Sign up for our Technology Newsletter 
nVidia NVLINK 2.0 Going In IBM Servers
Comments Off on nVidia NVLINK 2.0 Going In IBM Servers
On Monday, PCWorld reported that the first servers expected to use Nvidia’s second-generation NVLINK 2.0 technology will be arriving sometime next year using IBM’s upcoming Power9 chip family.
IBM launched its Power8 lineup of superscalar symmetric multiprocessors back in August 2013 at the Hot Chips conference, and the first systems became available in August 2014. The announcement was significant because it signaled the beginning of a continuing partnership between IBM and Nvidia to develop GPU-accelerated IBM server systems, beginning with the Tesla K40 GPU.
The result was an HPC “tag-team” where IBM’s Power8 architecture, a 12-core chip with 96MB of embedded memory, would eventually go on to power Nvidia’s next-generation Pascal architecture which debuted in April 2016 at the company’s GPU Technology Conference.
NVLINK, first announced in March 2014, uses a proprietary High-Speed Signaling interconnect (NVHS) developed by Nvidia. The company says NVHS transmits data over a differential pair running at up to 20Gbps, so eight of these differential 20Gbps connections will form a 160Gbps “Sub-Link” that sends data in one direction. Two sub-links—one for each direction—will form a 320Gbps, or 40GB/s bi-directional “Link” that connects processors together in a mesh framework (GPU-to-GPU or GPU-to-CPU).
NVLINK lanes upgrade from 20Gbps to 25Gbps
IBM is projecting its Power9 servers to be available beginning in the middle of 2017, with PCWorld reporting that the new processor lineup will include support for NVLINK 2.0 technology. Each NVLINK lane will communicate at 25Gbps, up from 20Gbps in the first iteration. With eight differential lanes, this translates to a 400Gbps (50GB/s) bi-directional link between CPUs and GPUs, or about 25 percent more performance if the information is correct.
NVLINK 2.0 capable servers arriving next year
Meanwhile, Nvidia has yet to release any NVLINK 2.0-capable GPUs, but a company presentation slide in Korean language suggests that the technology will first appear in Volta GPUs which are also scheduled for release sometime next year. We were originally under the impression that the new GPU architecture would release in 2018, as per Nvidia’s roadmap. But a source hinted last month that Volta would be getting 16nm FinFET treatment and may show up in roughly the same timeframe as AMD’s HBM 2.0-powered Vega sometime in 2017. After all, it is easier for Nvidia to launch sooner if the new architecture is built on the same node as the Pascal lineup.
Still ahead of PCI-Express 4.0
Nvidia claims that PCI-Express 3.0 (32GB/s with x16 bandwidth) significantly limits a GPU’s ability to access a CPU’s memory system and is about “four to five times slower” than its proprietary standard. Even PCI-Express 4.0, releasing later in 2017, is limited to 64GB/s on a slot with x16 bandwidth.
To put this in perspective, Nvidia’s Tesla P100 Accelerator uses four 40GB/s NVLINK ports to connect clusters of GPUs and CPUs, for a total of 160GB/s of bandwidth.
With a generational NVLINK upgrade from 40GB/s to 50GB/s bi-directional links, the company could release a future Volta-based GPU with four 50GB/s NVLINK ports totaling of 200GB/s of bandwidth, well above and beyond the specifications of the new PCI-Express standard.
Courtesy-Fud
IBM Goes Bare Metal
IBM has announced the availability of OpenPower servers as part of the firm’s SoftLayer bare metal cloud offering.
OpenPower, a collaborative foundation run by IBM in conjunction with Google and Nvidia, offers a more open approach to IBM’s Power architecture, and a more liberal licence for the code, in return for shared wisdom from member organisations.
Working in conjunction with Tyan and Mellanox Technologies, both partners in the foundation, the bare metal servers are designed to help organisations easily and quickly extend infrastructure in a customized manner.
“The new OpenPower-based bare metal servers make it easy for users to take advantage of one of the industry’s most powerful and open server architectures,” said Sonny Fulkerson, CIO at SoftLayer.
“The offering allows SoftLayer to deliver a higher level of performance, predictability and dependability not always possible in virtualised cloud environments.”
Initially, servers will run Linux applications and will be based on the IBM Power8 architecture in the same mold as IBM Power system servers.
This will later expand to the Power ecosystem and then to independent software vendors that support Linux on Power application development, and are migrating applications from x86 to the Power architecture.
OpenPower servers are based on open source technology that extends right down to the silicon level, and can allow highly customised servers ranging from physical to cloud, or even hybrid.
Power systems are already installed in SoftLayer’s Dallas data centre, and there are plans to expand to data centres throughout the world. The system was first rolled out in 2014 as part of the Watson portfolio.
Prices will be announced when general availability arrives in the second quarter.
HP’s Z-station Goes Nvidia
HP has added its Z Workstation family with a solution that delivers access via a virtual desktop route to workstation applications hosted in the data center.
Set to be available from next month, the HP DL380z Virtual Workstation enables organisations to provide remote access to workstation-class applications, even those calling for heavy-duty graphics, which allows them to keep data stored securely in the data centre wherever employees might be based.
As its name suggests, the HP DL380z is based on the same hardware as HP’s ProLiant DL380p server, a 2U rack-mount two-socket system based on Intel’s Xeon E5-2600 processors, which allows it to slot right into existing data centre infrastructure.
Where the HP DL380z differs is that it can be configured with up to two Nvidia Grid K2 graphics cards supporting the graphics firm’s Grid GPU virtualisation technology. This enables up to eight users to be hosted on each system, each with access to a virtual machine with GPU acceleration capabilities.
Jeff Groudan, worldwide director for HP Thin Client and Virtual Workstations, said, “For employees who work from A to B and everywhere in between, the HP DL380z allows them to access data that is securely stored in the data centre. Furthermore, the powerful HP DL380z is an always-on workhorse that can be used by businesses when not in use for virtual workstation sessions.
Remote access is delivered either by operating Citrix’s XenServer with its HDX 3D Pro technology, which the HP DL380z is certified for, or by utilising HP’s own Remote Graphics Software (RGS). The latest HP RGS release 7 adds the ability to have true workstation productivity from a tablet while bringing intuitive touch controls to non-touch applications, according to HP.
Either way, customers can provide engineers or other professional users with access to workstation-class applications from a variety of devices, including thin clients, laptops or tablets.
Pricing for the HP DL380z has yet to be confirmed.
HP & Foxcomm Head To The Cloud
May 20, 2014 by admin
Filed under Around The Net
Comments Off on HP & Foxcomm Head To The Cloud
HP and Foxcomm have announced a joint venture to create a line of cloud optimized servers for service providers.
The venture involving a non-equity, strategic commercial alliance will see the pair offering a range of products. Particulars and specifications are yet to be announced but the companies are aiming to target low total cost of ownership (TCO), scale and service.
This announcement is separate to the existing HP Proliant server portfolio, which includes the software defined server codenamed Moonshot.
HP CEO Meg Whitman said, “With the relentless demands for compute capabilities, customers and partners are rapidly moving to a New Style of IT that requires focused, scalable and high-volume system designs. [The partnership] will enable us to deliver a game-changing offering in infrastructure economics.”
News of the alliance will raise eyebrows at Apple, which reportedly returned an eight million unit shipment of iPhones to Foxconn last year, describing them as “dysfunctional” and “non-compliant”.
HP has had its own troubles recently, after settling two lawsuits this month, one to the former shareholders of Palm over its handling of WebOS, and another that revealed that HP executives were guilty of corruption in negotiations for lucrative contracts. Total payouts across the two settlements totaled $165m.
The HP joint venture with Foxconn will take effect from 1 May, when we hope to find out more details about what it will entail.