For related news, see:-
http://www.asus.com/news_show.aspx?id=1304
The Socket 939 nForce X16 header was added to the Products list
yesterday ( Sept 14 ) but no contents yet; presumably some detailed
product information is coming shortly. Shipment in September now
seems doubtful, regardless of the assertions of the news article.
See:-
http://www.asus.com.tw/products2.aspx?l1=3&l2=15
John Lewis
I've got a question for you - how will 4GB/sec worth of
Hypertransport bus, feed 8GB/sec worth of PCI Express
lanes ? I'm thinking those new lanes will only be
half busy at the best of times.
For Hypertransport - 16 bits up and 16 bits down:
1000MHz x 2 (DDR) x 16 bit bus / 8 bits_per_byte = 4GB/sec
There is a 4GB/sec upward HT bus and a 4GB/sec downward HT bus.
For PCI Express - separate up and down lanes are used:
250MB/sec PCI-Express lane x 16 = 4GB/sec
There is a 4GB/sec upward set of lanes and a 4GB/sec
downward set of lanes.
With two video cards running PCI-E x16, that is twice
the bandwidth available over the Hypertransport interface
to the processor (and the AMD internal memory controller).
Hypertransport appears to be a bottleneck here.
While a single card in the SLI card pair can burst at
the full rate, when the second card gets in on the action,
there are limits.
I await the first benchmarks for this "improvement".
On the Intel side of this innovation, at least the
Northbridge memory controller has more theoretical
bandwidth to offer. Two sticks of DDR2-667 offers 10.7GB/sec
bandwidth, according to the 955X Northbridge datasheet.
If two PCI-E x16 video cards were transferring information
at the same time, the memory controller might be a
better match for their 8GB/sec aggregate bandwidth in
one direction. Whether PCI Express video cards transfer
info in both directions at the same time, full tilt, is
an interesting question.
Judging by the benchmarks here, I wonder if these new
motherboards will even add 2% to overall performance.
Not much of an incentive to suggest throwing away an
existing SLI motherboard.
http://graphics.tomshardware.com/graphic/20041122/pcie-analyse-09.html
Paul