[AT] AMD's Seattle server SoC

[NTMBK]

http://anandtech.com/show/8362/amds-big-bet-on-arm-powered-servers-a1100-revealed

Seems like a very interesting part. A true server SoC that actually integrates things you would want (10Gb ethernet, SATA, PCIe, compression/decompression/crypto acceleration) while ditching stuff you don't need (GPU). When you consider everything that you can drop from a server motherboard with a part this integrated, system prices should be fairly low. No discrete I/O hub (and no need to cool it), no 10Gb NICs....

 

[monstercameron]

I guess it could be one of there http://en.wikipedia.org/wiki/Heavy_equipment#Types

maybe a cherrypicker or a backhoe.

and I also guess that dual 10Gbe could be exciting for the server folks.

 

[Vesku]

If the base board is $500-600 or less 2x10Gbit and 8xSATA with that PCIe expansion will be an attractive package. Age old question for AMD is will they have this out in retail in volume in a reasonable time frame? AMD initially planned to launch the A1100 development kit in April-May not August, so will they still have retail availability in Oct/Nov?

 

[piesquared]

http://www.theregister.co.uk/2014/08/11/amd_seattle_64_bit_arm/

http://www.enterprisetech.com/2014/08/11/amd-details-seattle-arm-server-chip/

http://www.tomsitpro.com/articles/amd-opteron-seattle-arm-soc,1-2105.html

http://www.anandtech.com/show/8362/amds-big-bet-on-arm-powered-servers-a1100-revealed

Nice timing by AMD. The demand for fast, low power, efficient throughput is huge, it seems AMD has built a solution to solve that. There are others also vying for a piece of the pie but with AMD's server and 64-bit experience they will be the front runners in a fast growing market.
The authors seem quite positive and comments are also quite positive.

PS Hey who won that Superbowl, Seattle or Denver... I forget!

 

[Shivansps]

I wonder how it compares to Avotons.

 

[Vesku]

I wonder how it compares to Avotons.

Has more built-in IO and crypto acceleration support should be quite good. Haven't seen anything regarding virtualization support.

Here is ARM's estimate for their standard A57 core, AMD is using stock A57:

http://www.arm.com/products/processors/cortex-a50/cortex-a57-processor.php

Performance should be at least OK, the small heatsink on AMD's reference board is a good sign for efficiency but then again it is the 4 core version and not the 8 core afaik. I'm guessing CPU performance will be a bit below Avoton especially given the suggested 2GHz speed. Might beat it in a handful of operations, I suppose.

 

[piesquared]

Avaton is behind Seattle in many areas. Supporting DDR4 and 128 GB is one key area. IO is better, perf/W is undoubtedly going to be better, dedicated hardware for crypto, TrustZone, dual 10Gb KR Ethernet ports. Lots of goodies in there, very nice.

There's also talk of AMD building custom ARM 64 bit server chips.

http://www.pcworld.com/article/2464080/amd-mulls-custom-arm-64bit-server-chips.html

 

[beginner99]

Avaton is behind Seattle in many areas. Supporting DDR4 and 128 GB is one key area. IO is better, perf/W is undoubtedly going to be better, dedicated hardware for crypto, TrustZone, dual 10Gb KR Ethernet ports. Lots of goodies in there, very nice.

There's also talk of AMD building custom ARM 64 bit server chips.

http://www.pcworld.com/article/2464080/amd-mulls-custom-arm-64bit-server-chips.html

How is DDR4 key? 128 GB might matter but not that it's DDR4. Current DDR4 isn't even faster but more expensive than DDR3. Actually DDR4 is probably more important for consumer platform due to iGPU than for server.

 

[Vesku]

How is DDR4 key? 128 GB might matter but not that it's DDR4. Current DDR4 isn't even faster but more expensive than DDR3. Actually DDR4 is probably more important for consumer platform due to iGPU than for server.

More power efficient, higher density both properties important for servers. Historically server buyers have been willing to pay early adopter premium on memory.

 

[dealcorn]

More power efficient, higher density both properties important for servers. Historically server buyers have been willing to pay early adopter premium on memory.

Until I see contrary evidence, I suspect Avoton/Rangerly will be more efficient. Avoton/Rangerly is built with more efficient transistors. That may be the reason why AMD has not disclosed TDP. Seattle's ability to use more memory is a clear advantage in certain applications. Density and efficiency may be intimately related. It is early to call a density winner.

 

[Vesku]

Until I see contrary evidence, I suspect Avoton/Rangerly will be more efficient. Avoton/Rangerly is built with more efficient transistors. That may be the reason why AMD has not disclosed TDP. Seattle's ability to use more memory is a clear advantage in certain applications. Density and efficiency may be intimately related. It is early to call a density winner.

Was pointing out why DDR4 support may be important to a server buyer, DDR4 is more efficient (lower voltage higher density).

 

[itsmydamnation]

Until I see contrary evidence, I suspect Avoton/Rangerly will be more efficient. Avoton/Rangerly is built with more efficient transistors. That may be the reason why AMD has not disclosed TDP. Seattle's ability to use more memory is a clear advantage in certain applications. Density and efficiency may be intimately related. It is early to call a density winner.

As a web server front end ( what this is attractive for) dedicated hardware crypto is a big advantage even compared to dedicated AES instructions . that alone will be a big win.

AMD have stated previously that they expect a pretty big jump from jaguar to A-57 so it should have a pretty good performance lead, perf per watt, who really knows........

that said i wish amd had been a little more bold, 16 cores would have created more of a differentiation.

 

[Shivansps]

Hang on a minute, a cryptographic coprocessor capable of doing sha256? mmmmmmmmm

 

[Erenhardt]

isn't that used for cryptomining?

 

[Mopetar]

isn't that used for cryptomining?

Or just anything that needs the functionality. For example, authenticating passwords or anything else that requires hashing is going to be a lot more efficient with dedicated hardware, both in terms of reducing the power draw and freeing up the CPU to do something else.

 

[piesquared]

So who among today's hardware vendors does Red Hat's chief ARM architect think is doing the job right? Not surprisingly, perhaps, at LinuxCon he gave special mention to AMD's "Seattle" SoC, which the chipmaker officially launched at the Hot Chips conference earlier this month and which Masters says is "done right in every single way."

"It's a standardized, server-grade SoC. It follows every single server design philosophy that AMD knows from working with x86, and it also follows all of the kinds of advice and guidelines that the industry has been working on over the last three years. It's a very, very nice design," Masters said.

http://www.theregister.co.uk/2014/08/22/red_hat_arm_server_standards/?page=1

 

[Abwx]

There s something that i can hardly understand , if someone has a clue
it would be welcomed..

According to AMD the Specint scores of Seattle and a quadcore 1.9 Jaguar
are 80 and 28.1 respectively :
http://www.anandtech.com/show/7724/...arm-based-server-soc-64bit8core-opteron-a1100

Now Intel said that it s top Avoton SKU with 8 cores score 106 but avoton has Silvermont cores wich are inferior IPC wise to Jaguar, yet Intel pretend that it has double the IPC, on a core per core basis, compared to Jaguar.

So how was Silvermont IPC miraculously doubled once they crammed eight of them in a single die..??..

 

[monstercameron]

There s something that i can hardly understand , if someone has a clue
it would be welcomed..

According to AMD the Specint scores of Seattle and a quadcore 1.9 Jaguar
are 80 and 28.1 respectively :
http://www.anandtech.com/show/7724/...arm-based-server-soc-64bit8core-opteron-a1100

Now Intel said that it s top Avoton SKU with 8 cores score 106 but avoton has Silvermont cores wich are inferior IPC wise to Jaguar, yet Intel pretend that it has double the IPC, on a core per core basis, compared to Jaguar.

So how was Silvermont IPC miraculously doubled once they crammed eight of them in a single die..??..

how did you get to this conclusion? Link?

 

[Abwx]

how did you get to this conclusion? Link?

Seattle and Kabini Specint scores are on the Anand s article i posted , Avoton score was published by Anand and other sites.

An Intel Avoton (top SKU) has a SpecInt score of 106 and a TDP of 20W.

http://hothardware.com/News/AMD-Ann...SoC-At-Facebooks-Open-Compute-Summit/#!bKor3Z

This score is exactly double of what it should be if we look at Jaguar vs Silvermont IPC in the Windows tests, so either it s a mistake or it is deliberate misleading of the general public, what is surprising is that no one noticed the oddity.

For the record a 1.9 4C Kabini scores 28.1 , so how could 8 silvermont cores scores 106 knowing that the top model SKUs cores IPCs and IPS are supposed to be on the same ballpark as this Kabini.?.

This imply that a 4C Silvermont score would be at least 53, double of a 4C Jaguar.

 

[ShintaiDK]

Its estimate. And its double the cores plus twice the memory channels with higher speed. Thats the key in this throughput bench.

I am sure the Seattle pulls 25GB/sec like the Avoton. While Kabini and Baytrail is down in the 10GB/sec area.

The X2150 is also 1.1Ghz to 1.9Ghz. So it may throttle.

 

[Abwx]

Its estimate. And its double the cores plus twice the memory channels. Thats the key in this throughput bench.

I am sure the Seattle pulls 25GB/sec like the Avoton. While Kabini and Baytrail is down in the 10GB/sec area.

Double the cores and 4 x the score.?.

Besides more memory channels, yes, but with double the cores to feed so your 25GB/s is not that more than 10GB/s once we look at the available bandwith/core, about 20-25% higher bandwith/core.

Also it s an integer bench, not a bandwith bench, i would say that this estimation is almost 100% too optimistic.

 

[ShintaiDK]

Double the cores and 4 x the score.?.

Besides more memory channels, yes, but with double the cores to feed so your 25GB/s is not that more than 10GB/s once we look at the available bandwith/core, about 20-25% higher bandwith/core.

Also it s an integer bench, not a bandwith bench, i would say that this estimation is almost 100% too optimistic.

Rate is also a bandwidth related bench. Its throughput.

And the Kabini is specified as 1.1Ghz-1.9Ghz.

 

[Abwx]

Rate is also a bandwidth related bench. Its throughput.

Indeed but the lower the IPC the less the needed bandwith.

And the Kabini is specified as 1.1Ghz-1.9Ghz.

Kabini s 28.1 score is at 1.9, you would think that if it was at 1.0 or somewhere between 1.0 and 1.9 AMD would be crazy to not use it instead of the A57.

Edit : i m inclined to believe that Avoton score is in the mid 50s,
a 8C Kabini AMD couldnt get much better scores so they relyied on the A57 since the apparent 50% higher score is way too much to match in the short term with revised Kabinis that wont come till next year.

 

[SlimFan]

Frequency matters. The Avoton is listed with a turbo frequency of 2.6GHz. So 2x the cores, 1.4x the frequency, more DRAM bandwidth available...

 

[Abwx]

Well actualy Shintaidk was right in that it was SPECint2006_rate that is effectively a measurement of bandwith more than anything else, the IPC is measured by SPECint2006 and the scores looks realistic, i just checked at SPEC site, last quarter Intel made a consumers CPUs submissions that included their own mainstream plateform as well as AMD s, so we have scores of popular CPUs while Supermicro did submit a C2750 2.4GHz as a welcomed basis.

http://www.spec.org/cpu2006/results/res2014q3/