(30 votes, average: 4.93 out of 5)
Loading ...Last week, I kicked off the first round of “20 Questions” by answering questions around our upcoming server platforms. This week I’ll focus on questions I’ve received related to virtualization and cloud computing.
When AMD pushed further into “Cloud\VM” computing there was talk of enabling GPUs as APUs in “future” sockets. Where is that initiative and where is it going?
Accelerated Processing Units (APUs) are essential to AMD’s product strategy and roadmaps. We still continue to see the potential for GPGPUs handling certain types of server workloads. However, this technology faces some challenges before becoming “the next big thing” for cloud clusters. The first is the programming model. One way we plan to address this is through the work that our ATI Stream team is doing around OpenCL, (ATI Stream technology is key in merging the GPU and CPU to form an APU)a framework for writing applications that execute across heterogeneous platforms consisting of modern multi-core CPUs, GPUs, and other processors
The second challenge is the power and cooling aspect of the solution. Customers are moving towards lower power solutions, especially in cloud, and CPUs are now approaching levels of performance/watt that still make them the leading choice for mainstream applications. The final area is the prioritization aspect. GPGPU is an emerging technology for servers. Integration into client processors has a much greater benefit for both customers and AMD so you will see much more emphasis on client integration first.
How will AMD address Green Computing?
We plan to continue to deliver low power processors, such as our HE and EE power bands. A new planned platform definition for Socket F, codenamed “Kroner” is focused as a “best practice for power management and design.” We expect to see a future “Kroner” follow-on platform that continues that thread. The C32 processors are really targeted at customers who want power efficiency, but we also plan low power options for the G34 processors as well. We are also planning enhancement to our AMD-P feature set with new technologies that are expected to help increase the power efficiency at the processor level, the platform level, and even at the data center level.
AMD has several initiatives targeting “cloud computing” and virtualization. Doesn’t this reduce the market for AMD products, since customers would buy only one server where they would have bought several?
Back when we are all in a single core world, everyone expected that dual-core would reduce the number of servers. It didn’t happen. Neither did the transition to quad-core. As virtualization and cloud computing help make IT more efficient, this can free up dollars for more investment in other areas. In any IT shop you’re likely to find that for every project that they are working on, there are dozens of other that they can’t fit into their budget. Add to this fact that world Internet usage is probably less than 25% today. As use of the Internet as a vehicle to deliver applications, data, and services continues to increase throughout the world, there is an obvious continued need for servers to power web and cloud clusters.
While AMD has made progress in the area of energy efficiency (e.g. AMD OpteronTM EE processors, Cool’n'QuietTM technology), does AMD feel there is more room for improvement in this area?
There is always room for more improvement. Power efficiency will continue to be an important focus for us.
John Fruehe is the Director of Business Development for Server/Workstation products at AMD. His postings are his own opinions and may not represent AMD’s positions, strategies or opinions. Links to third party sites are provided for convenience and unless explicitly stated, AMD is not responsible for the contents of such linked sites and no endorsement is implied.
#1 by Henry - August 4th, 2009 at 17:47
it is true that the design in each magnycours core will be same with current shanghai and istanbul (k10 core)? Is the chips will support new instructions like sse4b that comprises all new intel sse4?
#2 by John Fruehe - August 4th, 2009 at 22:11
The cores will be similar to what you see in the current processors, but they will not be exactly the same. There will be some additions to the core that will add some new functionality. I can’t go into those details now, but assume that Magny Cours will be similar but with new capabilities.
#3 by AJai dev - August 5th, 2009 at 15:21
What about a desktop hex core?? “965 is good for i7 920 but what about 950 and above??”
The Opteron 2427 Istanbul 2.2GHz costs $449.99 in newegg.
http://www.newegg.com/Product/…?Item=N82E16819105257
and the Core i7-950 Bloomfield 3.06GHz costs $569.99:
http://www.newegg.com/Product/…?Item=N82E16819115211
So why cant a Phenom x6 Istanbul 2.3GHz exist that costs around $500-$550??
Also will IPC be effected a lot with Magny Cours??
#4 by John Fruehe - August 5th, 2009 at 20:37
I am the server person, I can’t comment on desktop. They use a different socket and different infrastructure (1207 w/DDR-2 vs. AM3 with DDR-3).
There are always IPC tweaks with each new processor generation. The degree to which they impact performance is not something that we are releasing before launch.
#5 by Dante - August 5th, 2009 at 21:36
I have always wondered why it is so important to AMD to get the GPU onto the core of the CPU. For the time being couldn’t AMD leverage an onboard GPU and give it a hypertransport link to the cpu with cache coherency?
#6 by John Fruehe - August 6th, 2009 at 09:08
That is more of a client question, I am a server person. The key for getting the GPU into the CPU die comes down to having a better experience for end customers (better integration of graphics and processing for more capability) and better experience for the OEM (more efficient systems that are easier to design and bring to market.)
On the server side, you will see more heterogenous environments with our STREAM computing, especially now that OpenCL is coming to market, but these GPUs will be in add in cards vs. integrated into the CPU package because customers want the biggest, fastest GPUs and those will push the limits of packaging when included with a CPU in a single package. It makes more sense, based on physics and market demand to keep them outside of the server CPUs for now.
#7 by Stephen - August 5th, 2009 at 22:17
Magny-Cours looks promising, but I’m wondering what is connecting this MCM chip? Will it be HT3.0, 3.1 or higher? Will AMD maybe move HT aside and go for PCIe 3.0 for a faster and more common interconnect, especially when a GPU ends up a die? I know that will be mainly for desktop versions, but judging on past instances, it’s usually the server architecture that dictates the desktop design. Or, is this model going to change and say develope an ultra-effecient mobile part, and scale that up to the desktop and then server side of things? Will bulldozer be implemented on both G34 and C32? Ok, I’m done with the questions, and thanks again for this insight!
#8 by John Fruehe - August 6th, 2009 at 09:04
Desktop and server are different designs staring with Istanbul, so do not take any of the statements I am making to be a reflection of what is happening on the desktop side.
The MCM of Magny Cours will be connected with HyperTransport, which is the same industry-standard technology that we use to connect processors, chipsets and I/O devices together. PCIe 3.0 is a long way off, gen 2 is just now hitting the market. Having interconnected dies is critical for MCM scaling. Other companies have created MCM parts in the past by simply putting two dies in the same package. But if Die#1 needed to communicate with Die #2, it had to do so through the front side bus and the memory controller, creating more bus traffic and more bottlenecks. Interconnect MCM scales quite well, non-connected MCM does not.
When we get to bulldozer there will be 2 variants, one for the G34 infrastucture and one for the C32 infrastructure. Both will be compatible with the respective G34 and C32 sockets that you will see on the Maranello and San Marino platforms, so it should be an easy qualification for customers and a consitent platform. We have not released any other data on the bulldozer products, so you’ll have to wait for that