ARM based Opterons incoming!

Page 2 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

mrmt

Diamond Member
Aug 18, 2012
3,976
0
76
#26
Oh, this isn't going to make AMD much money. But it could be compelling if the price is right. A couple more ARM manufacturers jump in on this and you could have an old fashioned price war.
I don't think there is much room for price war here. The thing can't take a Xeon, and Avoton, Intel *current* offer for microserver, beats it both on raw performance *and* performance per watt. We should be talking about a bottom of the barrel product with correspondent low price and low margins, without much room for a price war.
 

Abwx

Diamond Member
Apr 2, 2011
8,777
106
126
#28
Price war in a tiny halfdead niche segment?

Calxeda didnt go bankrupt due to a growing profitable segment.

Who else is actually trying this segment, besides AMD and Intel?
There was a time when X86 servers were a niche in the server market.
 

Ancalagon44

Diamond Member
Feb 17, 2010
3,272
0
106
#29
Imagine if AMD used project Denver as the CPU instead of the Cortex A57?

I wonder if AMD is developing their own custom ARM CPU - all signs point to yes I suppose.
 

mrmt

Diamond Member
Aug 18, 2012
3,976
0
76
#30
Imagine if AMD used project Denver as the CPU instead of the Cortex A57?

I wonder if AMD is developing their own custom ARM CPU - all signs point to yes I suppose.
Yes, they probably are. There is no point in developing this subpar chip if they don't have something competitive in their R&D pipeline.
 

CHADBOGA

Golden Member
Mar 31, 2009
1,770
26
126
#31
Oh, this isn't going to make AMD much money.
I don't think it is going to make AMD any money.

But it could be compelling if the price is right.
I doubt AMD are in a position to provide the Contra Revenue that would be needed for this.
 

Abwx

Diamond Member
Apr 2, 2011
8,777
106
126
#32
I doubt AMD are in a position to provide the Contra Revenue that would be needed for this.
What you name contra revenue is an euphemism to not say
not competitive and not compelling products that must be
sold at loss to be of any interest , are you suggesting that
thoses chips are in the same position as intel s BT..??.
 

CHADBOGA

Golden Member
Mar 31, 2009
1,770
26
126
#33
What you name contra revenue is an euphemism to not say
not competitive and not compelling products that must be
sold at loss to be of any interest , are you suggesting that
thoses chips are in the same position as intel s BT..??.
Much, much worse position that BT.
 

mrmt

Diamond Member
Aug 18, 2012
3,976
0
76
#34
I doubt AMD are in a position to provide the Contra Revenue that would be needed for this.
Contra revenue makes sense when you have a good performance product that has some cost disadvantage, so you use contra revenue to bridge the cost gap and bring the customer towards you. This is not AMD's case. AMD ARM server is a vanilla design, that's probably as shoestring as it can get. The aim is to deliver the best performance/TCO for a significant part of the market, and build upon it.
 

krumme

Diamond Member
Oct 9, 2009
5,673
56
136
#35
There is no way a gf28nm is going to compete with a 22/14nm Intel solution for specint perf/eff. Quelle surprice. And there is no way this is a 100usd product. Its like selling Atom for the mobile market.

Besides the core integer speed difference between a15 and a57 is not that big according to arm own assessment, but memory speed takes a bigger hike.

Where is extreme low cost, seamicro fabric and high memory perf. an advantage?
What if tdp was say 9w and freq 1ghz?

An A15 is about 3.5mm2 sans l2/l3 what is an A57 aprox?
Do we have any assessment of the density of GF process?

Could the marginal production cost for this cpu go below 6usd?

What is the FP perf of A57, if strong, is there any customers for that?
 

Abwx

Diamond Member
Apr 2, 2011
8,777
106
126
#36
There is no way a gf28nm is going to compete with a 22/14nm Intel solution for specint perf/eff. Quelle surprice. And there is no way this is a 100usd product. Its like selling Atom for the mobile market.

Besides the core integer speed difference between a15 and a57 is not that big according to arm own assessment, but memory speed takes a bigger hike.

Where is extreme low cost, seamicro fabric and high memory perf. an advantage?
What if tdp was say 9w and freq 1ghz?

An A15 is about 3.5mm2 sans l2/l3 what is an A57 aprox?
Do we have any assessment of the density of GF process?

Could the marginal production cost for this cpu go below 6usd?

What is the FP perf of A57, if strong, is there any customers for that?
6$ for the raw chip , might be as much as 10$ once packaged
but then the dev. costs were minimal so this product will command
high intrinsical margin even if sold at double digit price.

FP perf is not of importance for its intended market , as for
the perf/watt it s 5W more than intels X86 Avoton but it has
much higher bandwith and has more capabilities.
 

Nothingness

Golden Member
Jul 3, 2013
1,874
30
106
#37
I'm not certain where Anandtech got the numbers in their table from as I haven't poured over AMD's PR materials, but they have a SPECint_rate Estimate for the A1100 of 80 with that estimated 25W TDP. Whereas according to Intel the C2750 gets a SPECint_rate_base of 106 with its 20W TDP.
Intel results were obtained with icc aka the SPEC/AnTuTu compiler. I'd like to see other benchmarks that are not overtuned by icc team. My experience with icc is that it usually is no better than gcc except for vectorization and for SPEC 2000/2006.

IMHO (and a lot of salt), given these results I'd say the chips have about the same performance (which means A57 IPC is higher, though who cares given the announced TDP...).
 

mrmt

Diamond Member
Aug 18, 2012
3,976
0
76
#38
What if tdp was say 9w and freq 1ghz?
Intel can reach 12W with 8 cores at 1.7Ghz, and that's with their 22nm. I think we can write everything AMD off when analysing raw performance and performance/watt. What remains to be seen is the kind of price AMD will be able to sell the thing.

Marginal cost isn't that relevant here. The heavy work in validation and debuging is that matters for server chips.
 
Mar 10, 2006
11,709
102
126
#39
Lower SPECint performance than currently-shipping Avoton (106 v.s. 80), higher TDP, and using a new ISA that doesn't have the maturity of x86 in this market?
 
Last edited:

krumme

Diamond Member
Oct 9, 2009
5,673
56
136
#40
Intel can reach 12W with 8 cores at 1.7Ghz, and that's with their 22nm. I think we can write everything AMD off when analysing raw performance and performance/watt. What remains to be seen is the kind of price AMD will be able to sell the thing.

Marginal cost isn't that relevant here. The heavy work in validation and debuging is that matters for server chips.
I have a dual core a9 at 650mhz in my little wd cloud nas. A completely new products 3 months old or so. The prior wd my book live had a single core.

A quad 5w a57 at 1ghz produced at 28nm is perhaps the next cpu in 2-3 years.

I dont know what debugging and validation is required on eg this market but marginal cost does for sure. Its hardly anything but cost.

And in more general about thd comments:

Everyone and his brother is going to the arm servermarket.

But as usual there is this usual whining and lol from the experts. The same goes for Intels new embedded products. The usual line of claims its the stupidest ever since 486.

As if its not professional making the decisions.

I think there is more perspective in trying to understand why those products is here, instead of showing our own competence. If it was anything worth we wouldnt be discussing it here.
 

ams23

Senior member
Feb 18, 2013
907
0
0
#41
Yes, they probably are. There is no point in developing this subpar chip if they don't have something competitive in their R&D pipeline.
AMD is many years and billions of dollars behind Apple, Qualcomm, Samsung, and NVIDIA when it comes to development of a fully custom ARM CPU. That is why their slides talk about 2019. Note that it takes about 5 years to develop a fully custom CPU. In the meantime, these other companies with better foresight from some years past will be marching ahead in the ARM space (while Intel will be marching ahead in the x86 space).
 
Last edited:

Vesku

Diamond Member
Aug 25, 2005
3,745
0
86
#42
As Khato pointed out, it doesn't have necessarily an edge on the I/O part. The only advantage they have over Avoton is that they can address double the amount of RAM, but that isn't helping them much in traditional server markets, I wouldn't bet that it would on microservers.
Actually one of the primary uses of high density servers is as big RAM farms.
 

NTMBK

Diamond Member
Nov 14, 2011
8,175
149
126
#44
AMD is many years and billions of dollars behind Apple, Qualcomm, Samsung, and NVIDIA when it comes to development of a fully custom ARM CPU. That is why their slides talk about 2019. Note that it takes about 5 years to develop a fully custom CPU. In the meantime, these other companies with better foresight from some years past will be marching ahead in the ARM space (while Intel will be marching ahead in the x86 space).
Those companies may have experience in developing an ARM CPU core, but they don't have experience in developing server fabric. And AMD has plenty of experience in developing x86 low power processors, which should be pretty transferable.

And yes, massive banks of memory is very important for this market. One of the main uses for microservers is memcached- having vast portions of databases hot in memory so that they can be quickly served up to users, instead of waiting for them to be loaded in from spinning disks. It's not a CPU intensive task, but it needs maximum density to cram as many DIMMs into the physical space as you can, as cheaply as you can, with a fast integrated network fabric.
 

Cerb

Elite Member
Aug 26, 2000
17,485
0
86
#45
There is no way a gf28nm is going to compete with a 22/14nm Intel solution for specint perf/eff. Quelle surprice. And there is no way this is a 100usd product. Its like selling Atom for the mobile market.
Which is working, surprisingly.

Besides the core integer speed difference between a15 and a57 is not that big according to arm own assessment, but memory speed takes a bigger hike.
Also, SPECInt is very likely going to give better results, relative to applications, on RISC CPUs, anyway, being mostly made up of loops that fit in various levels of caches. IoW, even AMD's own comparison against their Jaguar products should be taken with some salt.

Where is extreme low cost, seamicro fabric and high memory perf. an advantage?
Depends. I didn't see any costs for the CPUs, or supported servers (IE, HP), or bare-bones servers (IE, Supermicro). If they can get good SATA performance from all ports at once, while pushing enough data over the network to make 10GbE worth more than just not dropping down the speed at a switch, and the cost is low for having the 10GbE NICs, it could be quite compelling. Lots of ifs; not enough practical info.

What if tdp was say 9w and freq 1ghz?
Meh. The losses from fans, power supplies, drives, etc. make the SoC's power a secondary issue. it matters, but there's a point where reducing it isn't going to be worth the performance loss incurs from running at lower speeds. It's a balancing act, which includes up front costs, as well. You can bet AMD will offer slower models, though.

What is the FP perf of A57, if strong, is there any customers for that?
For what these are going to used for, nobody cares. Hardware FP support, to get low latencies for FP operations, may matter to some customers, but throughput of FP will not matter one bit to anyone even slightly interested. Xeon E5s, and even Teslas, aren't very expensive, if you have a business need for their FLOPS, and will generally be cheaper in the long run than many weaker processors.

These are targeted very much towards RAM caching, custom NAS/SAN infrastructures, shared web/DB hosting, network appliance makers, and so on. 99.999% int, and even then, very little throughput tends to be needed. Plenty of cores, for more active threads, and efficient interrupt handling, could matter, as could RAM capacity, but FP needs basically end at not being emulated in software, and having enough of the ISA's FP instructions supported to be able to use major distro X's binaries.

Lower SPECint performance than currently-shipping Avoton (106 v.s. 80), higher TDP, and using a new ISA that doesn't have the maturity of x86 in this market?

LOL!!! AMD is really done for.
Your username is showing. ARM is sufficiently mature for what these are targeted for. The software industry as a whole has made sure that this transition won't go down like IA32->x86-64 did. There's bound to be teething issues, but you can bet everything will be solid by the time 2015 rolls around, once real chips have been able to be tested with. Unlike with x86-64, most important software is already waiting on hardware, and devs and admins just need the hardware to test with to iron out the last kinks.

AMD may or may not be done for, but low-power server processors with support for many drives and gobs of RAM will make inroads, and there are companies hungrier than Intel, of which AMD is one. That, or Intel will to drop their ASPs and reduce their R&D, as part of trying to push others out, putting them on a path to not be the top-tier chip maker in the world, anymore, but just another Samsung, keeping up with the Joneses and abusing market positions. I have a feeling competition nipping at their heels is going to be better for us and them compared to that hypothetical scenario, so I'd rather it not go down like that. A short-term loss leader will not work. If they tried, as soon as they stopped losing money, a company that made ARM server SoCs for embedded use only would go ahead and remarket what they had, and start this all back up again.

Not going whole hog x86, however, is a bit worrisome. More because AMD has unique x86 designs to offer and improve upon, rather than any ARM software issues.
 

Nothingness

Golden Member
Jul 3, 2013
1,874
30
106
#46
Lower SPECint performance than currently-shipping Avoton (106 v.s. 80), higher TDP, and using a new ISA that doesn't have the maturity of x86 in this market?
Ashraf, you're smarter than that: don't use SPEC to compare Intel (or any company using their own compilers) vs others. Icc is overtuned for SPEC. It typically is about 10-20% over gcc on SPEC, while on most real non vectorizable code it's hardly faster if not slower, give icc a try it's free on Linux and compare against gcc. Just don't be afraid by the number of internal errors icc will spit at you :rolleyes:

IMHO the ISA is not an issue as this kind of chip will probably be used for servers running open source code.

But I certainly agree the TDP is very disappointing and the chip doesn't look good enough to switch away from x86. Just like Silvermont isn't good enough to switch away from ARM ;)
 

mrmt

Diamond Member
Aug 18, 2012
3,976
0
76
#47
AMD is many years and billions of dollars behind Apple, Qualcomm, Samsung, and NVIDIA when it comes to development of a fully custom ARM CPU. That is why their slides talk about 2019. Note that it takes about 5 years to develop a fully custom CPU. In the meantime, these other companies with better foresight from some years past will be marching ahead in the ARM space (while Intel will be marching ahead in the x86 space).
Actually, there is a practical reason for that. Most valuation models only cover 5 years, the rest of the value comes from perpetuity where there's not many changes in revenue composition and growth. That 2019 time frame and coupled with server estimates at the time fits that scenario dearly.
 

mrmt

Diamond Member
Aug 18, 2012
3,976
0
76
#48
That, or Intel will to drop their ASPs and reduce their R&D, as part of trying to push others out, putting them on a path to not be the top-tier chip maker in the world
Is that the only choice you can think? Drop margins should NEVER be your first choice. There's a lot of things that should come first, like Intel going for a far denser manufacturing node or using synthesis to make chip design cheaper. Intel is pursuing exactly that route.
 

mrmt

Diamond Member
Aug 18, 2012
3,976
0
76
#49
Actually one of the primary uses of high density servers is as big RAM farms.
Yes, but that restricts AMD to only... RAM farms, and that isn't a big advantage. Intel could just double the memory supported with Denverton and we're back to square 0. I'd wait for the TCO numbers. Only with final numbers we can see whether they have a business case or not.

I usually take AMD server marketing with truckloads of salt, mostly because of the old management shenanigans. New management was all right in servers, but it seems that old practices are back.

First the Warsaw launch, where we can find this pearl:

AMD said:
AMD Offers New Levels of Efficiency With 12- and 16-Core Additions to the AMD Opteron 6300 Series

AMD Opteron 6338P and 6370P Processors Deliver Optimized Performance Per-Watt Per-Dollar* for AMD Open 3.0 Open Compute Platforms
This tap dance to hide the fact that they shaved 14% of the clocks and got a 14% reduction in power consumption (and a whopping 22% reduction in turbo clocks). In fact, the new SKUs are less efficient than what AMD had on the market, it's just cheaper.

And now this, despite Puma arriving before ARM A57, it was Jaguar that was used to measure ARM performance.
 

erunion

Senior member
Jan 20, 2013
768
0
0
#50
Calxeda didnt go bankrupt due to a growing profitable segment.
Calxeda(and applied micro as well) was a vaporware company. It was more about cashing in on the demand to invest in ARM than the demand to purchase ARM.

Now that stock ARM v8 cores are available any startup can
bring a 64bit product to market at a fraction of the expense. In all seriousness, anyone can do what AMD did.
 


ASK THE COMMUNITY