THE LATEST VERSIONS of two critical PC I/O interfaces, 4.8Gbps USB3 and 6Gbps SATA3, are finally out in actual products - mostly the on-board interface chips rather than actual peripherals, both of which are still few and far between.
Since the current 3Gbps SATA2 still hasn't reached its performance saturation peak, save with a few high-end SSDs, the more important update is the USB one.
Why? Well, an order of magnitude higher speed for USB3 versus USB2 speaks for itself - 4.8Gbps versus 480Mbps. In addition, the irritating regular CPU polling by USB devices to increase CPU usage, a remnant from the days when Intel supposedly tried all means to create the demand for faster CPUs, is now finally gone. However, since I don't expect any manufacturers to build USB3 compliant keyboards and mice anytime soon, I'll still stick with PS/2 keyboard and mouse input for now, as they still seem to have better reliability and slight performance advantages over the USB2 ones.
Here's the X58A UD7 mainboard, the highest end of the new lot from Gigabyte, with both USB3 and SATA3 additions:
The new USB connector has both USB3 and USB2 pins, so either type of device can be used on the same connector without affecting compatibility. But then, the first implementations have a dedicated - currently an NEC 700200 chip - USB3 controller with its own pair of ports in parallel with the usual dozen USB2 ports coming out of the chipset South Bridge, so I doubt anyone will bother to plug any USB2 devices into the USB3 port at this time.
On the other hand, SATA speed doubling in the newest version was a far simpler matter, with the same connectors and cables still used, and a separate - mostly Marvell SE 9128 at this point - onboard controller for the dedicated 6Gbps SATA3 ports.
Of course, the sped-up USB3 will benefit a far greater range of devices, from ubiquitous flash memory sticks to high-speed scanners, wireless network adapters, even attached USB display connections. The SATA3 benefits will only be seen in high-end SSD drives where read speeds may exceed 300MBps, and even that not yet, as well as selected enterprise 15K RPM hard disks with large DRAM caches that can burst data at the full 6Gbps SATA3 speed. One of the same 15K RPM hard disks would mechanically barely challenge the old SATA1 speed at 1.5Gbps, though.
Now, in the absence of USB3 and SATA3 support in the current chipsets, the vendors had to add dedicated controllers to provide for the new interfaces on the current mainboard designs. Gigabyte was one of the manufacturers that snapped up a lot of these.
Right now, vendors - including Gigabyte - have started rolling out mainboards supporting these new USB and SATA specs on three major chipset platforms, all high end - Intel X58, Intel P55 and AMD 790FX. The X58 and 790FX have a plenty of PCIe V2.0 lanes at high speed that can provide sufficient transfer speed for both of the new I/O specs. Each two-lane or X2 PCIe v2 connection can give you up to a gigabyte per second in each direction, enough for a pair of the newest SATA3 or USB3 devices pushing their interfaces to the fullest, in parallel.
What about the P55? Well, the LGA 1156 Core i5 and Core i7 CPUs only have a total of - very low latency, yes - 16 PCIe v2 lanes coming directly out of the CPU. That's enough to drive one high-end graphics card, but not exactly enough for two high-end GPUs, and surely not enough for two GPU cards and new USB3 and SATA3 devices too.
One alternative is to use the P55 chipset's spare four X1 PCIe v2 lanes, or combine them into one X4 link. However, the 2.5Gbps speed of the links, plus the fact that all of these plus the chipset's own SATA, USB and Gigabit Ethernet connections all go to share a single DMI link to the CPU alerts us to an easily attainable congestion there.
Another approach is to use a PCIe v2 switch on those four chipsets' spare lanes, and provide v2 to v1 PCIe conversion along the way. The benefit is that, between those new generation SATA3 and USB3 devices, you could have full speed communication, as long as it is between these controllers and doesn't go via CPU or main memory. Otherwise, it would get slowed down by the half-speed v1 links to the chipset, and then the equally slow DMI link to the CPU.
Gigabyte tried something else on its UD7 version of the P55 board - it put in a more complex switch that combines using either the faster PCIe v2 lanes off the CPU if you're using just one GPU, or slower PCIe v1 lanes from the P55 chipset when GPU performance is a priority. So, in one GPU setup, you may let the 3D card live off just eight PCIe v2 lanes, and use the rest for the new USB3 and SATA3 controllers, or, when absolute graphics speed is a must, use automatic or forced manual switching to the chipset PCIe v1 lanes for the new I/O while the GPU regains the full 16 PCIe v2 lanes.
How do these approaches compare in actual device performance? SATA3 of course has somewhat higher total throughput potential than USB3, so it would saturate any bus bottleneck on the other side of the controller more easily. During my Taipei trip this past week, I've seen some early performance runs in the Gigabyte labs on its P55 SATA3 and USB3 enabled mainboard.
Now, by no means should these be considered cast in stone since the early SATA3 and USB3 devices don't use nearly the maximum interface potential yet. The controller used on these early SATA3 6Gbps SSDs is barely faster than the currently best SATA2 SSD equivalents. However, the differences are there. Here are the results on the prototype SATA3 SSDs for your own judgment:
Single SATA 3 6Gbps SSD
HD Tach Write
Direct connection - 202.6 MB/s
Bridge via P55 PCIe - 155 MB/s
HD Tune Pro Write
Direct connection - 193.8 MB/s
Bridge via P55 PCIe - 148.3 MB/s
IO Meter Write
Direct connection - 200.1 MB/s
Bridge via P55 PCIe - 152.6 MB/s
HD Tach Burst
Direct connection - 197.6 MB/s
Bridge via P55 PCIe - 182.1 MB/s
HD Tach Read
Direct connection - 215.9 MB/s
Bridge via P55 PCIe - 205.2 MB/s
HD Tune Pro read
Direct connection - 207.2 MB/s
Bridge via P55 PCIe - 195.6 MB/s
PC Mark Vantage index
Direct connection - 23469
Bridge via P55 PCIe - 21388
Total of five SSDs (one on SATA3, the other four on chipset SATA2)
HD Tach Burst
Direct connection - 345 MB/s
Bridge via P55 PCIe - 174 MB/s
HD Tach Read
Direct connection - 293 MB/s
Bridge via P55 PCIe - 174 MB/s
HD Tach Write
Direct connection - 200 MB/s
Bridge via P55 PCIe - 143 MB/s
Hmm, writes are impacted more than reads by the bandwidth and latency loss? And then, look at the massive, nearly half performance loss when overloading both the SATA3 and four SATA2 links all on the chipset lanes together when using five devices. This cries out for more bandwidth, clearly.
As you can see, there is a definite benefit in the direct PCIe V2 lane approach - something easy to achieve on X58 or 790FX chipsets, but not on the P55 platform. Gigabyte's direct switch does enable full SATA3 and USB3 performance even when multiple devices load it fully, but of course at some cost to the GPU performance due to the halved lanes. Still, if getting a new Intel platform mainboard now, and in specific need of fastest possible USB3 and SATA3 I/O, go for the X58 chipset mainboard like the Gigabyte X58A-UD7 available now, or the new mainboards from Asus and DFI that will show up in a month or two. Look for our own USB3 and SATA3 peripherals tests on these new boards, hopefully before Christmas.
In the meantime, with the ten-times speedup for USB and twice the SATA speed gain, Intel will have to re-architect its chipset connection for the "Sandy Bridge" follow-ups to the current LGA1156 mainstream CPU platform. Maybe a PCIe v3 based DMI link with three times to four times the real speed of the current DMI link while keeping a similar pin count would make sense, in the absence of simply adding more PCIe v2 lanes. By that time, we expect a new generation of SATA3 SSDs to hit as much as 500MBps in read performance, so a four-disk RAID array of these would overload the current DMI link nearly three times over. µ
We doubt Big Blue would take payments in voucher form
But the service still has 130 million members, so it's not exactly in trouble
Zhang could be jailed for up to 10 years if found guilty
America seethes as offers are replaced by error pages. With dogs