Interview with Intel’s Memory Guru, Christopher Cox

Being somewhat of a memory addict, I couldn’t help but be intrigued with OCZ’s announcement of Intel XMP modules early last month. It was the first I had heard of the technology, as Intel kept completely mum about it until recently. What we did know, though, was that it is essentially Intel’s version of NVIDIA’s EPP technology, which we explored in last year’s interview with Corsair.

As a quick primer before the interview, XMP is “Extreme Memory Profiles”, a technology that will become functional with Intel’s X38 chipset, due out soon. Similar to EPP, it will allow those desiring extra performance to use XMP profiles, instead of getting dirty with overclocking. The downside of EPP though, was that many true enthusiasts found the pre-overclocked settings to be laughable at best. XMP shows no sign of suffering the same fate of being labelled as “for newbies only”.

On the NVIDIA side, EPP-equipped modules are denoted with an “NVIDIA SLI-Ready” logo, which many memory manufacturers have used. Intel’s XMP however, has no special logo, although memory companies may choose to use the “Supports Intel Core 2 Extreme” logo, since all XMP modules are recommended for use with Extreme Edition processors, due to the fast frequency capabilities of both.

At last months Intel Developer Forum, I had the pleasure of meeting Christopher Cox, Sr. Staff Engineer for Intel’s Platform Memory Operations. After speaking with him for a few minutes, I quickly understood the passion that went into Intel’s memory technologies. He exhibited such great enthusiasm about what he was speaking about, that I knew an interview had to be in order. I’m pleased to say that we have one here today.

Techgage: Hi Chris. We appreciate your taking the time to answer some of our burning questions regarding Intel’s latest memory technologies. Can you first tell us a little bit about yourself and what your primary job responsibilities are?

Chris: A little about myself – First and foremost, I like to consider myself a gaming enthusiast, with deep interests in pushing my hardware as fast as I can (overclocking). I’ve been in the industry now for about 14 years, ~8 of which at Intel doing various things from EFI bios work to memory. Currently, I’m a DDRx Memory Architect and am responsible for writing the internal DRAM specs and owned the DDR3 specification at Intel.

Techgage: From what I recall, you are the lead designer behind XMP, so I’m sure you consider it your baby. When and why did Intel decide to work on such a technology, and was simply adopting NVIDIA’s EPP technology out of the question? Also, are there any direct advantages to XMP over EPP?

Chris: Yep, I guess you could say I’m the lead designer behind XMP. Fundamentally, the EPP and XMP concepts are very similar; think of XMP as an evolutionary change over EPP for DDR3. ; I’m just expanding on it for DDR3 and adding some new things that can’t really be seen by just the SPD portion of the spec (like new register definitions in our chipset or mobile centric power saving features).

Extreme Memory is really more than just the XMP spec; it’s reflective of Intel’s renewed focus on addressing the enthusiast segment. It’s also about adding new features to hardware, passing those registers to the BIOS teams and ultimately building a better environment for the end users.

We decided to focus on DDR3 because it allows us to take advantage of the better performance scalability of DDR3 technology and aligns with our DDR3 product schedules & plans. (This is also strategically aligned with our memory technology transition to DDR3).

Techgage: Since NVIDIA doesn’t yet support DDR3, we know that EPP will not be available there. However, will X38 ever allow XMP capabilities for DDR2, or is Intel going to focus solely on DDR3 and the future?

Chris: Intel has made a strategic decision to focus on a version for DDR3. The design of the X38 and the underlying architecture that drives XMP is compatible with most EPP parameters for DDR2, it’s just a matter of a ODM deciding to support it.

Techgage: At this point in time, Intel has laid out DDR3-1600 speeds as being the highest official available with XMP (due to 1600MHz FSB soon to be supported). Are memory manufacturers (OCZ, Corsair, etc) able to include a profile that goes beyond that frequency boundary?

Chris: Actually, that’s not really correct. XMP is completely agnostic to frequencies, latencies, voltages and such. It’s true we do have an upcoming platform that would utilize XMP at 1600 as a POR configuration to match the FSB but by no means is XMP limited to DDR3-1600. In fact, with some limited internal parts, we have gotten the FSB /MEM to 2380Mhz_ of course, that is definitely overclocked!

Techgage: Continuing in that vein, what I gather from the first XMP-capable modules on the market is that the overclocked settings are far from being considered weak. Some manufacturers are really trying to push the tight-timing/high-frequency limit. With that taken into consideration, are XMP profiles guaranteed to work on all X38-based motherboards, or will there be some instances where the motherboard will not be able to handle it?

Chris: This is a good question and one that really needs to be communicated right. XMP supports two profiles, the first is the certified enthusiast profile (great for those that want to flip a switch and have all the settings come up and give you some basic performance improvement). Intel has provided the memory suppliers with a set of test criteria that defines a certain amount of guardband for DIMM stability at XMP certified rating. This test condition also specifically identifies which platform it was tested on – so when an end user goes to the Intel XMP site, they will see a listing of which XMP modules were certified on which platforms (i.e. DIMM vendor 1 certified on the Asus P5K3).

The 2nd profile is what I call the Extremist profile, it typically will contain the absolute best settings that the DIMM mfg has seen run and will most likely contain settings that may not work on every platform (like command rate changes or command turn around optimizations). This is the bleeding edge profile and your mileage may vary.

Another key thing to keep in mind about XMP is that it doesn’t always mean higher frequencies or higher voltages. Sometimes, say like on a mobile platform, performance is sometimes rated in BW per Watt and therefore the XMP settings could be optimized to enable Extreme battery performance or other mobile-centric features.

Techgage: Does XMP affect the CPU frequency similar to EPP, and does it have the capability to adjust CPU/Memory voltages, or just the memory?

Chris: Short answer, yes it ‘can’. For discrete memory controllers, when the memory freq is set higher than the FSB, the FSB would also have to be set at that same level. Of course, this also then requires you to have a CPU that can handle this higher frequency (and again, anything above the spec FSB, your mileage may vary).

For example, buying XMP-1800 DIMM’s doesn’t guarantee you that your FSB will be able to handle running at 450Mhz (450×4=1800), it’s just an indicator that if you have one of these CPU’s that overclocks well, the memory will be able to keep pace. Some Board Mfgs may choose not to have the XMP profile actively change the FSB speed automatically but just inform the user they would need to do it… it all depends on how each board mfg implements XMP into the BIOS.

Techgage: On one of the slides for X38, it’s mentioned that an “Extreme Tuning Utility” is included. Is this referring to XMP options in the BIOS, or will users be able to view the various profiles (or tweak settings further) with a software-based utility, a la nTune?

Chris: Extreme Tuning Utility is another ingredient to our umbrella type Extreme program that is like nTune/uGuru but has native support for XMP hooks as wells as many others. Extreme Memory Utility is more of a framework product that allows the ODM’s/OEM’s to build their own tools on top of. Because it’s the foundation framework, you may never really see it as it was shown.

One feature that I really like is being able to build, save and even export profiles. So you can build a config that is setup for stability, one for power saving settings, and one for extreme (gotta get every ounce of perf out of this machine I can even if it crashes every hour or so!)… or as may profiles as you wish. I could envision people sharing profiles for certain cpu/board/gfx/memory combos to kind of let others share in knowledge of how to tune the platform – or at least, what they found was a good config.

Techgage: From what I understand, XMP allows up to three profiles. The first will be a “stock” setting, presumably one set by JEDEC, while the second is Intel’s profile and the last being the memory manufacturers. In the future, will be there expansions to this to say, include additional profiles for even finer setting tweaks? Would a larger SPD even be possible, or would that lead to increased latencies?

Chris: Pretty close. For backwards compatibility, we left the normal JEDEC settings there (which I guess you could call the first stock profile). The other profiles have been described, described above, so I’ll just point you back there. As for the possibility to include additional profiles, Yes and No. The design of XMP has a byte that is setup for allowing different profile configurations, as well as the actual ID byte that could be used by another company to define a complexly different set of profiles. It is also possible that if the eeprom size increases, we could expand the number of profiles or such but it may not be necessary, the utilities from all the board houses based on the Extreme Memory Utility framework could make it even easier. The SPD could be used as a good baseline for that utility.

Techgage: In closing, are there any new Intel memory technologies on the horizon that we should know about, and what particular Intel technology soon to be released has you most excited?

Chris: We are working on adding a few more speed bins to JEDEC’s standard DDR3’s definition (Specifically like 1866 & 2133) – even though these speed bins will first be covered by the XMP spec. Having a base product that natively supports 2133 at 1.5V means that going to that next step in performance is going to be that much easier – not to mention that the 1866 / 2133 JEDEC speed bins won’t be overclocked, so they will be about as stable as memory gets – as you probably know, overclocking (ala XMP) is more of an art form than a science, so while some XMP modules may get to 2000Mhz with CL7 soon, not every board, every CPU and every module works the same, and your mileage may vary. It’s like driving a Lambo at top speed… some drivers get 212Mph, others only get 199Mph.

As for other technologies, big Intel news is the integrated CPU, code named Nehalem. As you have probably already heard, this part will integrate the Mem Controller into the CPU. With our current desktop platforms, generally memory freq wins out over latency (this is primarily due to the FSB design-scaling FSB has a bigger impact); however, with Nehalem’s point to point bus, you’ll see even more benefit from faster memory latencies and the highest frequency / highest voltage combination won’t necessarily be the best performance or only stop in town. This concept is not really new to Intel, as we had at one time planned on a similar architecture (which was in fact my first project at Intel) but the economies of scale did not support bringing this to market at that time.

Of course, we are also actively working with JEDEC on the definition of DDR4, which is more of an evolutionary change over DDR3 –lower voltage 1.2V, etc. – and is being designed to bring even more performance to our future platforms.

Techgage: Thanks, Chris, for taking the time to speak with us and also for all the great answers!

It’s safe to say that XMP is more than a simple EPP competitor, which is a huge relief for enthusiasts and casual overclockers alike. Though I consider myself to be one of those enthusiast overclockers, I admit that if I were building a new computer, I’d likely take the lazy-mans way out and choose the best XMP profile on a given memory kit. Unlike EPP, the preset overclocks are very -good-, especially with DDR3-1800+ frequencies. It would be hard to select such a profile and still feel ripped-off.

Of course, the only real downside to DDR3 and XMP-equipped modules at this point in time is the price. DDR3 first launched this past May, and since then, kits have not dropped much in price, with mid-range kits still hovering around the ~$500 mark… and the highest-end kits nearly topping the $900 mark. Given that it’s still new technology, slow price decreases was to be expected. We hope to see the prices much more reasonable beginning in the new year, so that even more consumers can jump on the DDR3 bandwagon.

Thanks again to Chris Cox and Intel for interview!

If you have a comment you wish to make on this interview, feel free to head on into our forums! There is no need to register in order to reply to such threads.