AMD and Intel's Dual Core Demonstrations; The Race to Dual Core

The first time I, personally, had ever heard anything about "dual core" processors was several years ago at a regional conference when none other than Anand Shimpi gave a keynote about the advantages of parallelism, particularly with multiple core CPU applications.

Needless to say Intel and AMD must have come to the same conclusions earlier this decade since today we have seen server demonstrations with 90nm server chips from both companies. As the timeless adage states, "Many hands make light work."

We first got a hint that things would go toward dual core when we heard of a technology from Intel called "Hyperthreading." This precursor to truly multi-core processors emulated two processors on the same CPU. Many industry insiders almost took this as a warning for programmers to start optimizing and designing code for multithread operations. As quoted by Paul Otellini during Keynote Q&A of the recent Fall 2004 IDF, "over 90% of HT capable Xeons run with HyperThreading enabled."

With the dual core announcement, forum members (even in the AMD suite) were constantly correcting each other as to new nomenclature; no longer do we use "2-way, DP or 2-processor" to describe a server. Instead the correct terminology becomes a "2-Socket configuration." A dual socket, dual core Intel configuration actually refers to eight logical processors! Four physical processors exist on two sockets/dies with HyperThreading enabled on all four. Extremely threaded applications - such as web and database servers - are specifically tailored to take advantage of the additional threads, and things only get better from there.

During the Day 1 and Day 2 Keynote speeches, a technology that grabbed our attention was something called Vanderpool Technology or VT. VT is an upcoming feature on several processor families (including the dual core families) that allow processor emulation and/or virtualization. As explained by Intel, with Vanderpool Technology, we can actually run two operating systems simultaneously on the same machine, and even dedicate one processor to a given operating system, and another processor to a different operating system.

Of course, no Intel Developer Forum is ever complete without a stop by the AMD hospitality suite usually within a block or two of the convention center. In fact, AMD has an extremely similar technology to Vanderpool dubbed "Pacifica." According to our AMD correspondents, Pacifica technology will allow us to run multiple operating systems and dedicate specific CPUs to those operating systems. AMD also unveiled to us a technology called "Presido", which looks identical to Intel's Lagrange Technology (LT).

Dual Core Intel Solutions

Intel's dual core demonstrations came in two forms; the first being a demonstration of Itanium 2 during the Day 1 Keynote. Day 2's Keynote demonstrated a first look at a 65nm mobile processor, dubbed Yonah.

We got an extremely detailed analysis of the Montecito core from the introduction keynote. Below you can see the general layout of the core.

Things we do know about Montecito are listed below:

• Codename: Montecito
• Process: 90
• Operating Wattage: 100W
• Clock Speed: 1.7GHz
• Two cores, Four threads - First Itanium processor with HyperThreading

Although Montecito had quite a bit of promise around it, it only exists in A-0 (Alpha) silicon. Intel was extremely pleased to state that Montecito boasted 1.5 to 2 increases in performance on first silicon. Obviously, average performance improvement should increase even further as programs become more threaded and as more efficient arbiter algorithms are produced. It is important to remember that performance from a dual core processor can never exceed 2x (and even getting a specially optimized benchmark to see full 2x improvement is very impressive), so it looks like Intel is on a very solid track.

The Day 2 Keynote, which revealed the "Yonah", was Intel's introduction to "Personal Parrallel Computing" - a term introduced during the Day 1 Keynote. Applications like realtime facial recognition require on average 50 gigaflops with today's software. These enormous requirements are not out of reach of near-future processors, claims Intel. We were then introduced to Yonah, a 65nm dual core successor to Dothan.

We outlined both cores in red, but many of the elements found on Montecito are clearly visible in Yonah as well. The topmost portion of the the image is the L2 cache and the arbiter unit is placed in the center. The Yonah cache (green) and Dothan cache (black) do not look proportionate. It may be the case that Yonah processor and the Dothan core utilize the same amount of cache.

Although difficult to tell from the quality images we have of Dothan's core, Yonah and Dothan do not appear to share many qualities with each other. This may be due to the fact that Intel uses an automated layout process that is trained to minimize hotspots and efficiently route signals. This may have just been the result of their optimizations, but it may also reflect alterations or enhancements Intel has made to the Dothan core. Things we do know do know about Yonah are listed below:

• Codename: Yonah
• Process: 65nm
• Operating Wattage: 50W
• Clock Speed: Unknown
• Launch Q3'05

Don't forget, a few weeks ago we touched upon Smithfield; Intel's dual core desktop processor. As we mentioned in our Day 1 coverage, expect to see shipments of Smithfield, Yonah and other dual core processors pan out in this fashion:

Multi-Core Processor Shipping Expectations

And just how is Intel producing chips with higher clocks and lower wattages, and then sticking additional cores on them? Many of the answers lie in our earlier IDF 65nm coverage. For the most part, moving from the 130nm to 90nm process seems to result in the largest changes. Variants of Enhanced Speed Step are also showing up on all processors in the Intel arena as well.

Dual Core AMD Solutions

AMD's dual core announcements were actually made slightly before Intel's and in more public fashion.

The hinge of AMD's dual core strategy revolves around the "System Request Interface", or SRI. The SRI simply acts as an arbiter between the core and the crossbar switch - which in turn controls the memory controller and HyperTransport resources. Like Intel's solution, cache (L1, L2 and L3) are not shared between cores. New AMD servers will have dedicated memory banks per socket, although each memory bank can be accessed globally by any of the other CPUs.

AMD's dual core implementation required no hardware upgrades other than a BIOS flash on the servers running dual core processors. One four socket HP Proliant DL585 server ran Windows 2003 (32-bit) while the other four-socket machine ran a RedHat Linux variant on the 2.6 kernel. However, details of the demonstration - particularly clock speed - were not clear. The important thing to remember for both the AMD and Intel demonstrations is that upgrading to dual core processors requires only a BIOS update from hardware available today.

Things we do know do know about Eygpt/Italy/Denmark are listed below:

• Codename: Egypt (8 socket), Italy (2 socket), Denmark (1 socket), Toledo (Desktop)
• Process: 90nm SOI
• Strained Silicon
• Operating Wattage: Unknown
• Clock Speed: Unknown
• Launch H2'05

Unfortunately, AMD would not quote on exact thermal or clock speeds for either machine. We were generally under the impression they were running at lower clock speeds, however Pat Patla, AMD's Director of Server Marketing, was quick to point out that both systems were running in standard hardware configurations without additional cooling and shipping dual core product should be rated at the same wattage as single core solutions. Again, the large message behind AMD's dual core push was not thermals or clocks, but that the technology exists today and that deployment will not be hindered by core logic or motherboard changes.

The AMD roadmap has also been slightly updated to reveal the presence of the dual core chips form AMD.

AMD or Intel, the better solution?

The largest question on everyone's mind was Intel's future plans for desktop and mobility dual core processors. A large amount of evidence suggests each processor family will continue its development cycle on existing architecture - i.e. Smithfield is a third generation 90nm Pentium 4 architecture (NetBurst), Yonah is a third generation 90nm Pentium M architecture (P6). Of course, the dual core Montecito processor demonstrated on Day 1 of the show coverage is a 90nm IA-64 processor. Smithfield, the desktop dual core Pentium 4, will remain electric and pin compatible with Socket 775.

Even with similarities cropping up between the two CPU manufacturers, there are some architectural differences that will force the entire industry to reconsider how performance should be handled. Obviously, AMD is advantaged in a dual-socket, dual-core scenario when compared to an Intel dual-socket, dual core configuration. Whereas 8 logical CPUs need to share one memory controller on an Intel server, 4 physical CPUs share two memory controllers. Intel may be countering this issue with independent memory controllers on future chipsets such as Twin Castle. Then again, on the other hand Intel is capable of producing processors that can emulate twice as many threads.

Today things between AMD and Intel look extremely competitive. They both have first generation strained silicon 90nm dual core processors demonstrations, both utilize a "security" module and a "virtualization" module and both are capable of 64-bit addressing. To sum things up; both AMD and Intel expect to ship dual core Server, Desktop and Mobility processors by 2006. AMD's release schedule anticipates a release of dual core server processors, followed by desktop processors and finally DTR and mobility CPUs - with each family of processors releasing a few months apart from each other. Intel's strategy also includes server processors as the most prevalent CPU product SKU, but actually weights higher priority on mobility P6 (Pentium M) Yonah processors.

As if the last two years have not been exciting enough in the AMD versus Intel arena, do not be surprised if the next several months show similarities between the "Race to 1GHz" competition several years ago. And of course, as the "Race to Dual Core" picks up, AnandTech will be there with all the details.