Basic question about interface.

[Denithor]

SATA 3.0 provides 3Gbit/s transfer rates (375MB/s) that seems to limit today's SSD drives to around 270MB/s reads.

PCIe 2.0 provides 500Mbit/s (62.5MB/s) per lane. i7 offers 32 lanes so 2GB/s total. So if you have one GPU using 16 lanes you still have 16 remaining available (with 1GB/s bandwidth - nearly three times what is available through a SATA channel).

Why don't we have ExpressCard-type drives that can be plugged into an unused PCIe slot on the motherboard? It seems to me this kind of setup would provide significantly faster throughput and would move the read limit up considerably for SSD drives.

This could potentially improve even further with i5 and the reduced latency due to the PCIe lanes being incorporated into the CPU.

Or is there something obviously wrong with this concept I'm overlooking?

 

[MerlinRML]

Originally posted by: Denithor
SATA 3.0 provides 3Gbit/s transfer rates (375MB/s) that seems to limit today's SSD drives to around 270MB/s reads.

PCIe 2.0 provides 500Mbit/s (62.5MB/s) per lane. i7 offers 32 lanes so 2GB/s total. So if you have one GPU using 16 lanes you still have 16 remaining available (with 1GB/s bandwidth - nearly three times what is available through a SATA channel).

Why don't we have ExpressCard-type drives that can be plugged into an unused PCIe slot on the motherboard? It seems to me this kind of setup would provide significantly faster throughput and would move the read limit up considerably for SSD drives.

First, your PCIe numbers are wrong. It's 500MBytes per lane, not Mbit.

Second, the practical side of things. Even though SATA currently only provides 3Gb/port, it's still more than enough for a single mechnical hard drive. It might be limiting burst rates slightly, but burst rates don't play a big factor in real world use. The only cases where bandwidth is limited are SSD and port multiplier. Luckily, 6Gb SATA should be available in the near future (my personal guess is that we'll see it next year included in the chipsets from all the regular vendors).

And in reality, you are using your PCIe bus to access your storage. The SATA ports that your storage connects to are run from the ICH of your chipset which usually has a PCIe 4x connection to the rest of the system.

Still, your idea is not a bad one, and Fusion-io has made a PCIe SSD. What they've done is they've essentially built an I/O controller into their storage device. It's great because they can control the entire path from storage controller to storage device and everything in between. They can optimize everything to get the best performance. The problem with their design is cost. They're building a proprietary storage controller into every one of their devices. While it provides great performance, you have to pay them for their work on designing and building the controller and the storage. I wouldn't be surprised if there were other niche vendors out there doing the same thing.

 

[taltamir]

second reason is that until current SSD came about, no regular drive could even saturate SATA1, not even the velociraptor. There is not a single spindle drive on earth that can saturate SATA1!
SATA2 has been forward thinking in this regards, and arguably useless at the time...
This is why SATA3 is coming about, with double the speed of SATA2

 

[Denithor]

So a single PCIe lane provides 1/3 more bandwidth than a SATA 3.0 channel? Wowsers.

That makes my point even more apt.

If we only had the SATA 1.5 interface today we would look at SSD performance and think they were equivalent to the best mechanicals available (VR nearly saturating the available bandwidth, SSD severely restricted by limited bandwidth - these would therefore be turning in very similar performance numbers).

I wonder just how fast today's SSD drives could go if we already had SATA 6.0 or even a direct PCIe link. Or are the controllers designed with the maximum available bandwidth in mind - limiting speeds through firmware?

By far the slowest piece of any computer is the HDD. I just think it would be cool if you could buy an SSD that plugged directly into an available PCIe x1/x16 slot and gave you a considerable boost in performance.

 

[jimhsu]

You can but not for a "consumer-friendly" price - e.g. 80GB for $900 (ioExtreme)

 

[lopri]

I think Intel's Braidwood technology will attempt to take advantage of SSD+PCIe. It's due early next year.

Don't forget that you need drivers for PCIe devices. It won't be fun having to have drivers per different HDDs. (vendors, models,..)

 

[taltamir]

Originally posted by: Denithor
So a single PCIe lane provides 1/3 more bandwidth than a SATA 3.0 channel? Wowsers.

That makes my point even more apt.

If we only had the SATA 1.5 interface today we would look at SSD performance and think they were equivalent to the best mechanicals available (VR nearly saturating the available bandwidth, SSD severely restricted by limited bandwidth - these would therefore be turning in very similar performance numbers).

I wonder just how fast today's SSD drives could go if we already had SATA 6.0 or even a direct PCIe link. Or are the controllers designed with the maximum available bandwidth in mind - limiting speeds through firmware?

By far the slowest piece of any computer is the HDD. I just think it would be cool if you could buy an SSD that plugged directly into an available PCIe x1/x16 slot and gave you a considerable boost in performance.

what? we are just finally reaching speeds high enough to be limited by SATA2.