h1

Shanghai Economics 101

April 30, 2009

Before the release of the Istanbul 6-core processor we wanted to preview the CAPEX comparisons we’ve been working on between today’s Opteron (Shanghai) and today’s Nehalem-EP. The results are pretty startling and mostly due to the Nahelem-EP’s limited memory addressing capability. Here are the raw numbers for comparable performance systems (i.e. high-end):

Nehalem-EP Configuration Street $
Shanghai HT3 Configuration Street $
Savings $ Savings %
2P/8C, Nehalem-EP, W5580 3.2GHz, 6.4GT QPI with 24GB DDR3/1333 $7,017.69   2P/8C Shanghai, 2393 SE, 3.1GHz, 4.4GT HT3 with 32GB DDR2/800 $5,892.12   $1,125.57 16.04%
2P/8C, Nehalem-EP, W5580 3.2GHz, 6.4GT QPI with 48GB DDR3/1066 $7,755.99   2P/8C Shanghai, 2393 SE, 3.1GHz, 4.4GT HT3 with 48GB DDR2/800 $6,352.12   $1,403.87 18.10%
2P/8C, Nehalem-EP, W5580 3.2GHz, 6.4GT QPI with 96GB DDR3/1066 $21,969.99   2P/8C Shanghai, 2393 SE, 3.1GHz, 4.4GT HT3 with 96GB DDR2/667 $11,968.72   $10,001.27 45.52%
2P/8C, Nehalem-EP, W5580 3.2GHz, 6.4GT QPI with 144GB DDR3/800 $30,029.19   2P/8C Shanghai, 2393 SE, 3.1GHz, 4.4GT HT3 with 128GB DDR2/533 $14,300.92   $15,728.27 52.38%
               
2P/8C, Nehalem-EP, W5580 3.2GHz, 6.4GT QPI with 96GB DDR3/1066 $21,969.99   4P/16C Shanghai, 8393 SE, 3.1GHz, 4.4GT HT3 with 96GB DDR2/800 $17,512.87   $4,457.12 20.29%
2P/8C, Nehalem-EP, W5580 3.2GHz, 6.4GT QPI with 144GB DDR3/800 $30,029.19   4P/16C Shanghai, 8393 SE, 3.1GHz, 4.4GT HT3 with 192GB DDR2/667 $28,746.07   $1,283.12 4.27%
2 x 2P/8C, Nehalem-EP, W5580 3.2GHz, 6.4GT QPI with 144GB (288GB total) DDR3/800 $60,058.38   1 x 4P/16C Shanghai, 8393 SE, 3.1GHz, 4.4GT HT3 with 256GB DDR2/533 $33,410.47   $26,647.92 44.37%

Even the 4-socket Shanghai 8393SE averages 23% lower implementation cost over Nehalem-EP and produces 16 “real” cores versus 8 “real” cores in the process. Even at 50% theoretical efficiency using Nehalem’s SMT, the 4P Shanghai represents a solid choice in the performance segment. An Istanbul drop-in upgrade spread’s the gulf in capabilities even wider.

Based on today’s economics and the history of seamless vMotion between Opteron processors, 4P/24C Istanbul is a solid will be a no-brainer investment. With 2P/24C and 4P/48C Magny-Cours on the way to handle the “really big” tasks, a Shanghai-Istanbul Eco-System looks like an economic stimulus all its own.

11 comments

  1. Updated: corrected typos and added emphasis to slight configuratin differences in table data (i.e. RAM and system qty.)

    Like


  2. Comparison based on similarly configured systems from Tier-1 barebones providers.

    Nehalem-EP: Supermicro 6026T-NTR+, 8xSAS Hotswap, 720W Redundant, 2U

    Shanghai: Tyan TA26B2932-SI, 8xSAS HotSwap, 600W Redundant, 2U

    Like


  3. I checked your link and your claim that DDR3 costs $10,000 is blatantly dishonest and assumes people are dumb enough to buy from the most expensive vendor that marks up prices 800%.

    I ran a quick google search and found http://www.provantage.com/kingston-technology-kvr1… 4GB DDR3 DIMMs with ECC for $102.26. That means with 18 4GB DIMMs for a total of 144 GB of RAM, you’ll spend $1840.68 on all new memory.

    Like


    • George: please check your math before claiming that we’re being dishonest. You can not make 144GB using 18 4GB DIMMs. It requires 18 units of 8GB DIMMs to create a 144GB pool. These are available for $1,168/each directly from Kingston:

      http://shop.kingston.com/partsinfo.asp?promo=PRCGRBR&ktcpartno=KVR1066D3Q4R7S/8G

      If you can find a much better deal, please let us know and we’ll update our statistics. While you can make 72GB with the 4GB parts, this is NOT what the comparison systems (VMmark, etc) are using. Note, as we disclose, that this memory drops to DDR3-800 in the 18-DIMM configuration – significantly lowering Nehalem-EP’s performance marks.

      (Note: our 72GB Nehalem-EP system is based on the $138/DIMM part.)

      So, we can either print the truth about how much “tested” performance costs, or skew the parts to accommodate cost and have no reliable indication of performance (about 25% would be appropriate based on AnandTech’s statistics). Again, this changes the effective claims of the vendor.

      Like


  4. It doesn’t make much sense to buy either DDR2 or DDR3 8GB DIMMs these days. You’re looking at $580 for DDR2 and $1030 for DDR3 8GB sticks though the prices will come down dramatically within months. For now, it costs more than the actual server to buy that density of memory and it’s going to stay that way for a while.

    So it really depends on what you’re looking for. If you’ve got to have that 144 GB VM host, it will be a little less insane with an AMD DDR2 platform. However, a lot of server architects prefer the hybrid approach to virtualization where you don’t buy into it wholesale. Heck, most shops aren’t even using a whole lot of virtualization. You might put 90% to 95% of your servers in virtualized environment and you keep the most CPU/IO/Memory intensive servers on bare metal. For those applications which cost $10K to $100K per CPU socket, I’m not really in to mood to share any hardware which is cheap by comparison. I want that application to be running on dedicated hardware.

    The bottom line is that I don’t really need a VM host with that much memory. I’m perfectly happy doing an 8 to 1 or 16 to 1 consolidation of physical servers.

    Like


  5. Thank you, George, and we’re not advocating the purchase of “outrageous” memory configurations – it’s economics, and that was the point of the blog entry. If you read our follow-up, you’ve seen where the right balance between price and performance come for either Intel or AMD platform.

    However, when you start to talk 8:1 or 16:1 consolidation – and I’ll assume that is VM:HOST not vCPU:Core then you are correct: 16 VM’s per Nehalem-EP host is about the sweet spot (assuming average of 1 vCPU and 1.5GB memory.

    What we’re demonstrating is the “awesome consolidation ratios” enabled by multi-core densities pushes the economics of memory beyond the breaking point. If you use vCPU:Core that limit is 30-40 VM’s per host – memory allowing.

    As core begin to multiply, memory exhaustion begins to become a factor in consolidation (just like earlier Intel architectures limited to 6-12GB) and not CPU availability. Finding the right balance is a problem for the consultant.

    We agree that today’s architecture lends itself towards hybrid environments: retaining bare metal for specific, performance sensitive applications. Regardless of your system foundation, performance is cheaper today that it ever was. However, we’re quickly approaching 100% virtualization – is that 99.999%? Probably, but the cost of NOT virtualizing a platform is growing steeper as companies become more entrenched in virtualization eco-systems.

    Like


  6. Just one year ago, 8 GB DIMMs were at some insane level of cost that would have cost as much as a house if you tried to buy 144 GBs. Today, it’s no longer insane but it’s no bargain either. By the time 144 GB configurations are routinely necessary, it will be reasonably affordable and 32 GB DIMMs will be the insanely priced DIMMs.

    I don’t see the penalty of not going 100% on everything. In fact, I see a huge penalty for insisting on purity for the mere sake of purity. Some might argue that bare metal configurations are not portable and can’t be backed up, but there are tools that let you image bare metal configurations and allow you to migrate those images to different hardware.

    The fact of the matter is, most people aren’t even doing virtualization on the majority of their systems yet so it’s quite silly to be worrying about 100% virtualization.

    Like


  7. […] 101 – Continued Shanghai Economics 101 – Conclusion May 6, 2009 In the past entries, we’ve looked only at the high-end processors as applied to system prices, and we’ll continue to use those as references through the end of […]

    Like


  8. […] Shanghai Economics 101 […]

    Like


  9. […] path to Intel’s (rip and replace) where VMware infrastructures are concerned. That article, Shanghai Economics 101, was one of our most popular AMD-related postings yet, and – judging from what we’ve […]

    Like


  10. […] he finally comes to the same conclusion we’ve presented (in an economic impact context) about Shanghai’s virtualization value proposition. While his results are consistent with […]

    Like



Comments are closed.

%d bloggers like this: