Haswell at IDF 2012: 10W is the New 17W
by Anand Lal Shimpi on September 7, 2012 2:52 PM EST- Posted in
- CPUs
- Intel
- Haswell
- Trade Shows
- IDF 2012
Earlier this week Intel let a little bit of information leak about Haswell, which is expected to be one of the main focal points of next week's Intel Developer Forum. Haswell is a very important architecture for Intel, as it aims lower on the TDP spectrum in order to head off any potential threat from ARM moving up the chain. Haswell still remains very separate from the Atom line of processors (it should still be tangibly faster than IVB), but as ARM has aspirations of higher performance chips Intel needed to ensure that its position at lower power points wasn't being threatened.
The main piece of news Intel supplied was the TDP target for Haswell ULV (Ultra Low Voltage) parts is now 10W, down from 17W in Sandy and Ivy Bridge. The standard voltage Haswell parts will drop in TDP as well, but it's not clear by how much. Intel originally announced that Haswell would be focused on the 15 - 25W range, so it's entirely possible that standard voltage parts will fall in that range with desktop Haswell going much higher.
Intel also claims that future Haswell revisions may push even lower down the TDP chain. At or below 10W it should be possible to cram Haswell into something the thickness of a 3rd gen iPad. The move to 14nm in the following year will make that an even more desirable reality.
Although Haswell's design is compete and testing is ahead of schedule, I wouldn't expect to see parts in the market until Q2 2013.
Early next year we'll see limited availability of 10W Ivy Bridge ULV parts. These parts will be deployed in some very specific products, likely in the convertible Ultrabook space, and they won't be widely available. Any customer looking to get a jump start on Haswell might work with Intel to adopt one of these.
The limited availability of 10W ULV Ivy Bridge parts does highlight another major change with Haswell: Intel will be working much closer than it has in the past with OEMs to bring Haswell designs to market. Gone are the days when Intel could just release CPUs into the wild and expect its partners to do all of the heavy lifting. Similar to Intel's close collaboration with Apple on projects like the first MacBook Air, Intel will have to work very closely with its PC OEMs to bring the most exciting Haswell designs to market. It's necessary not just because of the design changes that Haswell brings, but also to ensure that these OEMs are as competitive as possible in markets that are heavily dominated by Apple (e.g. the tablet market).
Don't expect any earth shattering increases in CPU performance over Ivy Bridge, although I've heard that gains in the low double digits are possible. The big gains will come from the new GPU and on-package L4 cache. Broadwell (14nm, 2014) will bring another healthy set of GPU performance increases but we'll likely see more than we did from IVB with the transition to Haswell on the graphics side.
Configurable TDP and connected standby are both supported. We'll also see both single and dual-chip platforms (SoC with integrated IO hub or SoC with off-chip IO hub), which we've known for a while. We'll get more architectural details next week, as well as information about all of the new core and package power states. Stay tuned.
Facebook
Twitter
RSS
43 Comments
View All Comments
umbrel - Friday, September 7, 2012 - link
I don't think they are aiming at people with 1~2 years old computer. The target for this product should be 3 or more year old PCs.Compared to those the improvement in performance, TDP and iGPU are not so marginal.
blandge - Friday, September 7, 2012 - link
Power and graphics. Haswell is all about low power in the ultrabook and convertible segments, and a sizable incrase in GPU power.But to answer your question, there is not much reason to upgrade from an i7 2600k @ 1155 to and i7 4xxxK @ 1150.
Granted, Intel doesn't really care about that because ultrabooks are their main concern with Haswell.
jhoff80 - Friday, September 7, 2012 - link
... is there ANY possibility of Microsoft getting these in the Surface Pro? I assume it won't have Haswell or better until its successor, but damn I'd love if they ended up delaying Surface Pro just barely long enough to use one of these chips.amdwilliam1985 - Friday, September 7, 2012 - link
I like your idea. Delay Surface Pro to get these chips in.And remember to keep that price in balance too.
Anand Lal Shimpi - Friday, September 7, 2012 - link
Not in Surface Pro, but bet on it being in the second version :)dagamer34 - Monday, September 10, 2012 - link
Why shouldn't the Surface Pro get the special 10W Ivy Bridge parts. It makes delaying that product all the more believable if it isn't just another run of the mill 17W ULV Ivy Bridge CPU. And you know Microsoft has to get the battery life right to make that product really worth investing in.Shadowmaster625 - Monday, September 10, 2012 - link
There really is no difference between 17W IB and a hypothetical 10W IB. With configurable TDP and p-states and all that other jazz you should be able to just throttle down the 17W part so it fits in whatever envelope you need. A good windows tablet would have a 50Wh battery.It shouldnt be too tough to dissipate 13W without a fan. I consider 13W to be the sweet spot.Keep in mind that even the iPad 3 can easily burn up 10 watts while doing heavy gaming. (I'll bet I could push well past 10W if I ever got my hands on one.) So it shouldnt exactly be rocket science to simply dissipate 30% more heat than the ipad.
EnzoFX - Friday, September 7, 2012 - link
I long for smaller form factors. ITX being even more streamlined. Smaller everything, PSU's, even cables. Also the idea of cramming these babies into something super small, a nettop I could run a full OS on is always so attractive to me.With that excitement said, it seems that CPU power is taking a backseat, makes sense as most things aren't cpu bound whatsoever. Are we entering a period where software needs to catch up? Or are we waiting SoC's to catch up? In which case then we'll start seing more well rounded growth on both hardware and software?
dcollins - Friday, September 7, 2012 - link
As a software engineer, I'm incredibly excited about Haswell's support for Hardware Transactional Memory. Please get all of the details you can on that. For the layman, HTM has the potential to radically simplify and speed up mutli-threaded programs since Transactional Memory is currently one of the most promising parallelism techniques and software implementations have a 2-4x slowdown on a per thread basis.Loki726 - Friday, September 7, 2012 - link
It is interesting to see companies going in this direction, but don't expect miracles out of HTM. Research efforts have been looking at these issues for at least the last decade.The progression of these ideas is as follows:
OOO execution [Pentium Pro] -> aggressive OOO (multi-cluster, trace/loop-caches, prefetchers/run-ahead, etc) [Sandybridge] -> HTM [Haswell] -> HW-assisted compiler speculation [1] -> Multi-scalar [2] -> Speculative multi-threading [3]
[1] - A Real System Evaluation of Hardware Atomicity for Software Speculation, ASPLOS 2010
[2] - Multi-Scalar Processors, ISCA 1995
[3] - Bluegene Q
Each one of these comes with diminishing returns for programs with sequential dependences, and are not necessary for truly parallel programs.
For a software engineer, I would suggest exploring either better sequential algorithms for your problem (which are likely to yield much better performance gains than these HW techniques), or explore truly parallel algorithms (conflict/dependency free) that have the potential to scale on a lot of simple cores.