Details of Intel 45nm Processors Leaked 104
DCC writes "TechARP has gotten some juicy news from Intel. This time, it's the top secret details of the Intel 45nm desktop processors, both Yorkfield and Wolfdale with benchmarks and pricing included! 'As promised earlier, Intel will launch their 45 nm processors by the end of this year. In fact, we have been told that the launch date had already been set at November 11, 2007, so mark your calendars. [...] Code-named Yorkfield XE, the Intel Core 2 Extreme QX9650 will be a quad-core processor built from two 45 nm Wolfdale processor dies. It will displace the Core 2 Extreme QX6850 (Kentsfield) processor as the top desktop processor model until Q3, 2008'"
Not all that new (Score:4, Informative)
Still FSB and dual dual-core (Score:4, Informative)
And The amd 4x4 system with 2 amd quad cores with desktop ram will be alot better then intel Skulltrail with FB-DIMMS and poor chipset io Full sever chipset + 2 nvidia chipset linked by a pci-e x16 bus 1.1 from the intel chipset to the nvidia chip and HT from nvidia to the other nvidia chipset with 2 x16 pci-e 1.1 sli slots. Amd system will cost less with cheaper ram and
a less costly MB.
The amd system will likey have the choice of a nvidia based system with 2 Full sli x16 slots pci-e 2.0 slots + other pci-e 2.0 slots with HT links form the cpus to the nvidia or a
ATI one with
* Codenamed RD790
* Dual or single AMD CPU configuration
* Supports socket AM2+ and socket F CPU
* Allowing maximum of four physical PCI-E x16 slots at x8 lanes bandwidth or 2 PCI-E x16 slots at maximum bandwidth (16x-16x or 8x-8x-8x-8x CrossFire)
* Discrete PCI-E x4 slot
* Providing a total of 52 PCI-E lanes [4], 41 lanes in Northbridge
* Two to four cards CrossFire, with reported 2.6 times of performance than single card
* Support of HyperTransport 3.0
* Support for HTX slots
* Support of PCI-E 2.0
* Supports Dual Gigabit Ethernet, and teaming option
* Discrete chipset cache memory of at least 16 KB to reduce the latencies and increase the bandwidth
* Reference board codenamed "Wahoo" for dual-processor (Quad FX) reference design board with three physical PCI-E x16 slots, and "HammerHead" for single socket reference design board with four physical PCI-E x16 slots, also notable was the reference boards includes two ATA ports and only four SATA 3.0 Gbit/s ports (as being paired with SB600 southbridge), but the final product with SB700 southbridge (see below) should support up to six.
* Northbridge runs at 3 W when idle, and maximum 10 W under load
http://en.wikipedia.org/wiki/AMD_700_chipset_series [wikipedia.org]
Re:Still FSB and dual dual-core (Score:5, Informative)
However the more interesting number was that it took Intel's FSB 77ns to transfer data between the dual-dies, and if the data were only going between cores on the same die that time was only 26ns. So the upshot was, that the worst case scenario for Intel's data latency was less than 2%, while the better case scenario (which is not too hard to achieve) gave Intel a 50% reduction in data latency. If you want to talk about 4 socket+ systems then Hypertransport is a winner, but on a desktop I wouldn't obsess over it too much.
Re:Time for a new naming convention? (Score:3, Informative)
Okay, now how do you mark different versions of that? Ones with different sized caches? Different FSB speeds?
I'm not claiming that the Intel numbers make all that much sense, but they still manage to convey a fair bit of information. Higher "hundreds" digits are faster clocks. (The Q6600 and E6600 both have the same clock speed.) Numbers with the same leading digits, e.g. the E6700 vs. the E6750, are different revisions. (The E6700 has a 1066 MHz FSB, and the E6750 1333 MHz.) The E prefix says that it is dual core; quad cores have Q. If the thousands digit is 2, then things shift around a bit, but that information alone tells you that you're working with one of the budget chips. (Slower clock, no VT, smaller cache.)
Re:Time for a new naming convention? (Score:3, Informative)
Of course, that would insufficient; You would need some other indicator to mention that fact that it is, say, a Wolfdale instead of a Conroe (Wolfdale's being, say, 10% faster). Also imagine that another axis has to be considered; power-efficient, or non-power-efficient. That would make your model name even more complicated: "Conroe4x2200PE". That's quite a mouthful. This is only an example to indicate that specification-based model numbers have a tendency to get prohibitavely complex. The spec-based model number becomes more and more complicated as each axis of of variability is added. So, most companies settle for more arbitrary model numbers (NVidia, ATI, BMW, Lexus, etc).
Consider BMW; for many years, their model numbers were usually of the form *model**displacement**feature* (an incomplete schema would be [3,5,7][15,20,25,30,35,40,50][i,ci]). However, at some point, they realized that there might be several different models with the same engine size, and rather than making the feature indicator more complex, they made displacement inaccurate (sometimes pro-rating it to indicated a higher performance version of an identical displacement engine).
All that to say, I don't think that having arbitrary model number is such a bad state of affairs; just have a publically accessible lookup table that relates a single model number to all the relevant specs. Saying "NVidia 7600GT" is a heck of a lot easier than saying "NVidia 256-128-22.4-6.7-700" (for memory, memory interface, memory bandwidth, fill rate, and vertices-per-second).
Re:here we go again (Score:4, Informative)
Of course, there are many parts of a CPU that traditionally don't scale as well as the basic transistor, so with continued work, we can probably keep shrinking CPUs. But we'll be doing it in small increments with increasing marginal costs, not by the factors of 2 we've been seeing for the past 20 years.
Flawed Analysis (Score:2, Informative)
The problem is right there in the Author's analysis. For example:
"If you extrapolate the data, then the Yorkfield processor is really about 12-21% faster than the Kentsfield at the same clock speed. This is almost entirely due to the 50% larger cache in the Yorkfield processor. The very large 81% boost in DivX 6.6.1 is again mostly due to SSE4-optimized code in DivX."
But But But!!!! Changing the clock speed doesn't make the cache any bigger! You can't then assume a linear relationship between performance and clockspeed if the difference is primarily how long you are going to have to wait to fill the cache!
The article isn't too flawed. They give actual results. But do yourself a favor as you read the aritcle and completely discount any "extrapolation" done by the author to get "really" numbers. When comparing processors, the "really" numbers are always the hard cold facts, not the "I wish" numbers generated by speculating what would happen if you changed the processors in some way.
Re:Not all that new (Score:3, Informative)
I know -- it's been a weird release, to say the least. I haven't really heard very much about them at all, and for a chip this neat that's kind of surprising.
You can buy them now on Newegg here [newegg.com] -- they were up a few weeks ago for about $800, but then they were taken down, and now they're back up. Who knows, eh?
Re:Time for a new naming convention? (Score:5, Informative)
For a 1MB cache (per core) cpu, it's exactly 2x the clock speed in megahertz. The X2 4000+ is 2000MHz. This continues every 200MHz all the way up to the top cpu, the 6400+ (3200MHz, 1MB cache).
For a 512kB cache (per core) cpu, it's 200 lower than that. The X2 3800+ is 2000MHz as well, but 512kB cache. This continues every 100MHz all the way up the line to the 5400+ (2800MHz, 512kB cache).
For a 256kB cache (per core) cpu, it's 200 lower again. The X2 3600+ is ALSO 2000MHz, but has 256kB cache. There is only one 256kB cache X2 cpu. There is also a X2 3600+ that is 1900MHz and 512kB cache, which still fits the pattern.
The single core Athlon 64s seem to have a similar numbering scheme, but with more factors affecting it, including hypertransport speed (800MHz/1000MHz), and socket (754/939). Some of the cpus were numbered slightly differently, but this is 99% accurate:
The base is a 512kB cache socket 754 hypertransport 800MHz 2000MHz cpu, which is rated at 3000.
Socket 754 cpus were rated 200 higher for every 200MHz higher cpu speed.
1MB cache versions were mostly 200 higher (one was 300), and 256kB cache versions were 100 lower.
Socket 939 cpus were rated 200 higher than socket 754. (Due to the support for dual-channel ddr, they were better).
1000MHz HT cpus were rated an additional 100 higher for every 200MHz higher cpu speed than the base (2000MHz). The cpus that were 200MHz slower than the base didn't get an additional 100 points deducted though.
Again, 1MB cache versions were 200 higher.
This doesn't cover the 1500+, which was only used in a HP Blade PC.
The AM2 cpus were mostly the same as the 1000MHz HT S939s, except for the 4000+, which was a 2600MHz/512kB cache instead of 2400MHz/1MB, and the details above would have scored it at 4100+.
As you can see, the numbers are mostly arbitrary, and mostly derived from the features of the cpus instead of a comparison against intel.
Re:Yep (Score:3, Informative)
There is a perception that AMD's solution is more elegent and regardless of benchmarks is somehow "better". Intel's design is a "throw cache at the problem" sort of solution--but it works for most normal usage.
I suspect that many people would like to see what Intel could do if they got off their seats and really did something original...like if they can do this good with half-assed lashups, how good could their cpus be if they actually did some novel *design*!?!