We hadn't planned to use the old test method anymore, but reality made us reconsider our decision. We got ourselves an engineering sample of a new Core i7 processor, and we decided to test it with our time-proven method to be able to compare it with other CPUs. And as soon as we adopt the new test procedure, our first step will be to retest this processor and other models from this and other series.

Now what's so interesting about the Core i7-950? As we have brought it up to the title of the article, it's an original processor, but it's not surprising. Why original? The Core i7 family, just like Xeon 3500, initially consisted of only three models at 2.66GHz, 2.93GHz and 3.2GHz, that is going up by 266MHz steps (or the doubled reference frequency). The Xeon 5500 series is tighter -- it starts at lower frequencies and goes to 2.93GHz using all integer multipliers. What does it imply? It would have been logical to expect Core i7 930 operating at 2.8GHz as the reviewed Xeon X5560. But the company has rolled out Core i7-950 instead, operating at 3.06GHz. That's original, there is no such thing in the other families.

But it's not surprising. TDP restrictions may explain why Xeon 5500 with this clock rate does not exist. To all appearances, the company hasn't managed to 'squeeze' processors above 3GHz into the limits of 95W TDP. So the flagship of Series 5500 (Xeon W5580) has the 130W TDP. It's a flagship, after all. But two tigers cannot live on the same mountain. The X5570 has a recommended price of $1386, while W5580 has the MSRP of $1600. So what should the price of the hypothetic W5575 be, considering that it consumes as much power as a top processor and its performance is somewhere in between these two models? So, there is no point in this processor, it will bring only confusion (Intel hasn't even reserved a model number for it).

Core i7 is another pair of shoes. All processors from this family (even the 920) have the TDP of 130W, so heat release is not a problem. The 930 would have messed everything up -- this inexpensive model would have had to stand between the 920 and 940. An offcast. But the 950 is welcome here. The fact is, the delayed Core i5/i3 and 'squeezed dry' Core 2 force the company to expand Core i7, even though it wasn't planned. In particular, Core i7 Extreme Edition 975. It differs from the previous top processor (EE 965) by just 133MHz, and by 400MHz from the 940 model. In these conditions it's apparently dangerous to put the EE 965 down to the 960 by locking core/memory multipliers, cutting down the price, but preserving the same official frequency -- its performance level will be too close to the top product. And the EE 965 is not interesting in its current implementation either -- the EE 975 traditionally has a disbursing price of $999, that is identical to the old price for the EE 965. But the gap between the 940 and the EE 975 is too big. There is an elementary solution to this problem -- if the clock rate of the extreme model is increased by 133MHz, the same must be done to the 'regular' top processor. Prices should be preserved, that is the 950 will cost just like the old 940.

Meet Core i7-950. Does it differ from the 920 and 940, except for the frequencies? No, it does not. QPI bandwidth is still 4.8GT/s (it's higher in top Xeons and Extreme modifications -- 6.4GT/s), the multiplier is locked to 23, there are only two memory multipliers -- 6 and 8, which provide support for DDR3-1066 maximum, not DDR3-1333 as in the X55x0. They could have done better here, as manufactures of three-channel memory kits usually offer DDR3-1333 or higher. Alas, it's not destined to happen.

Let's digress a little to the memory issue. There has been recently a post in our forum with an assumption that Core i7 will make sense even after the appearance of Core i5 on the desktop market, because theoretically, it can be faster in tasks that require high memory bandwidth. Is it really so? Let's think about it.

Three channels right now vs. two channels tomorrow

It's all logical on the surface of it: the i7 has a three-channel memory controller, while the i5 will have a dual-channel controller. Hence, memory system of the former is 1.5 times as fast. But it's only on the face of it -- the number of channels is not the only memory property (otherwise, SiS R658 with its four RDRAM channels would have still been a model desktop chipset). Throughput of the entire memory system depends on each channel, which is almost directly proportional to memory frequency. What about this characteristic? It does not shape up very well for the hypothetical opposition of the i7 and i5.

1. To all appearances, Intel is not planning official support for DDR3-1333 in its Core i7 processors.
2. Core i5 will support at least DDR3-1333, very likely DDR3-1600.

The first assertion does not need proof -- if the company wanted to do it, it would have done it already. Suppose that the company had some problems in the end of last year, which interfered with official DDR3-1333 support for Extreme Edition (it physically supports even higher memory frequencies, but at one's own risk). But there are no problems with that now, even Xeons work with this memory type officially. And the Core i7 upgrades to the new core stepping (even the 920), so Multiplier 10 can be easily legalized. But it hasn't been done. Probably to avoid competition with the Core i5 in future.

The latter will certainly support the latest DDR3-1333 memory, very likely even DDR3-1600. Not because Xeon supports DDR3-1333. Everything is much worse: this memory is officially supported by the AMD AM3 platform. DDR3-1333 is officially supported even now, 1600 -- unofficially. And it's strategically wrong to give a competitor such a noticeable advantage as higher numbers in specifications (most users still hunt for gigahertz). And the delay in the rollout of new processors also plays into the idea to support faster memory. Even right now budget memory kits often include DDR3-1333 models; slower modifications (800 and 1066) may be extinct by summer. So most forecasts agree that even Core i3 will support DDR3-1333, so the top family should work with even faster memory. It's an idea made in heaven.

So, the probable outcomes are: two channels for 1600 or three channels for 1066. They seem to be on a par. But if we dig a little further, we'll see that the first option is faster. Just because absolute timing values abate with the increased frequency. So the resulting performance grows faster with frequency than with the attempt to raise memory bandwidth with more channels. Moreover, reduced latencies accelerate even those applications that do not need high memory bandwidth. Thus, if future processors for LGA1156 support DDR3-1600, they will outperform cheaper Core i7 models (officially, even top models) in applications critical to memory performance, even despite the dual-channel memory controller. If they support only DDR3-1333, the situation will be more complicated -- sometimes the i5 will be faster, sometimes it will be the i7. And finally, if the company 'legalizes' faster memory than DDR3-1066 for the Core i7, there will be no competition between these two families -- the i7 will always be no slower. But we don't see it now! The company hasn't seized this ideal moment (the new stepping and new models on the background of the recent announcement of Xeon.) Thus it isn't planning to do it at all.

And now let's get back to the product under review.

Testbeds

(*) When Turbo Boost is enabled (by default), the real clock rate of separate cores grows by 133-266MHz depending on the current load.
(**) The maximum frequency officially supported by CPU memory controller.
(***) Unlocked for overclocking needs.

We've selected these competitors for a reason -- the Core i7-950 will noticeably outperform the 920 and lag behind the Extreme 965. We are interested in how much: on one hand, their clock rates differ only by 133MHz; on the other hand, slower QPI may influence the results. Theoretically it shouldn't, but things happen. The 950-vs.-X5560 standoff is much more interesting. The former's clock rate is faster by 266MHz (by almost 10%), the latter can boast of extra 266MHz of memory frequency (it makes them different by 25%). This advantage will be partially degraded by higher relative timings. We'll analyze the effect of memory parameters on Nehalem in future. This article is about rough estimates. Besides, Xeon has a full-speed QPI controller. But it theoretically shouldn't make any difference in a single-socket system. Tests will demonstrate the real situation.

Write a comment below. No registration needed!