500,000 cores already shipped since September

AMD launched its much anticipated Bulldozer architecture for the consumer market last month, but many were disappointed at the performance numbers. Now the company has officially launched new processors using the same architecture for the server and workstation markets, but things have changed significantly.

The key difference is in the software used to process instructions. The consumer side is reliant upon Windows 7 and earlier operating systems, which are unaware of the shared nature of the Bulldozer architecture. Resource sharing is inefficient at best, and the full possibilities of higher Turbo Core frequencies are missed.

AMD has worked to ensure optimization and/or support on many commonly used server operating systems. Linux 2.6.37, Windows Server 2008 R2 SP1, Xen 41, Ubuntu 11.04, and VMware vSphere 5.0 already have OS hypervisor support for Bulldozer, while others such as RedHat Enterprise Linux 6.2 and Windows 8 Server are currently in development.

AMD is specifically targeting the High Performance Computing (HPC) segment, with over 500,000 Bulldozer cores already shipped to this market since September. The AVX, FMA4, and XOP instructions require software to be recompiled in order to take advantage of their performance enhancements. Java 7 was mentioned as a program that was being worked on.

The Opteron 6200 series was formerly codenamed Interlagos. It is scalable to 4 sockets supporting 16 Bulldozer cores each. The fastest model is the 6282 SE at 2.6Ghz, with a maximum Turbo Core frequency of 3.3GHz and a TDP of 140W. The Opteron 4200 series was formerly codenamed Valencia. It is the most similar to the FX series (Zambezi) launched in October, but it will support up to 2 sockets with 8 cores each.

Both series support DDR3-1600 memory natively, but there will be official support for DDR3-1866 through specific OEMs. Opteron 6200 CPUs have quad memory channels, while the Opteron 4200 chips have dual channels. 1.35v low voltage memory and 1.25v ultra-low voltage memory is also supported, as are Load Reduced DIMMs (LRDIMMs).

The L1 cache is arranged as 16KB data per core and 64KB instruction per module, while the L2 cache is 1MB per core. Opteron 6200s have a shared 16MB of L3 cache per socket, while Opteron 4200s only have a shared 8MB per socket.

In order to speed time to market and lower validation costs, AMD has designed its new Opterons to function on its previous platforms using the G34 and C32 sockets. The company believes that its lower total platform costs over Intel’s Xeon platforms impart a significant advantage. For example, the AMD Opteron 6276 will ship at the same price as the Xeon E5640, but will outperform it by 89%.

Cloud computing requires high throughput, scalability, density, and power efficiency. AMD thinks that it can gain significant market share by claiming the lowest x86 watts/core in the industry at 5.3W for Interlagos and 4.375W for Valencia. The new C6 power state reduces power consumption at idle by up to 46% over the previous generation by enabling core power gating When a core is halted, its context is exported to system memory and voltage is removed from the core.

Intel will be launching new server and workstation products based on the Sandy Bridge architecture next year, but AMD also has plans for the future with its Piledriver architecture. Sepang will use the C2012 socket and replace the Opteron 6200 series, while Terramar will use the G2012 socket. Both new platforms will support PCIe 3.0.


[Submitted by Christo [PCD]]



Click to discuss this news item in the forums

Samsung, Hynix, and Micron preparing for the next generation

DDR3 SDRAM is now the standard memory for servers and workstations, but its supremacy was never assured. The introduction of Fully-Buffered DIMMs was slated as a substitute, but its high costs and power consumption heralded its doom. 1333MHz has been the top speed for servers using DDR3 during the last two and a half years, but things are about to change. A number of new product launches from Intel and AMD are leading to some faster memory from Samsung, Hynix, and Micron.

Although DDR3 was first introduced in 2007, it didn't make its way to servers until March 2009. Intel's 5500 series servers were the first to make use of DDR3's higher clocks speeds and lower power consumption, while AMD followed with its Socket AM3 quad-core Opterons, code-named "Suzuka."

AMD's FX series of CPUs using the Bulldozer core are the first to officially support DDR3-1866 at stock speeds. Although its performance was not as high as anticipated, there is a lot of interest in using the FX series in the workstation market due to ECC support in its dual channel integrated memory controller. Intel's Xeon processors are dominant in that market, but they are much more expensive.

Samsung is the largest DRAM manufacturer in the world, and is currently mass producing several DIMMs to address this opportunity. The most interesting is an 8GB unbuffered ECC DIMM (M391B1G73BH0-CMA), which uses 4Gb chips running at DDR3-1866 (PC3-14900). A fully populated AM3+ motherboard like the ASUS Crosshair V Formula would be able to address 32GB of this RAM. 4GB (M391B5273DH0-CMA), 2GB (M391B5773DH0-CMA), and 1GB (M391B2873GB0-CMA) capacities will also be available.

Hynix is currently sampling a similar 8GB ECC module (HMT41GU7MFR8C-RD), but 4GB (HMT351U7CFR8C-RD) and 2GB (HMT325U7CFR8C-RD) DIMMs are already in mass production. Micron, the last U.S.-based DRAM manufacturer, has similar products in the works as well. They declined to comment at this time, but they do have 2Gb DDR3-1866 chips in mass production. All three firms are utilizing a CAS latency of 13 at this speed.

AMD's upcoming 16 core Interlagos servers and some of Intel's Sandy Bridge-E server CPUs will feature quad-channel DDR3-1600 (PC3-12800) integrated memory controllers. This is the most lucrative market for Samsung, Hynix, and Micron. All three already have 16GB and smaller capacity ECC Registered DIMMs in mass production at that speed. Hynix is sampling a 32GB DIMM, while Micron has the only DIMMs with a CAS latency of 10.

Lower voltage (1.35v) versions of these DIMMs will be the most in demand due to their reduced power consumption. However, the largest servers will use Load Reduced DIMMS (LR-DIMMs). These are the successors to FBDIMMs, but without the problematic serial interface. In addition to the address signals normally buffered by registered DIMMs, LR-DIMMs also buffer the datalines. This allows the number of memory chips that can be used per channel to be greatly increased at the cost of additional latency. Servers using as much as 2TB of memory have been proposed using this technology.

Future server CPUs based on Ivy Bridge and Piledriver architectures are expected to make use of DDR3-1866 ECC Registered DIMMs or LR-DIMMs, and all three firms have plans for those as well. DDR3-1600 is a 20% increase in memory bandwidth over DDR3-1333, while DDR3-1866 is a 40% increase. Given the pace of memory advances in the server and workstation markets, these will have to do for the next couple of years.


[Submitted by Christo [PCD]]



Click to discuss this news item in the forums

The new future for Intel?

“Intel is transforming from a leading server technology company to a comprehensive data center provider that offers computing, storage and networking building blocks,??? said Kirk Skaugen, Intel vice president and general manager, Data Center Group. “Fulcrum Microsystems’ switch silicon, already recognized for high performance and low latency, complements Intel’s leading processors and Ethernet controllers, and will deliver our customers new levels of performance and energy efficiency while improving their economics of cloud service delivery.???

Looks like it could be a really good deal to me! I say watch this space as Intel as has the resources and man power to make news in the industry with this acquisition!

Source
[Submitted by Siversmith[PCD]]

Click to discuss this news item in the forums