Additional Details on Sandy Bridge-E Processors, X79, and LGA2011by Kristian Vättö on April 26, 2011 12:46 PM EST
- Posted in
- Sandy Bridge
The latest Intel roadmap has come out, and it's already being discussed elsewhere, so we're going to weigh in with our own analysis of the content as there's plenty of interesting bits of information to sift through. We’ll be looking at other areas over the coming days, but today we’re going to start with the Sandy Bridge-E (SNB-E) processors. Sporting a new socket and chipset, the SNB-E CPUs will start showing up in Q4 this year. None of this is new, as we’ve known the general timeframe for the launch since our Sandy Bridge review, but we can now add some concrete specs. According to the roadmap, the initial SNB-E lineup will consist of three CPUs: two hex-core processors and one quad-core. We don’t have model numbers yet, but we do have most of the other pieces of information.
|The Sandy Bridge-E Lineup|
|Family||Core i7 Extreme||Core i7||Core i7|
|Max SC Turbo||3.9GHz||3.8GHz||3.9GHz|
|Overclocking||Fully unlocked||Fully unlocked||Limited unlock|
The new chips will all use the LGA2011 socket with Intel’s X79 chipset, scheduled for simultaneous release with the CPUs. The platform replaces the current LGA1366 with X58 chipset, providing an upgrade path for high-end enthusiasts and workstation users. Memory support will move up to quad-channel DDR3-1333, so where the current Bloomfield can provide up to 25.6GB/s of bandwidth at the specified tri-channel DDR3-1066, LGA2011 kicks that figure up to 42.7GB/s—a 66% increase. The additional memory bandwidth should be particularly useful with certain workloads on the hex-core chips.
One interesting piece of information is that the roadmaps make no mention of integrated graphics or Quick Sync, suggesting the platform will be for discrete graphics only. That makes perfect sense on one level, as users likely to upgrade to such high-end systems are almost sure to have discrete GPUs. On the other hand, Quick Sync has proven very effective for video transcoding, providing up to a four-fold increase over CPU-based encoding, so the loss of the feature is unfortunate.
Intel hasn’t disclosed all of the various Turbo modes yet, but they have listed the maximum single-core Turbo speeds. Both the hex-core 3.3GHz and quad-core 3.6GHz top out at a maximum speed of 3.9GHz, and likely the hex-core chip can do 3.6GHz on QC workloads making it equal to or better than the QC chip on every potential workload. The 3.2GHz hex-core steps the maximum clocks speeds down 100MHz, along with cutting the L3 cache size. As with other i7 processors, all the new chips support Hyper-Threading, and while the hex-core chips will be fully multiplier unlocked the quad-core offering will be a “limited unlock”. The roadmap states that the limited unlock will allow up to six bins of overclocking above the maximum Turbo frequencies, which means that even that chip should be able to hit up to 4.5GHz (with appropriate cooling, motherboard, etc.)
Intel makes no mention of pricing at this time, but the new chips should follow familiar patterns. The i7 Extreme will replace the current i7-990X and target the familiar $1000 price point. Moving down, the 3.2GHz hex-core replaces the current i7-980 (which is set to replace the i7-970 in the near future), taking over the $550~$600 range. At the bottom of the SNB-E lineup is the quad-core 3.6GHz chip, which will take over from the i7-960 as well as providing a competitor to the i7-2600 in the sub-$300 market.
|PCIe Graphics||2x16 or 4x8 (chipset)||2x16 or 4x8 (CPU)|
|PCIe Based Uplink to CPU for Storage||No||Yes (x4)|
|USB 2.0 Ports||12||14|
|SATA Total (6Gbps)||6 (0)||14 (10)|
One final area to discuss is the chipset. We’ve included X58 in the above table as a reference point, and we can see that X79 improves a few areas but still fails to support a few newer technologies. While the X79 chipset will include native support for SATA 6Gbps (up to 10 ports, with four additional SATA 3Gbps ports), USB 3.0 support is still missing, similar to the current 5- and 6-series chipsets. X79 natively supports dual x16 PCIe graphics, or quad x8 graphics, but this time the PCIe lanes come directly from the CPU instead of from the chipset, providing lower latency GPU access. There’s another extra, as the CPU (chipset) has the option to use four additional PCIe lanes from the PCH dedicated to storage bandwidth, presumably to help with performance on fast SATA 6Gbps devices (e.g. SSDs).
Given the 2x16 PCIe lanes for graphics and quad-channel memory, we can account for most of the pinout increase relative to LGA1366 and LGA1155, and adding in these remaining storage PCIe lanes with a DMI link to the chipset should take care of the rest. Intel doesn't state whether they're using DMI or QPI, but DMI 2.0 only provides up to 20Gbps between the CPU and chipset, so supporting 10 SATA 6Gbps ports with fast SSDs would certainly saturate that.
That wraps up the consumer side of the SNB-E platform. Note that Intel will also have SNB-E Xeons launching in a similar timeframe. The bigger concern for us is that SNB-E continues the strengths of the Bloomfield/Gulftown processors but doesn’t address some of the weaknesses (i.e. lack of Quick Sync). SNB-E looks like a very capable processor, but if you’re willing to forego the current SNB lineup and wait for SNB-E, you’ll then have to contend with Ivy Bridge. That will be Intel’s first 22nm CPU and it’s scheduled for release in the first half of 2012, but that’s a story for a separate article. We’ll also have additional information on Atom CPUs and Intel SSDs in the near future.
Post Your CommentPlease log in or sign up to comment.
View All Comments
DanNeely - Tuesday, April 26, 2011 - linkDo you have any confirmation from Intel about x79 using a QPI link instead of DMI? The older rumor/leaks I've seen on LGA2011 all listed it as using the latter to connect to the southbridge.
JarredWalton - Tuesday, April 26, 2011 - linkThe information we have doesn't specify the link type/speed, which is why we say "assuming it uses a similar QPI link..." If it's DMI, then the additional storage-dedicated lanes make a lot more sense, but then the PCIe links would have to come straight from the CPU or they'd saturate the DMI link, right?
The Sandy Bridge and Ivy Bridge platforms both use DMI, because the CPU has the PCIe graphics link, but I believe that's the whole point of LGA2011. Also, as noted above there's a huge increase in the pinout of LGA2011 relative to LGA1366, and I can't think of where they would use all those extra pins if they're not doing a full QPI link.
JarredWalton - Tuesday, April 26, 2011 - linkOkay, discussed with Anand and updated. It's no *entirely* clear whether SNB-E will use QPI or DMI. Given the extra four lanes for storage, DMI would be very likely. What does this mean for Xeon, though? Maybe the chips have the option to do QPI or DMI, but they're using DMI on single socket configurations and consumer models, and QPI will only be on multi-socket. Or maybe they'll use a different die for Xeon. Lots of possibilities, but the text now reflects the fact that some of this isn't entirely clear, and we make some educated guesses as to what's going on.
DanNeely - Tuesday, April 26, 2011 - linkI've been scratching my head about the pincount as well. LGA 1567 is quad channel and 4x QPI, so the only new on die features adding pin count would be the PCIe lanes, and DMI link (4x PCIe). Each lane is 4 data pins, so 40 pcie lanes (32 +4 sata + 4 dmi) is 160 pins for data, and a few more for control/clock/etc (Wikipedia isn't clear which control signals are per slot which are shared), for maybe 200ish total. I'm going to assume that intel will also be using this socket for their forthcoming 4/8way boards and will keep the four QPI links. If they drop back to only two there're another 164 pins unaccounted for.
Maybe the remaining 250ish are just extra power/ground connections in order to support bigger/faster chips?
Kristian Vättö - Tuesday, April 26, 2011 - linkhttp://vr-zone.com/articles/intel-x79-roadmap-leak...
That says that the SATA ports will have total bandwidth of 8GB/s (64Gb/s) due to the four extra PCIe lines. That would be enough even for 10 SATA 6Gb/s SSDs. A big upgrade from X58 though, since it tops out at around 660MB/s from what I have heard.
This is starting to be very complex lineup for Intel if some parts use QPI and others DMI, let alone what LGA 1356 will add to the mixture :D
DanNeely - Tuesday, April 26, 2011 - linksome parts using QPI and others DMI is the same situation as we had with LGA1156/1366, so I don't see anything changing.
LGA 1356 rumors are operating at a much lower level than LGA1155/2011 rumors have. It's an open question if it's a real product or not. My guess is that it either always was vaporware, was a real part at one point but completely discontinued to only have 2 sockets vs 3 for the last architecture, or is a plan B part if bulldozer turns out fast enough that quad core/dual channel Ivy Bridge CPUs aren't fast enough to compete in the $200-300 range, and hexcore dual channel DDR3 IB chips would be badly memory bottle necked.
Kristian Vättö - Tuesday, April 26, 2011 - linkThe LGA 1356 is pretty much confirmed from what I have read and heard from my sources. Currently it looks like it will be meant for low-end DP systems so it is unlikely that we will see any consumer parts for that socket (i.e. only Xeons).
That would mean that there will be three sockets of SB Xeons (don't forget E3 lineup for LGA 1155 socket).
DanNeely - Tuesday, April 26, 2011 - linkAnything publicly available on highly trustworthy sites? I don't think I've seen anything on sites I consider A list for reliability. This isn't to say I think all the sites reporting the rumors are questionable. some might quality, but there're far more sites reporting rumors than I normally follow and can evaluate.
Casper42 - Tuesday, April 26, 2011 - linkI was in a briefing by Intel in Feb and one of the things I remember complaining about was the complexity of the SB lineup in servers.
I remember seeing the following:
1 way with Dual Channel
1/2 way with Triple Channel
2/4 way with Quad Channel (4 way design here would need 2 hops for some chip-to-chip communication due to limited QPI links so it was like the poor man's 4 way box)
4/8 way with Quad Channel (this had more QPI links thus making it a proper 4 way box)
I wanna say they were socket H, B2, R and lastly EX-B for the big boys. (same order as specs above)
I straight out called them out on it as AMD has boiled their lineup to a simple 1/2 Dual Chan socket and a 2/4 Quad Chan socket. Much easier to build servers when the chips are clearly delineated.
So assuming 2011 is R and I know 1156 is H, that certainly leaves room for 1356 to come in as B2. EX-B will likely be the last addition as it will supplant the soon to be released Westmere-EX chips. The monster 10 core 30MB cache ones.
Kristian Vättö - Wednesday, April 27, 2011 - linkhttp://www.cpu-world.com/news_2011/2011020803_Deta...
Socket B2 is same as LGA 1356. Obviously, I cannot confirm that this is 100% accurate. It does make sense though. Currently Intel doesn't offer much in the low-end DP market due to their high prices. Maybe LGA 2011 parts are simply too expensive for Intel to manufacture, so they couldn't replace the E5502 for example.
LGA 1356 with crippled features would let Intel stay competitive, even in lower-end markets. Cut down the PCIe lanes, stick with triple-channel, possibly no ECC and stuff like that.
Server grade stuff is not usually reported by these regular news sites so getting information that is available for everyone is harder. Subscription is usually required to get access to information like this. I don't have access like that so this is just what I heard from a friend of mine who has access to this kind of information.