both support up to 128 GB RAM.
Seattle has a lot more cache (4 MB L2 + 8 MB L3) vs. the D-1520 (1 MB L2 and 6 MB L3).
Both have integrated dual 10 gb ethernet. Seattle supports 8 SATA3 drives while the D-1520 is limited to 6. However Broadwell D supports 24 lanes of PCIe 3.0 and 8 lanes of PCIe 2.0. Seattle is much more limited with 8x PCIe 3.0.
As far as pricing an performance goes, D-1520 and the top bin seattle are priced identically (ark intel prices are almost always overstated). As far as performance goes I think they will perform pretty similarly.
Clocks will be higher on Broadwell D-1520 at 2.2 base, 2.6 Boost compared with (what I expect) 2.0-2.2 on Seattle. Broadwell of course includes hyperthreading.
Both have a dual channel memory controller but intel's supports 2133 mhz DDR4 while seattle is limited to 1866 mhz DDR4. Both support up to 128 GB RAM.
Seattle has a lot more cache (4 MB L2 + 8 MB L3) vs. the D-1520 (1 MB L2 and 6 MB L3).
Both have integrated dual 10 gb ethernet. Seattle supports 8 SATA3 drives while the D-1520 is limited to 6. However Broadwell D supports 24 lanes of PCIe 3.0 and 8 lanes of PCIe 2.0. Seattle is much more limited with 8x PCIe 3.0.
As far as pricing an performance goes, D-1520 and the top bin seattle are priced identically (ark intel prices are almost always overstated). As far as performance goes I think they will perform pretty similarly.
No, no they don't look viable. The upfront cost for the processors is going to be pretty small compared to the life-time energy costs. Plus, presumably to make up for the performance gap you'd have to buy 20% more processors. Thus the overall upfront cost savings is marginal (~15%) , and the energy gap is going to be huge (at minimum >30% even given your generous estimate that real world differential per chip is around ~5%).
Define "lifetime"
Lesser upfront cost can be a good reason to upgrade to latest and greatest earlier. It will be obsolete soon. Simple math, but have 0 data to make it work.
Why you are not isolating the extra power consumption of the processor and mixing it with the system? It's the power consumption of the processor alone that you should factor in order to decide which one to use. So no, it's not 5%, but more likely 25% as the difference in TDP.
The energy gap at minimum will be ~0% (is this a magnitude thing?)
So yes, the imaginary AMD solution will be much better with the lower upfront costs? Where are you picking your numbers from..?
We are talking about the power the buyer pays for. You know, the power of the computers/racks/datacenters.
By your own numbers the AMD systems will be ~5% more power hungry? Maybe more under full load, but less with real world usage.
Last edited:
The energy gap at minimum will be ~0% (is this a magnitude thing?)
So yes, the imaginary AMD solution will be much better with the lower upfront costs? Where are you picking your numbers from..?
We are talking about the power the buyer pays for. You know, the power of the computers/racks/datacenters.
By your own numbers the AMD systems will be ~5% more power hungry? Maybe more under full load, but less with real world usage.
Clearly we are misunderstanding each other.
We have unverified numbers for Seattle that put it at:
80% performance of Avoton (based on SpecInt)
25% more power consumption (based on TDP)
30% less upfront costs (based on prices)
You suggested that, if those numbers are true, they represent a competitive product offering and that Seattle wouldn't be DOA. I disagreed with that assertion.
100 Seattle chips would equal 80 Avoton chips in performance.
Upfront price would equal 100 * .7 for the AMD chips and 80 * 1 for the Intel chips. 70/80 = .875. Meaning 12.5% upfront cost savings for the same performance.
Energy consumption = 100 * 1.25 for the AMD chips and 80 * 1 for the Intel chips. 125/80 = 1.56. Thus total energy consumption of just the chips is 56% greater for the AMD chips once equalized for performance.
Tell me again how that solution is competitive for high volume customers? Clearly the AMD product might find certain niche uses with one off consumers. But for a server product targeted at the commercial market, that looks DOA to me.
Again, all of this is based on our unverified numbers. Seattle may end up being much more competitive.
n.b., using full system power usage is a misleading statistic. It's especially misleading when you aren't also using full system prices for your comparison.
You are measuring the system, not the individual processor. That's not how you calculate TCO to a given CPU.
Sure, which is why my comment wasnt aimed at the accuracy of the estimate, but whether the estimate represented a competitive product. As i said, seattle may perform better.
AtenRa
Lifer
- Feb 2, 2009
- 14,003
- 3,362
- 136
Those SoCs are targeting the Data Centers. That means that those Cores are not going to work at 100% 24/7 365.
So when the 4 Core SoC will work with all four cores, the 8 Core SoC will only operate 4-5 cores, the rest will be power down. So at the end, the power consumption varies depending on the workload.
Just for those that use the TDP to estimate power consumption
So when the 4 Core SoC will work with all four cores, the 8 Core SoC will only operate 4-5 cores, the rest will be power down. So at the end, the power consumption varies depending on the workload.
Just for those that use the TDP to estimate power consumption
DrMrLordX
Lifer
- Apr 27, 2000
- 23,204
- 13,289
- 136
That's just SPECInt. Like I mentioned before, it might actually be better to have the 8-core A57 chip for something like webserver duty where you have numerous light tasks constantly demanding attention, rather than a raw-throughput situation where the CPU is constantly pegged at 100% and all resources are under constant utilization. I could see where the Seattle might be more useful for light VM terminal work as well. Under those circumstances, the higher TDP would justify itself, and the lower cost-to-entry would seal the deal.
Again, it's just SPECInt. Everyone throws out SPECInt numbers for server/workstation products. The POWER8 promoters threw out less-than-amazing SPEC scores, too, and then spent a lot of time arguing about why performance/watt didn't matter (ironically, the SPEC numbers they threw out there showed very good perf/watt for POWER8, but that's another matter).
You're probably right. I would think that most situations where throughput really matters will go to the D-1520, and probably to some Avoton solutions as well.
The issue in in scenarios where CPUs with high IPC/clockspeeds often go idle from lack-of-utilization. People don't want to buy more iron than they need, hence the interest in microservers with boatloads of cores so they can respond to a lot of requests (webserver, database server) without needing a lot of raw computational power per thread to retire tasks. Sure, you still have latency issues and such.
Don't get me wrong, I'm not all that bullish on the A-series Opterons: they suffered too many delays. AMD really needed to hit the market with these things BEFORE Broadwell-D launched to turn a few heads. Launching them side-by-side with Carrizo means AMD is late to the party.
Some people may still find a use for them where AMD's current server offerings make no sense, so that could be a plus. Maybe.
Though your post appears to be somewhat troll-ish, there's more here that bears mention: marketing matters.
Some Intel resellers could copy/paste mrmt's posts (with minor editing perhaps) and use that material in sales presentations to convince server customers to stay the heck away from A-series Opterons.
Their SPECInt numbers are lower than dated Avoton products, their listed TDP is higher, and they're delayed. Throw in some unit-distorted bar graphs for effect.
Plenty of middle manager and procurement types will look at that and go "thumbs down!" without even thinking about how the hardware actually performs in a work environment, considering the implications of going with an ARM-based system rather than x86, or anything else that actually might matter. Double this effect if the people holding the purse strings have a good working relationship with the vendor already.
I ran into this effect YEARS ago when interviewing for a tech/sales job I didn't land with an independent reseller that liked to sell people Intel-based machines. The k6 had just come out, I had one, and I was like "do you sell k6 for office apps? Boot magazine had some great benchmarks" and he was all "yeah they aren't compatible, I have a box of k6 chips in the back from replacements I've had to do".
I really wanted to see that box of chips too. I would been happy to take them off his hands! Greedy bastard didn't want to share.
Anyway, AMD is behind the 8-ball (again), this time because their rep has been tarred by the Construction core chips. Unless they have something that is balls-to-the-walls better than what Intel is offering in at least some significant way, it's going to be an uphill battle to get a buyer to actually consider the possibility that A-series Opterons might actually be better than an Avoton quad or a Broadwell-D quad in the same general price range at certain datacenter workloads. Let's face it, Seattle probably won't turn many heads.
Last edited:
I would expect AMD marketing team to do the sensible thing and throw the benchmarks to the press that show their product in the best possible light, and if they threw SPECInt, SPECInt it is. I expect other workloads and benchmarks to show the processor in even poorer light than SPECInt.
DrMrLordX
Lifer
- Apr 27, 2000
- 23,204
- 13,289
- 136
Why? Do you really think Seattle is going to show up comparatively WORSE in light webserver duty?
I have no idea why they stuck to a SPEC benchmark. My guesses would be:
1). It's standard procedure, and they did it without really thinking about it much
2). Thanks to A-series Opterons being ARM-based, they have less software available for benching (there's some open source stuff, but not everyone wants to use that)
3). The platform is immature and not performing properly on non-synthetic benchmarks.
It could be one or more of the above, who knows? Regardless, the performance/watt delta is MUCH smaller here than the comparative delta between Piledriver-based CPUs and competitive Intel offerings, which is something nobody here really seems to want to mention. Seattle's delays suck hardcore, but the scant data available shows it being pretty close to a similarly-priced Avoton solution. AMD can't even get within the same ballpark as Haswell-E using standard Opterons.
For anyone willing to take the leap into ARM server territory, I'm pretty sure Seattle will have its uses for light webserver, VM work, and maybe datacenter duty.
I have no idea why they stuck to a SPEC benchmark. My guesses would be:
1). It's standard procedure, and they did it without really thinking about it much
2). Thanks to A-series Opterons being ARM-based, they have less software available for benching (there's some open source stuff, but not everyone wants to use that)
3). The platform is immature and not performing properly on non-synthetic benchmarks.
It could be one or more of the above, who knows? Regardless, the performance/watt delta is MUCH smaller here than the comparative delta between Piledriver-based CPUs and competitive Intel offerings, which is something nobody here really seems to want to mention. Seattle's delays suck hardcore, but the scant data available shows it being pretty close to a similarly-priced Avoton solution. AMD can't even get within the same ballpark as Haswell-E using standard Opterons.
For anyone willing to take the leap into ARM server territory, I'm pretty sure Seattle will have its uses for light webserver, VM work, and maybe datacenter duty.
If AMD numbers are correct for a 100% utilization scenario then we have 25% more power consumption for 20% less performance, which translates into a ~50% extra power consumption for a given performance level... not a sane choice for a Datacenter, this, on top of the headache of switching software, which is not something trivial. If all AMD has to offer to counter these issues is a 15% rebate on the processor then it's just a bad business proposition.
But let's say that these numbers are wrong and that in reality Seattle is much better than what I'm saying here. Well, that would mean AMD marketing guys are dumb beyond any recovery, because the numbers they bothered to release aren't showing their products in the best possible light, and they would have been even dumber in not speeding up the ramp up of the chip as soon as possible, especially because AMD is in dire need of extra revenue.
Regarding this last topic I would like to point out that AMD hasn't committed yet to the production ramp up of the Seattle chip:
Lisa Su said:
She doesn't know when it is going to full production.
DrMrLordX
Lifer
- Apr 27, 2000
- 23,204
- 13,289
- 136
1). AMD's marketing has already demonstrated itself to be subpar in the past
2). Aren't SPEC scores composites of several scores from a suite of benchmarks? I tried looking them up on spec.org but there's nothing published for Q1 or Q2 2015 related to A-series Opterons, so I can't see a breakdown of how Seattle performed in each part of the benchmark suite. Wherever AMD published these numbers, they didn't go through the trouble of getting them listed on spec.org (which is disappointing).
As much as I like to knock AMD's marketing dept., I don't think they have much say in when Seattle will be ready to ship.
Her comment is (perhaps purposefully) vague. I do remember seeing some material from AMD not too long ago comfirming the Carrizo launch around Computex which also covered 300-series GPUs and A-series Opterons. They've recently made the claim that Seattle would finally be ready by then. If they back off of that release date, they may as well throw the chip into whatever dark closet that swallowed Berlin and their other cancelled products from days gone by.
I use open source software... So if I compile it for ARM then your suggestion have much worse support
Even if you use open source software, the support for Avaton is much greater. Do you want to use a full software stack that people have been running and debugging for years, or one that you had to work through to get to work for the first time... OSS isn't a pancea, it merely aides in porting efforts, but does not production ready make.
It matters because it means that the developers are intimately familar with the whole stack.
Have fun running a completely new virgin network stack with unknown issues vs one that has been running for years... These things actually matter, esp if you run into an issue. Its like saving a couple bucks running a Marvell based SAS controller vs and LSI one. Post a support question using a Marvell SAS controller and the first 100 responses will be why aren't you using LSI...
Do keep in mind that just about anyone that would be in the market for a Seattle like part has probably already bought and put into production an Avaton part. So far the data points to the Seattle part not really being an upgrade vs Avaton.
Those SoCs are targeting the Data Centers. That means that those Cores are not going to work at 100% 24/7 365.
So when the 4 Core SoC will work with all four cores, the 8 Core SoC will only operate 4-5 cores, the rest will be power down. So at the end, the power consumption varies depending on the workload.
Just for those that use the TDP to estimate power consumption
Which would be great except in this comparison the 4core SOC runs rings around the 8 core soc in performance and by all historical data has significantly better power management...
As a counter argument, actual customers have found Avaton which has roughly the same performance as Seattle to be severely lacking in performance for actual production environments outside of storage controllers and network switch controllers(Facebook is the prime example here). These actual customers however appear to be buying Xeon-Ds by the truckload ATM.
Actually, real actual customers are going with D-1540s for the light tier (aka front end) of their server infrastructure. The cost differential between D-1520 and D-1540 is dwarfed by the cost of the ancillary hardware (memory, storage, enclosure, floor space, power, etc).
The problem is the microservers as you envision them don't have enough single thread performance. Hence why people like facebook are using D-1540 instead of Seattle/Avaton for their light tiers.
In at least one public case, the experience with lower ST performance is that it requires significantly more development effort which dwarfs any potential saving from microservers.
The problem is that the kinds of workloads viable for Seattle have also been viable for Avaton, and Avaton has been available for almost 2 years and in already in production enviroments.
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads
- Started by Tigerick
- Replies: 24K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
-
Facebook
Twitter