Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/qualcomm-begins-commercial-sampling-of-worlds-first-10nm-server-processor.8733/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021370
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

Qualcomm Begins Commercial Sampling of World's First 10nm Server Processor

mbello

Member
Qualcomm Begins Commercial Sampling of World’s First 10nm Server Processor and Reshapes the Future of Datacenter Computing | Qualcomm

48-core processor, shipping in H2/2017

To me it felt like Qualcomm was taking forever to release their server processor, but the fact that they will launch it on 10nm process means they did not want to take any chances.

I just hope Qualcomm is also investing in everything else that is required for their processor to perform well, starting with the open source compilers. For reference, take a look at this article that showed how much of an improvement one could get for Cavium ThunderX performance just by using the latest compilers (unreleased at the time of launch of Cavium's ThunderX) Cavium ThunderX Benchmarks Part II: Why enterprise ARM developers need these machines.

Well, 2017 will be an exciting year to watch, on the server/datacenter side Intel will be under attack by ARM, Power9 (search for OpenPower to understand how Power is becoming relevant) and AMD with Zen. And Zen I think is the least dangerous threat to Intel's Datacenter Business. I just wish Power9 would launch on 10nm process, but it seems unlikely.
 
Last edited by a moderator:
Zen is actually the biggest threat to Intel's sever business. It doesn't have the penalty of requiring the software to be ported - it all runs on it out of the box and AMD have much more experience with servers than Qualcomm.
 
Qualcomm Begins Commercial Sampling of World’s First 10nm Server Processor and Reshapes the Future of Datacenter Computing | Qualcomm

48-core processor, shipping in H2/2017

To me it felt like Qualcomm was taking forever to release their server processor, but the fact that they will launch it on 10nm process means they did not want to take any chances.

I just hope Qualcomm is also investing in everything else that is required for their processor to perform well, starting with the open source compilers. For reference, take a look at this article that showed how much of an improvement one could get for Cavium ThunderX performance just by using the latest compilers (unreleased at the time of launch of Cavium's ThunderX) Cavium ThunderX Benchmarks Part II: Why enterprise ARM developers need these machines.

Well, 2017 will be an exciting year to watch, on the server/datacenter side Intel will be under attack by ARM, Power9 (search for OpenPower to understand how Power is becoming relevant) and AMD with Zen. And Zen I think is the least dangerous threat to Intel's Datacenter Business. I just wish Power9 would launch on 10nm process, but it seems unlikely.

I chatted with some folks at IEDM about this and was told that QCOM is really focusing on China. Remember QCOM is partnering with Guizhou Province on a $280M joint venture (Guizhou Huaxintong Semiconductor Technology). I was also told that FaceBook, Amazon, and Google are working with QCOM on this as an alternative to Intel.

You should also note that Anand Chandrasekher (SVP of Qualcomm Datacenter Technologies) spent his career at Intel so there is some experience behind this effort. You can see his 10nm chip blog here:

Meet Qualcomm Centriq 2400, the world’s first 10-nanometer server processor | Qualcomm

So I agree, the ARM server chip market is getting interesting again.
 
Zen is actually the biggest threat to Intel's sever business. It doesn't have the penalty of requiring the software to be ported - it all runs on it out of the box and AMD have much more experience with servers than Qualcomm.

AMD also has a $371 million Chinese venture:

AMD Creates Joint Venture With China's Nantong Fujitsu - Bloomberg

IBM OpenPower is in China as well. I'm not sure if this is a "winner take all" strategy or "cover all bases" but clearly China wants an alternative to Intel for their massive cloud.
 
This is from LinkedIn:

"I don't understand why ARM servers are not getting adopted. Can somebody explain why ?"

Basil Varghese
There are no viable product available at the moment, and there will be none shipping in next 12-18 months. Qulacomm is starting to sample, but typically it will be 12-18 months before design and volume would materialize and products will be qualified. Lack of products is due to a combination of technical and business constraints. ARM is yet to create to server class CPU. So, a company with ARM v8 ISA license had to create a custom CPU to make product with Xeon class performance. Not many companies have the expertise, deep pockets and patience to develop such a CPU architecture and a product. Even if a company manage to tweak the CPU to generate the performance, it will take couple of iterations to create a product with decent power-performance combination to compete with Intel at the hyperspace customer(GOOG, FB, AMZN), which is where the money is. To profitably manufacture an advanced node (16nm and below) with a decent gross margin, companies need good volume at foundries, which is a club with very few members. Also, the hyperspace customers will not be interested unless there is a long term roadmap from a company with history of successful execution, generation after generation. QCOMs product is promising, but it will be 2-3 year


I agree with this completely. From what I have heard QCOM has already taped out a 7nm version of this chip at TSMC. The 10nm is proof of concept and the 7nm will be the production version in 2018.... my opinion.
 
AMD also has a $371 million Chinese venture:

AMD Creates Joint Venture With China's Nantong Fujitsu - Bloomberg

IBM OpenPower is in China as well. I'm not sure if this is a "winner take all" strategy or "cover all bases" but clearly China wants an alternative to Intel for their massive cloud.

Just to point out that the JV you mentioned isn't in any way related to CPU, that JV is about a testing, assembly and packaging factory owned by AMD but not critical for their operations

Key highlights of the deal include:
  • NFME’s affiliates have purchased an 85 percent share of AMD’s Penang, Malaysia and Suzhou, China ATMP operations and act as controlling partner for the new joint venture business.
  • AMD received from NFME approximately $371 million, excluding purchase price adjustments, with net cash proceeds of approximately $320 million after payment of taxes and other customary expenses. AMD retains a 15 percent ownership of the Penang and Suzhou operations.
AMD and Nantong Fujitsu Microelectronics Co., Ltd. Close On Semiconductor Assembly and Test Joint Venture

AMD has a chinese JV regarding a CPU licensing deal, but previously disclosed that it was x86 related, probably their new ZEN core

Advanced Micro Devices (AMD) has created a JV with THATIC (Tianjin Haiguang Advanced Technology Investment Co., Ltd.) to develop SoCs for the Chinese server market. Essentially, AMD will license its x86 processor technology and the IP (intellectual property) related to designing an SoC (system on chip), so I expect this involves IP like memory controllers, input/output and caching, but not GPU technology.
Neither THATIC or AMD is commenting on which flavor cores these will be or any more technical details, but I have to expect that it is Zen
Forbes Welcome

As for the reasons for ARM servers not having take off, well it might be a chicken and egg problem, you can't have hardware without Software support, and by SW i mean OS and hypervisors to run it.

Until now the only OS you could run would be OpenSUSE or Fedora, but starting the beginning next year you will have SUSE Enterprise Linux and Red Hat Enterprise Linux Edition supporting out of the box ARMv8

SLES for ARM is part of the SUSE Linux Enterprise 12 common code base. This means that the versions, and thereby the source code, of all core packages of the SUSE Linux Enterprise product family are the same—from the desktop to the mainframe. The tool chain, like compilers and libraries are the same across the supported hardware architectures. The common code base guarantees product consistency and a persistent look and feel, which lets you leverage the skills of your IT staff. It also provides for the highest code quality, better supportability and preemptive code maintenance.
SLES for ARM will become available later this year and initially will support SOCs from AMD, Applied Micro, Cavium, NXP, and Xilinx. A number of solution vendors will exploit SLES for ARM to deliver servers to support a wide range of workloads. SLES for ARM will initially be available for partner exploitation. SUSE will continue to expand the list of supported SOC and solution vendors as the ARM 64-bit server community expands.
Introducing SUSE Linux Enterprise Server for ARM - SUSE Blog | SUSE Communities
 
The software ecosystem is developing slowly, but it has continued to evolve. The thing about disruption is that it happens slowly, and then all of a sudden. I believe ARM will one day become a serious threat to Intel's datacenter business, but the threat will take a few more years to be realized. I'd bet 7nm will be the turning point, at which point the foundries will have a decisive process lead. Once the foundries are a node ahead of Intel, it will be impossible to ignore the performance/cost advantage that the ARM/foundry business model will offer.
 
The software ecosystem is developing slowly, but it has continued to evolve. The thing about disruption is that it happens slowly, and then all of a sudden. I believe ARM will one day become a serious threat to Intel's datacenter business, but the threat will take a few more years to be realized. I'd bet 7nm will be the turning point, at which point the foundries will have a decisive process lead. Once the foundries are a node ahead of Intel, it will be impossible to ignore the performance/cost advantage that the ARM/foundry business model will offer.

Going by the latest analysis from Scotten-Jones, Intel 10nm and TSMC 7nm are very similar processes (same CPP/MMP) and have similar timescales, so I wouldn't agree that "the foundries will have a decisive process lead" -- unless you use the marketing-BS labels instead of real geometries.

It is of course possible that TSMC will ramp up their large-die yield faster than Intel given last-node experiences at 14nm (Intel) vs. 10nm (TSMC), but this is just an assumption -- as they say, past performance is no guarantee of future performance ;-)
 
Yes, that's fair to say. TSMC 7nm and Intel 10nm will probably be comparable in density and performance, but my guess is that TSMC will be a much lower cost. So by decisive lead, I mean in terms of performance/cost rather than just performance. On the other side, I see Intel moving into the foundry space to be a big positive as if Intel foundry volumes are large enough it can help bring down their unit costs.
 
Back
Top