by Anton Shilov
03/29/2010 | 01:14 PM
Advanced Micro Devices on Monday unveiled its new lineup of server microprocessors designed for 2-way and 4-way servers. For the first time in recent history of servers AMD’s new Opteron chips for dual-processor and multi-processor servers are priced equally. Even though the move is somewhat unexpected, AMD claims that it makes a great sense to make multi-processor servers more affordable.
In order to find out more about AMD’s new server plans, X-bit labs recently asked John Fruehe, director of product marketing at AMD’s server and workstation division.
X-bit labs: You are claiming that by the launch of AMD Opteron 6000-series processors and Maranello/G34 platform you are removing something what you call the “4P Tax”. But does it make sense to remove this tax at all?
John Fruehe: Yes it does. The 4P market is in decline. It used to be 10% of the total market, and now it represents only 4% of the total market as 2P servers have accelerated their performance and reliability. If the 4P market is going to change, something needs to happen. Left alone, it may whiter away to only a point or two. We have had a lot of customers approach us with a desire to purchase more 4P capable systems, but because of the price premium (for generally identical silicon), they have been hesitant to buy 4P. Instead, they compromise and choose 2P because of the more favorable price. We believe this change makes sense today, just as, in the late 90’s, the reduction in 2P price points drove a large increase in units.
X-bit labs: What is the difference in pricing between 2P and 4P platforms for G34 processors? Just a very rough number, if you can.
John Fruehe: Today, you can buy a top bin 2P processor for ~$1100, but the top end 4P capable processor is ~$2500. This is more than a two times premium for what is, essentially, the same silicon.
X-bit labs: Who do you think will be the new customers for relatively low-cost 4P systems?
John Fruehe: One of the first places will be in HPC. Today customers are buying 2P platforms because they are much lower in price. However, when you buy two 2P systems, you are also buying two interconnect cards (which can be very expensive) as well as doubling the number of fabric switches. In addition, there are two chassis, two power supplies, as well as the power to run the two nodes. A single 4P node would reduce the interconnect costs by half, and would make cluster management easier as you would have half as many nodes to manage. In addition, virtualization customers are buying 2P platforms more than 60% of the time, because of the cost. However, they would really appreciate being able to get into a 4P to take advantage of the larger numbers of corese.
X-bit labs: Don’t you think that more cost-efficient 4P systems will actually affect high-end 2P systems?
Have a look at the following:
- Two AMD Opteron 6176 SE (12-core, 2.30GHz, 105W) cost $2772 and the customer gets 24 cores, up to 24 DIMMs, and 210W TDP.
- Four AMD Opteron 6134 (8-core, 2.30GHz, 80W) cost $2092 and the customer gets 32 cores, up to 48 DIMMs, and 320W TDP.
The cost of such 4P system is likely to cost as much as the cost of a high-end 2P machine (or even lower!). But they have greater scalability and performance due to higher memory bandwidth.
John Fruehe: Exactly. When you look at that difference in silicon cost it is clear that some of that price difference helps offset any higher costs for systemboards, power supplies and other platform components. The key is that those four mid-bin processors will perform better than two top bin processors. Do we think this will impact the 2P market? Of course! But the difference is that we own the platform with these tremendous economic benefits and our competitor does not. Even if they were to try to match our pricing, they would be far behind us because their 4P solution (Nehalem EX) has high infrastructure costs and high power consumption, so the economics are not necessarily in their favor.
X-bit labs: Do you think that it makes sense to actually lower the pricing of server processors? Large customers can afford almost any price. Smaller customers are hardly considering 4P due to various reasons. So, does it make sense to lower your own profit margins by removing the so-called 4P tax?
John Fruehe: I have never run into a customer that wanted to pay more for their processors. If you do find some, please let me know. When prices go down, customers buy more. Look at the market after dual core hit. Did customer purchases drop by half? No, they increased. Customers have IT budgets and by maximizing their budgets they can purchase more technology. That is a good thing.
X-bit labs: You claim that your new platform has greater memory “design” per socket. However, considering the fact that that the DCA 1.0 server platform has 6 cores and up to 8 DIMMs per socket, whereas the DCA 2.0 server platform has 12 cores and up to 8 cores per socket, it seems that this is a step back… What do you think?
John Fruehe: I disagree. What matters to customers are two clear vectors: total memory footprint and total memory bandwidth. With our new products, you actually get more of both so it is a step forward. Our old platforms supported DDR2 memory, and 4GB DIMMs were the most common; 8GB DIMMs were disproportionately expensive. With our new products, customers will use DDR3 memory. 4GB DIMMs are cost effective, 8GB DIMMs will soon be cost effective and 16GB DIMMs will eventually be available. So memory scalability is much greater with 12 DIMMs per socket vs. our old 8 DIMM architecture. Considering the “maximum affordable memory per core”, when 8GB DIMM prices come in line, that will be 8GB. With our older platforms, eight 4GB DIMMs meant a little more than 5GB per core. Speed of DDR3 is 1333MHz vs. the DDR2 speed of 800MHz as well.
X-bit labs: You introduced average CPU power (ACP) metric back in 2007 in order to show the difference between CPUs with and without integrated memory controller. However, you are now comparing TDP of Intel processors with integrated memory controller against your ACP. Don’t you think it is somewhat unfair (especially considering the fact that you do provide TDPs for your platform partners)?
John Fruehe: TDP means something different for both of us, that is why TDP has been so difficult to use as a comparison. For years, TDP was our max power and for our competitor it was a mere fraction of their total power. I always recommend that customers focus on power at the wall instead. A cheap $20 power meter will tell you more about the actual consumption than all of the spec sheets in the world.
X-bit labs: Thank you, John!