TWO MONTHS AGO, certain websites in the Far East showed early sample runs of Gulftown, the upcoming six-core 32nm processor that Intel will release in 2010. Originally expected to appear in the second quarter of next year, this flagship of the Westmere generation should ensure that the LGA1366 socket parts keep the substantial performance and price advantages over their more mainstream LGA1156 socket brethren.
On top of that, it would still be pin and - BIOS allowing - plug compatible with most current LGA1366 mainboards, but allow quite a bit of extra system level performance if used with newly tuned mainboard revisions and updated Tylersburg chipset steppings. In fact, you might not even need new mainboards in many cases, as the memory and QPI speed improvements are usually mostly related to the CPU uncore parts anyway.
Earlier site previews showed this new CPU running at only 2.4GHz with DDR3-1066 memory, surprisingly slow knowing it's a new process and basically an improvement on very well known and tested core and uncore designs by now. To me, that didn't sound exactly right since the combination of a newer manufacturing process and more time the designers had to further tune up both the core and uncore portions of the CPU should have resulted in more performance with more cores and more clock speed too.
So, when visiting Taipei a week ago, and speaking with two large vendors involved, I was happy to have them share their first experiences with the recent A-0 stepping deliveries of the Gulftown CPUs in both UP desktop and DP workstation slash server versions. Each of the two vendors has received around two dozen samples, and it was interesting to compare the frequency headroom they saw as well as stability and performance spread across the samples they had on a variety of mainboards.
Good news, most of the samples hit close to 3.5GHz with regular voltages, and around 4GHz with basic "push up" measures (we'll avoid saying "overclocking" among the vendors). And this is with six cores and 12MB cache per chip, a full 50 per cent increase in both categories from the current LGA1366 45nm processors, with roughly the same thermal design power (TDP) rating as well. Oh yes, also don't forget the encryption acceleration as well as little reductions to latency penalties all over the system.
The memory controller actually does perform better, too. Expect fairly easy triple channel DDR3-2133 operation on the desktop - and some more if you're a hard core overclocker - as well as standard DDR3-1600 ECC server memory at three channels per CPU socket. In fact, most of the new server and workstation mainboard revisions, according to our designer friends, will support a working "Force DDR3 1600 speed" option across all 12 DIMM sockets in a dual CPU system. If the memory is good enough, of course.
For the expected one or two dual-CPU overclockable board entries to take the domain previously covered by the Intel Skulltrail, the potential, with desktop high-end memory, is to have six channels of DDR3-2133 memory in parallel, on a 4.27GHz capable pair of Gulftowns with the right cooling. Our favourite - and hated by some commenters - Sandra memory bandwidth benchmark would surely yield some fabulous results then, north of 60GBps measured bandwidth.
Other good news, for Intel and the OEMs at least, was that the Gulftown release might be pushed slightly forward, to March of next year, just in time for the othewise cold and increasingly boring CeBIT show. This will be just after the desktop and mobile Clarkdale and Arrandale, the initial 32nm Westmere generation dual core parts, will be arriving.
The reasons? The good yield and clock speeds are obvious, and so is the benefit of selling high-end, high-margin parts earlier. After all, the initial Core i9 XE series desktop Gulftowns and "Westmere-EP" dual processor workstation and server Gulftowns will most likely sell in the $600 to $1,600 price range, if they follow the present LGA1366 CPUs. There will be no equivalent of the cheap Core i7 920 here this time, as the simpler LGA1156 CPUs have taken the mainstream role, even in Xeon UP servers.
So, it will benefit Intel's coffers, as well as the average selling prices of the matching high-end mainboards from vendors such as Asus, Gigabyte, DFI, Tyan and others. After all, if you're spending nearly two thousand quid on a pair of Gulftowns, you won't even blink an eye at spending another half thousand on a matching mainboard. And, the Westmere Gulftown refresh with all the CPU and memory timing tune-ups will come just in time to implement a more finalised USB3 and SATA6 Gbps controller solution on the spare PCIe V2 lanes of the Tylersburg IOH North Bridge, those which the P55 chipset doesn't have.
These mainboards will also be among the first to, finally, most likely have DSP-based sound, either the expected Realtek DSP or a high-end part like the Creative X-Fi. What remains is to use higher-end Ethernet controllers with better TCP/IP processing offload like the Intel 82576, as well as hardware-accelerated RAID rather than the CPU processing XOR logical ops for RAID5 or even the mirror checks. Intelligent I/O is coming to the desktop, back after 20 years since being last seen on the Commodore Amiga.
In summary, the Taiwan vendors and global users alike will be happy - a good speedup matched with useful tuneups at the system level, and unmatched power for the desktop, workstation or server. A double-precision 160 GFLOPS dual CPU deskside box will be a reality now, before we count anything from the GPUs.
AMD, however, will have pressure on its high end increased further with these chips, both on the desktop and on the workstation and server. Even if we get to tweak the Istanbul to somehow reach 3GHz on six cores, it'll have a hard time competing against a say 3.33GHz Gulftown, and that's before we even turn on any Turbo or multithreading gadgetry. The Gulftown can support 50 per cent more memory capacity and nearly twice the bandwidth due to faster DDR3 grades supported, and AMD's dual die Magny Cours can't go into the same price bracket as simpler, single-die Intel offerings. So, it remains to hope that the next generation Bulldozer-based high-end CPUs will see the light of the day - and press cameras - in 2010 rather than later.
As for Intel, the issue remains that, with the next generation core out of the Haifa labs, the Sandy Bridge with its doubled floating-point throughput, ring bus architecture and substantially reduced uncore latencies, the attention will be on it soon after these Westmere parts are out. Sandy Bridge was already trumpeted during the past Intel IDF, and Intel will have a task on its hands to manage everyone's expectations and avoid an unlikely, but still possible, mini-Osbourne debacle where people hold off for the next CPU coming too soon after the just announced one. In this respect, speeding up the Gulftown rollout makes perfect sense on both the high-end desktop and the workstation and server platforms.
Plus Extra performance across the board, plug compatible.
Minus AMD falls even farther behind, Sandy Bridge may be closing in. µ