David Kanter (dkanter@realworldtech.com) on 1/5/12 wrote:
---------------------------
>George Baker (george_baker@yahoo.com) on 1/4/12 wrote:
>---------------------------
>>Intel's announced pricing for dual-socket 8-core SandyBridge-EP ranges from $1106
>>(1.8 GHz) to $2057 (2.9 GHz) per chip!
>
>That sounds about right.
>
>>There is little benefit of x86 compatibility
>>in servers
>
>What makes you say that? Have you ever tried to run SQL server on PowerPC? Or VMware?

Other architectures have their own virtualization tools (PowerVM is excellent) and DBMS. For the moment, MS SQL Server runs on both x86 and IPF.

>
>x86 compatibility is fairly useful.
>
>Moreover, what is the pricing of comparable chips using other ISAs?
>
>1. PowerPC - much more expensive
>2. SPARC - Lower performance
>3. Itanium - Probably lower performance
>4. ARM - Lower performance (servers don't currently exist)

Poulson and T5 should be comparable to or faster than contemporary multi-socket Xeon processors. T5 also adds support for 8S systems. T4 has done well on some benchmarks as well, although the methodology has been criticized.

Additionally, Power Systems really aren't significantly more expensive than xSeries - take a look at 720 vs x3620, for example. Other commodity vendors are admittedly cheaper.

>
>>so this market seems ripe for a transition. Nvidia is working on Project
>>Denver, which is an ARM chip for servers.
>
>Unfortunately sprinkling magic ARM dust on a chip doesn't make it high performance.
>It merely makes it ARM compatible. There is no reason to believe that Nvidia can
>design chips with comparable performance to AMD, let alone Intel.

They've done okay with graphics and HPC chips. If nothing else, they should have expertise with multicore designs and high-performance memory subsystems.

>
>>I think other companies will make similar
>>chips (perhaps even AMD).
>
>AMD won't. They have everything to lose and nothing to gain. Did you see my article about why AMD won't use ARM?

If they don't stay competitive in x86, what are their options? Pulling out of the enthusiast and server industries entirely? I agree that AMD ARM is highly unlikely.

>
>>As the economic benefit of Moore's Law slows down,
>
>It's not clear that the economic benefit of Moore's Law is slowing down at all.
>
>>companies
>>will have no choice but to use more efficient chips to provide continued price/performance
>>improvements. It wouldn't surprise me if the majority of >annual server shipments use ARM processors within 5 years.
>
>That's impossible. There aren't even any credible ARM server chips coming out till 2013.

>
>>AMD has intellectual property and expertise for cache-coherent HyperTransport,
>>PCI Express, caches and high-speed DRAM interfaces. All of these could be applied
>>to an ARM server chip.
>
>Or you could apply it to an x86 chip and make more money. It's pretty obvious what AMD will do...
>
>>In an ideal world, AMD and Nvidia would work together on
>>such a chip and would be second sources for each other on >the finished product.
>
>Why would anyone do that? AMD has far more experience with server CPU design and
>has no incentive to share the expertise. Moreover, Nvidia doesn't have an x86 license and AMD cannot grant one.
>
>>Perhaps they could get some funding from Google and/or >Facebook for developing this chip.
>
>Why? Both facebook and google seem happy with existing Intel and AMD solutions.

Facebook has displayed significant interest in Tilera.