Sockets, cores and threads, oh my
Intel's silicon brick roadmap
Intel gave us lucky hacks a post-Nehalem launch briefing to outline aspects of its near-term processor roadmap. Here's a fast-paced tour through the briefing content.
Intel reassured us that it could see Moore's Law progression being maintained for another ten or so years with the process shrinks and other developments in its roadmap. The inexorable rise in transistor counts on dies, 731 million with Nehalem quad-cores, should continue over that period. The big issue is people comprehending how to use parallelism.
Intel Server Processors
Intel's server platforms groups covers several markets: high-performance computing (HPC), internet data centres and the cloud, RISC migration, storage and the virtualised data centre. Its building blocks are CPUs, chipsets, the LAN, SSD/RAID, boards/KDKs and software. Its processor platforms are Itanium (9000 series) for mission-critical servers, Xeon 7000s for high-end enterprise servers, Xeon 5000s for mainstream enterprise quad- and dual-socket servers, and Xeon 3000s for single socket SME servers. Obviously Xeon 7000s and Itanium compete in everyone's eyes but Intel's.
Much was made of Intel designing in virtual server support features to hopefully get apps running in virtual machines running as fast as apps running in a non-virtualised server.
We had a quick romp through the Xeon 5500's features and then got to the interesting bit, Westmere-EP, the next-generation Nehalem (5000 series) processor built with a coming 32nm process. Westmere is basically a Nehalem shrink onto the 32nm process.
It will have six cores, cache enhancements and better energy efficiency, and will be announced in 2010. Also it will be socket-compatible with the 5500, so server vendors can use existing server boards and plug Westmere processors straight in. The graphic showed a PCI 2 bus linking Westmere to an Intel X25-E SSD, and an Intel 10GbitE controller.
Intel has a tick-tock analogy for its engineering developments. A "tick" is a process shrink using the current micro-architecture. The last tick was the move to the current 45nm process technology. A "tock" is the introduction of a new processor micro-archictecture and Nehalem is the latest tock event.
It will be followed by the 32nm tick which will, in turn, be followed by the next micro-architecture after Nehalem.
On the 7000 front the 7400 is a 6-core 45nm processor that's socket-compatible with the previous 7300. It will be followed by Nehalem-EX in the second half of this year and then Westmere-EX some time later.
Nehalem-EX, built on the 45nm process, is an 8-core processor designed for 4-socket platforms, meaning 64 threads (2 threads/core still). It will use an Intel Scalable Memory Interconnect with Buffers and what Intel called next-generation I/O technology with virtualisation.
There will be Intel VT technology in the processor (VT-x), the chipset (VT-d) and for the network (VT-c). The idea is to get these components working better together so that server vendors will go all-Intel instead of using third-party chipsets and communications interfaces. Thus there's no need for a Fibre Channel over Ethernet adapter, a CNA, because these VT features plus Intel's software FCOE initiator will get the Intel 10GbitE NIC spurting out FCOE packets at near wire speed to direct-attached devices.
There will be Virtual Machine Device Queues (VMDq) to improve throughput by offloading data sorting to the NIC. In a virtualised server the virtual machine (VM) will more or less talk directly to the NIC without having to go through the hypervisor. (Naturally that means the individual VMs in the server use CPU cycles to do the FCOE work though. Why Intel doesn't go for FCOE-offload is beyond me. It tried to do the same thing with TCP/IP offload, saying there was no need, and pretty much lost that argument.)
Intel has introduced Flex Migration so VMs can be readily transferred between different Xeon servers. Previously, notwithstanding instruction set compatibility between 3000s, 5000s and 7000 Xeons, VMs could only be moved between similar processor types.
We can expect a Nehalem version of the 3300, probably the 3400 and probably later this year or in 2010.
Next page: Desktops
COMMENTS
Itanium taking share from IBM and Sun
Itanium fills a role at the head of the high-end server class, despite its detractors there is no better option for mission critical computing. Read more at: http://blog.itaniumsolutions.org/
@Mad Mike: Power 6+ & 7 will probably be screamers!
I guess I was wrong... The IBM Power6+ was not a screamer...
http://www.theregister.co.uk/2009/04/28/ibm_shipped_power6_plus_already/comments/
Silently released, no significant performance boost.
RE: Projected Intel, IBM, AMD, and Oracle/SUN/Fujitsu processors
Matt Bryant says, "And the Nehalem and Barcelona cores are full-bodied cores with a real pipeline to keep those threads spinning"
But the cores are not always processing, waiting for memory accesses on cache misses.
Matt Bryant continues, "whereas the whole Sun design is a capitulation to the idea that you can't have all the threads going at once, because Sun couldn't design a bus or core powerful enough to"
Actually, Sun designed a core powerful enough to run an instruction for every clock cycle, not stalling. The difference is that when the Sun thread stalls, it does not stall the core, while with other processors, the core stalls.
The Sun design was brilliant, attacking the core not running 100% of the time, using a different strategy to keep it working. It was a very unique solution in the marketplace. It does not work well for all workloads, but it is superior for many workloads. (i.e. on web servers, a single socket T2 processor will handle encrypted traffic almost as well as a quad socket Intel machine!)
Matt Bryant suggests, "T3 will still have the same scale problems of all the Niagara lines"
A single socket octal core T2 performing slightly slower than a quad socket hex core Intel is not a bad scaling problem to have - especially when the T2 will scale to 4 sockets... and easily outrun 12 hex-core Intel sockets in encrypted middleware bus or web-server loads.
http://spec.org/cgi-bin/osgresults?conf=web2005&op=fetch&proj-COMPANY=256&proj-SYSTEM=256&proj-PEAK=256&proj-HTTPSW=256&proj-CORES=256&proj-CHIPS=256&critop-CHIPS=0&crit-CHIPS=1&proj-CORESCHP=256&proj-CPU=0&proj-CACHE1=0&proj-CACHE2=0&proj-CACHE3=0&proj-MEMORY=0&proj-NETNCTRL=0&proj-NETCTRL=0&proj-NNETS=0&proj-NETTYPE=0&proj-NETSPEED=0&proj-TIMEWAIT=0&proj-DSKCTRL=0&proj-DISK=0&proj-SCRIPTS=0&proj-WEBCACHE=0&proj-OS=0&proj-HWAVAIL=0&crit2-HWAVAIL=Jan&proj-OSAVAIL=0&crit2-OSAVAIL=Jan&proj-SWAVAIL=0&crit2-SWAVAIL=Jan&proj-LICENSE=0&proj-TESTER=0&proj-TESTDAT=0&crit2-TESTDAT=Jan&proj-PUBLISH=256&crit2-PUBLISH=Jan&proj-UPDATE=0&crit2-UPDATE=Jan&dups=0&duplist=COMPANY&duplist=SYSTEM&duplist=CORES&duplist=CHIPS&duplist=CORESCHP&duplist=CPU&duplist=CACHE1&duplist=CACHE2&duplist=CACHE3&duplist=NETTYPE&dupkey=PUBLISH&latest=Dec-9999&sort1=PEAK&sdir1=-1&sort2=SYSTEM&sdir2=1&sort3=CORESCHP&sdir3=-1&format=tab
Matt Bryant predicts, "SPARC... I predict Larry will cut them lose without a second thought."
So far, both Oracle & SUN indicate that SPARC is staying around... Larry seems to disagree with you on his first thought since merger announcement.
If RocK is significantly late (again) and T3 is late (CoolThreads have been on-time or early since 2005) - I agree this may be a reasonable prediction. Perhaps Larry's second thought will be align with your prediction under these conditions.
I guess, we will see!!!
@Mad Mike: Power 6+ & 7 will probably be screamers!
Mad Mike says, "Power 6+ is already out, at least according to IBM. Power 7 is on target and will be released on time. It's also not a 16 core design."
I have not been able to find any information, outside of speculation on Power6+ or Power7.
Register suggests the Power6+ is not out, as of "23rd April 2009".
http://www.theregister.co.uk/2009/04/23/ibm_power6_plus/
Wikipedia says there are 2 chips per module with 8 cores per chip - which looks like a total of 16 cores per socket, assuming 2 chips in a single multi-chip module per socket. Wikipedia is a little unclear.
http://en.wikipedia.org/wiki/POWER7
I look forward to the processor releases, I am sure they will all be screamers!!!
RE: Projected Intel, IBM, AMD, and Oracle/SUN/Fujitsu processors
One problem with old Novatose's marketeering - did he call Larry first to get his approval? Fact is, all the "coming" Sun chips mentioned are waiting on Axeman Larry to decide whether Oracle want them or not, and the evidence so far is not. First off, Oracle haven't committed to ANY current Sun hardware, let alone future products. Wonder what the two-week delay was in Sun announcing the Nehalem kit? It wasn't just Sun being slower to market than the other vendors (as they usually are), it was probably because Sun had to ask Oracle for permission to launch the kit. Until Larry stands up and says Rock and T3 are go they are just vapourware, and even less convincing vapourware than when the SPARC fanclub was running Sun.
And even if they do get released, Intel and AMD are hardly going to be quaking in their boots. T3 will still have the same scale problems of all the Niagara lines - Xeon and even the old Barcelona Opteron already out-scales and out-perfroms it. And the Nehalem and Barcelona cores are full-bodied cores with a real pipeline to keep those threads spinning, whereas the whole Sun design is a capitulation to the idea that you can't have all the threads going at once, because Sun couldn't design a bus or core powerful enough to. All T1 and T2 did was cannibalise the existing low-end SPARC base, and even then Sun and FSC finally had to produce a single-SPARC64 M-series server (the M3000) as their customers kept screaming at them that Niagara just didn't do what they wanted. And the final nail in T3's coffin - even if Oracle drop the stupid Sun insistance of pitching Slowaris as the OS of choice instead of Linux, it still can't run Windows, which means it can't compete with the cheaper, more flexible and faster x64 options which can Linux, Windows or even Slowaris x86. Niagara's little webserving niche is not going to be enough to keep T3 alive in Oracle.
Rock? Way too little and far too late, even if they can fix the bugs. Going on the info Sun have released, Rock will be out-performed by the Power5 and old Madison Itanium2 cores it was originally supposed to go up against. Power6 and the latest Montecito Itanium2 cores will comfortably out-perform it, and then there are Power7 and Tukwila Itaniums waiting in the wings. 32-thread Power7 pricing is anyone's guess but likely to be keenly driven as IBM and hp scrap over who gets to migrate all those Sun accounts McNeedy and Ponytail have left in the lurch. Tukzilla-based Integrity servers from hp will come with the big price advantage that Tukwila and Nehalem allow more sharing of components than Power7 and xSeries can, so hp will be able to leverage the economies of scale of that massive ProLiant bizz to keep Integrity costs down. Oracle doesn't have a massive x64 bizz (Galaxy can't be described as a tier1 x64 bizz), and the Galaxy servers don't share components with the Niagara or M-series servers to anything like the same extent. So, Rock and T3 will offer poorer performance and at uncompettive prices. Yeah, I can see Larry jumping with joy at that prospect - not!
IBM aren't keeping the Power6+ as a spoiler to the Rock/T3 announcements, they don't need to. They are keeping it ready for the arrival of Tukzilla this Summer. With hp's new servers due around this September, IBM will want to throw a spoiler in there to stop hp pointing out to all those Sun customers they can have a shiny new Tukzilla server in September or wait another six months plus for Power7. After all, if IBM play it the same way they did Power6, they could start with just one part of the range and that could mean customers actually waiting for Power7 in the servers they need as late as 2011.
Interesting though that Novatose is no longer singing the praises of the Fujitsu SPARC64 chips. Are the Sunshiners just sulking because Fujitsu didn't ride in and save their fantasy world like they insisted FSC would? Or is this a silent admission that even the next gen SPARC64 - if it ever arrives - is going to be just as uncompetitive as Rock?
Larry Ellison isn't blind, unlike McNeedy and Ponytail, and he will know better than to follow their blinkered and comic faith in SPARC. Larry may want the Galaxy servers for his new storage stack, but the rest are just a profits blackhole, and I predict Larry will cut them lose without a second thought.
/Novatose - the comedy gift that keeps on giving.
