Playing 20 Questions (Part 3)
To wrap up the first round of “20 Questions,” below are additional questions I received. In the first post, I addressed questions around our upcoming server platforms. The second post focused on virtualization and cloud computing. And the final post will answer questions related to future technology trends/adoption.
If you have additional questions or subjects you’d like me to address, let me know and we’ll play another round of “20 Questions.”
Are there any plans to support the HyperTransportTM Technology High Node Count (HNC) Specification in future CPUs?
HyperTransportTM technology is a key cornerstone of the AMD OpteronTM processor and is expected to be integrated into future AMD Opteron processors. HyperTransport High Node Count (HNC) is a very interesting new specification that supports the development of highly scalable systems. HyperTransport technology is an open industry standard and the HNC specification was developed in conjunction with various HyperTransport consortium members. While not integrated into our public roadmaps at this time, it is a technology in which we see potential.
Is there any news on Torrenza?
Torrenza was an umbrella for a wide range of programs, many of which continue today under our Accelerated Computing initiative. Accelerated Computing takes into account the hardware and software evolution necessary for new combinations of integrated and discrete products designed to deliver a superior user experience across a broad range of usage scenarios. Utilizing the GPU for certain computationally intensive workloads via ATI Stream is an example of that, as are new software development tools like OpenCL that makes it easier to take advantage of the CPU and GPU capabilities in a system. Many of the initial Torrenza technologies were based on the Socket F, so as we move to new sockets (G34 and C32) with Accelerated Computing, we may see different implementations, whether it is socket-based or not. Just like the rest of the market, AMD continues to evolve as we learn more about how customers want to solve computing problems.
Does AMD foresee convergence of desktop and server products/platforms in the future, perhaps when CPU power is no longer the limiting factor in software progress?
Actually, we are seeing the opposite. The biggest driver for change in servers that we are seeing is the need for lower power consumption, not higher power parts. As we look at future processors, we expect to continue to see the divergence of the desktop and server roadmaps. In a desktop, the impact of a higher power processor is less profound than in a server where multiple CPUs can be consuming much more electricity.
Are there plans to release the current Six-Core AMD OpteronTM processor in a quad-core variant in the future?
There are currently no plans, but there are plans for quad-core models of the C32 processor in the first half of 2010.
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.
POSTED IN: AMD Opteron
TAGS: Accelerated Computing, ATI Stream, Socket C32, Socket G34
SHARE:
COMMENTS: 13
I want to use new upcoming AMD platform using G34 socket. Is there any information from AMD to get incentives to existing AMD’s Server Processor user if they want to upgrade their servers early?
The cuurent products have a different socket than the upcoming G34 products.
I recommend getting in touch with your local AMD representative to discuss options.
John Fruehe, you still have not replied to a question, when you make a processor as automatic AK-74. Disassemble the machine and see if you have a processor must also work, namely, when he was mistaken, and it should send the information to the kesh.
You must remove the brake:
1. Kesh second level.
2. Make the exchange between the nuclei. When error is not returned in kesh, and analyzed, the rest of the kernel must fit all you need.
This is the increment. I see it as a processor works and what it lacks. I just give you advice that can be improved without losing time on the development of a bulldozer. You have a new threat I5, as you are going to fight with him??? You need a modified architecture. Wa does not squeeze out of the phenomenon.
Unfortunately, something is being lost in translation. I speak English, some German and a little Spanish. However, your question appears to be translated through some type of online translator and it is not making sense to me. If there is a way that you can find someone to do the actual human translation to either English or German, I might be able to answer it. Unfortunately the machine-level translation from Russian to English is not working well for a technical discussion.
I think he’s says decrease L2, make communication of Core and L3 paramount and do everything else that does not fit in the L2, should be taken care of in the OS environment.
Our engineers have spent a lot of time evaluating the relative cache sizes, cache latencies and caching algorithms in order to choose the optimal configuration for our architecture.
You have to consider die size as well in that equation. When I was in college, and studying economics they would always hold one or two variables constant in order to reduce the matrix. But in real life, you have to think about the entire architecture. Varying only the 2 cache sizes is a very simplistic approach, but there are literally hundreds of variables that you have to take into consideration when making that decision.
Your note on HNC is intriguing, i have read about future of QPI an note similarities.
On the whole quad core thing i wanted to ask why. Why make a quad core when the hex core is quite effective as it is. Quad core consume almost the same as a hex core in ideal and only when loaded hex cores eat a bit more. But in a hex core 2P system vs a 4P quad core system i am 100% sure the 2P would win in the energy department.
I dont see the point of having a 4P EE Quads when almost the same amount of work can be done with a normal 2P Hex ??
“As we look at future processors, we expect to continue to see the divergence of the desktop and server roadmaps.”
Does that mean that the current trend is going to continue/accentuate? I’m referring to the fact that currently, AMD’s best desktop CPU is high-frequency 4-core processor, while the best server CPU is a moderate-frequency 6-core one, and a low-frequency 12-core one is coming.
Well, since I am a server guy and not a client guy, I will give you a response from my point of view. Client workloads tend to be far less threaded than server workloads. I watch my CPU all the time at home, and with the exception of transcoding a video, I rarely see more than 2 cores spiking at any given time. I am hardly a power user, but my workloads probably represent a majority of the market. I noticed a big performance going from 2.3GHz to 2.9GHz; the move from dual core to quad core (which was lower frequency) gave me far less performance improvement.
We’ll be moving to 8 and 12 cores, those designs definitely favor servers over clients, as there is an inverse proportion between clock speed and core count. The higher the core count, generally the lower the clock speed. A 3.4GHz quad core will run circles around a 2.6GHz 6-core for most client workloads, but when it comes to a well-treaded server workload, expect the 6-core to have the advantage.
Beyond the Magny Cours product, our next core, bulldozer will be a completely new architecture. Both server and client will leverage this core, but in different ways. Don’t expect to see similar dies between the server and client products, but they will continue to use the same underlying archtiecture because we strongly believe that the core logic of our processors is a significant competitive advantage.
Hence the modularity of the design, I suppose.
It’s interesting how for a time, it seemed that client and server CPUs were converging and now, they’re diverging again.
Thank you for taking the time to answer.
There will always be pieces of the technology that are leveraged across the server and client processors – it would be crazy not to take advantage of the technology that the other teams have. However, as discrete silicon designs, you’ll see server and client on two different tracks. Their cadence and customer needs are different than ours. When we get to bulldozer, you’ll see us both leverage the technology together, but our modularity will allow for us to take a path that focuses on more cores and greater threading (up to 16 cores) and their path will focus more on higher clock speeds.
Your strategy of partitioning the market into C32 and G34 processors seems quite reasonable, but there is a detail that was not mentioned and that is very important for your customers. There are indeed a lot of cases where C32 processors will be the right choice, but there are also the cases for which your customers are considering Xeon 5500, with which C32 processors cannot compete in performance due to their slow memory interface. G34 processors might have adequate performance, but we do not know if they will also have an adequate price. The reason is that you say G34 will replace both the 2000 and the 8000 series but you do not say how the price of G34 will be positioned compared with the current two series. If AMD will not offer, for example, a G34 processor with at least 2.6 GHz and at no more than $1000, then it will not be able to compete with Xeon. What I mean is that if the prices of the G34 processors will be at the level of the 8000 series, then they will not be competiitive for most applications.
We never release pricing or bin speeds prior to launch, so I cannot comment on those. However I will say that I firmly believe that we will be very competitively priced relative to our competitor.
When you say that we must be at 2.6GHz are you considering that we will have both 8 and 12 core processors? One will have 33% more cores, the other will have 50% more cores than Intel’s product at the time. As for the $1000 price tag, with the average server processor selling closer to the $300-600 range, I think the $1000 price point is fairly optimistic. Few buy 2P processors in the $1000 price range.