upcoming SSDs

[offandon]

I am going to build an Ivy Bridge PC when it is released and I am trying to find the best deals on the components I want in the rig. Some things I think it makes sense to get now if a great deal presents itself (case, power supply, OS at the very least). As things stand now I am targeting the Samsung 830 256GB SSD or perhaps the Intel 520 to be my boot and apps drive. I have not seen anything which would indicate that there is some great new thing coming in the next month that will change my opinion on what to buy. I am considering buying one of these SSDs if a great deal presents itself but wanted to see if I am just out of the loop too much on any new potential SSDs coming out before May. I would kick myself if I buy something now and there is a "must have" drive that comes out in the next 4-5 weeks.

Anyone see or hear anything?

 

[groberts101]

While there are surely more models coming into the market based on existing controllers.. the only thing coming soon which would be running a newer controller would be the Vertex 4. And while I think that will be a decent drive.. not sure I would call it a "must have".

Others based on newer Marvell/Sandforce controllers will come several months from now and many of them will be scoring well over 1000 in AS SSD because of huge randoms and faster writes.

 

[Coup27]

I would stick with your two choices.

As Roberts as said, the Vertex 4 is due out which will be based on the Indilinx Everest 2 but beyond that nothing has been announced. I certainly think for the rest of this year at least the two you have highlighted will the best choices.

 

[offandon]

Thank you for your replies ... I will go with one of these when a good deal pops up.

 

[Burner27]

Just when i finally secured 4 x 256GB M4s too......

 

[exdeath]

SATA 6G is going to remain the bottleneck for the next foreseeable year or two, so don't lose sleep over new SSDs just yet.

We are stuck right now at 550 MB/s sequential and 300 MB/s randoms for now . Yes SATA 6G limits randoms too, remember its a serial bus. There is a ton of unseen command and protocol overhead over the SATA wire requesting thousands and millions of 4Ks individually.

 

[groberts101]

While 6G may certainly limit the random performance.. it will not be anywhere near a 300MB/s limit. Especially since I've already seen that assumed single port limit smashed all to hell and think we may even see 400MB/s capability by the end of the year.

And aside from compressible data on Sandforce.. there is a huge amount of improvement available from the "other" next gen 6G controllers on the sequential writes.

Only newer controller's, tweaked firmware, and time will tell, though.

 

[exdeath]

While 6G may certainly limit the random performance.. it will not be anywhere near a 300MB/s limit. Especially since I've already seen that assumed single port limit smashed all to hell and think we may even see 400MB/s capability by the end of the year.

And aside from compressible data on Sandforce.. there is a huge amount of improvement available from the "other" next gen 6G controllers on the sequential writes.

Only newer controller's, tweaked firmware, and time will tell, though.

Well we aren't going to see 550 MB/sec randoms no matter what is my point. Millions of 4k random requests from the host to the SATA controller are inefficient and command and control packets are going to utilize a significant portion of available SATA wire bandwidth regardless how fast the controller is.

I agree there is room for improvement still, but not much. Anyone investing in a SF 2281 with toggle NAND right this minute is going to be sitting very pretty for at least a year or two, excepting for cheaper and higher capacity drives of course.

Personally the most significant thing I see in SSD improvement before SATA 12g is controllers that are able to handle more channels and deal with the higher work load of sorting through 512GB worth of NAND such that 512MB drives become as fast or faster than current performance leading 256GB class drives. There isn't much room left with SATA 6G in the way, but 512GB drives are not even competitive with class leading 256GB drives; they are often rated at HALF the IOPs on the same controller.

I look forward to getting the performance out of a single channel single drive that I currently do on 4 x RAID... but then I would just RAID 4 of those as well Something about right click drag copy here on a 5 GB file and not even seeing a dialog box appear has a way of twisting a person's sense of reality.

 

[exdeath]

Just when i finally secured 4 x 256GB M4s too......

You'll be fine A 2012 Ferrari isn't anything to hate just because a 2013 Ferrari was announced

 

[KingFatty]

Well we aren't going to see 550 MB/sec randoms no matter what is my point. Millions of 4k random requests from the host to the SATA controller are inefficient and command and control packets are going to utilize a significant portion of available SATA wire bandwidth regardless how fast the controller is.

Does the control request info really take up much space at all?

If it's a small amount of information, then even if you have millions of them, it's still relatively small compared to the overall capacity of SATA 6G right?

I just can't help visualizing it like tiny drips of requests compared to a fire-hose of data.

 

[exdeath]

Does the control request info really take up much space at all?

If it's a small amount of information, then even if you have millions of them, it's still relatively small compared to the overall capacity of SATA 6G right?

I just can't help visualizing it like tiny drips of requests compared to a fire-hose of data.

Well SATA 6G is 600 MB/sec raw physical bandwidth.

6 gigabit = 6000 megabit, with 8/10 encoding = 4800 megabit = 600 Megabyte /sec

The fastest any SSD can sustain sequential is only around 550 MB/sec for SATA 6g, 260-280 MB/sec for SATA 3g, etc. So we know that even for simple large block transfers, we get some solid evidence of significant overhead with the best case scenario (over 8% of available bandwidth unaccounted for). So even for large sustained sequential transfers with the least amount of command and control overhead, a whopping 50 MB/sec of SATA bandwidth is still lost to protocol overhead for relatively simple LBA start-end streaming block requests.

So now extrapolate to millions of separate requests and you can see that it's going to be worse. I don't know SATA protocols by the book, but you can be sure there are command packets from the host to the drive to request the data, interrupt commands back to the host controller, "data ready" commands and status updates, "window" throttling for tx/rx buffers/shift registers full, and finally the data itself encapsulated by whatever headers ,etc. Having that data duplicated millions of time for individual 4k requests consumes a serious chunk of bandwidth, not to mention the stop/start overhead of flooding the controller on both ends with commands. The data sent per packet is the same regardless if you only need a dozen packets or a 10 million packets. Obviously it consumes a higher percentage of user bandwidth in the latter.

It's the same story with SMB over Ethernet, jumbo packets, etc. 20 GB file is going to copy much faster than 20GB worth of 4k files because there is going to be so much wire protocol overhead occupying physical wire bandwidth even if you have infinite memory, I/O, and processor speed on both ends of the wire. Transaction overhead stealing user bandwidth is one of the drawbacks to serial buses compared to parallel buses with separate address/data/command lines that can keep the data channel saturated at the maximum physical bandwidth under all conditions, random or otherwise.

 

[KingFatty]

I want to believe what you are saying, and I like the analogy to Ethernet.

But when I look at the real-world facts, I don't see how it could be correct?

For example, if the SATA 6G is limiting the 4K randoms, why do different SSDs perform so differently? I mean, look at this for example, you see a pretty big range of performances:
http://www.tomshardware.com/charts/ssd-charts-2011/AS-SSD-4K-Random-Read,2784.html

So because the numbers seem to improve from one SSD to the next, I can't help but think there is still much room for improvement in SSDs to get better 4K random scores?

 

[hal2kilo]

Well SATA 6G is 600 MB/sec raw physical bandwidth.

6 gigabit = 6000 megabit, with 8/10 encoding = 4800 megabit = 600 Megabyte /sec

The fastest any SSD can sustain sequential is only around 550 MB/sec for SATA 6g, 260-280 MB/sec for SATA 3g, etc. So we know that even for simple large block transfers, we get some solid evidence of significant overhead with the best case scenario (over 8% of available bandwidth unaccounted for). So even for large sustained sequential transfers with the least amount of command and control overhead, a whopping 50 MB/sec of SATA bandwidth is still lost to protocol overhead for relatively simple LBA start-end streaming block requests.

So now extrapolate to millions of separate requests and you can see that it's going to be worse. I don't know SATA protocols by the book, but you can be sure there are command packets from the host to the drive to request the data, interrupt commands back to the host controller, "data ready" commands and status updates, "window" throttling for tx/rx buffers/shift registers full, and finally the data itself encapsulated by whatever headers ,etc. Having that data duplicated millions of time for individual 4k requests consumes a serious chunk of bandwidth, not to mention the stop/start overhead of flooding the controller on both ends with commands. The data sent per packet is the same regardless if you only need a dozen packets or a 10 million packets. Obviously it consumes a higher percentage of user bandwidth in the latter.

It's the same story with SMB over Ethernet, jumbo packets, etc. 20 GB file is going to copy much faster than 20GB worth of 4k files because there is going to be so much wire protocol overhead occupying physical wire bandwidth even if you have infinite memory, I/O, and processor speed on both ends of the wire. Transaction overhead stealing user bandwidth is one of the drawbacks to serial buses compared to parallel buses with separate address/data/command lines that can keep the data channel saturated at the maximum physical bandwidth under all conditions, random or otherwise.

Won't Thunderbolt surpass Sata 12 ?

 

[exdeath]

I want to believe what you are saying, and I like the analogy to Ethernet.

But when I look at the real-world facts, I don't see how it could be correct?

For example, if the SATA 6G is limiting the 4K randoms, why do different SSDs perform so differently? I mean, look at this for example, you see a pretty big range of performances:
http://www.tomshardware.com/charts/ssd-charts-2011/AS-SSD-4K-Random-Read,2784.html

So because the numbers seem to improve from one SSD to the next, I can't help but think there is still much room for improvement in SSDs to get better 4K random scores?

There is room for improvement in SSD controller and NAND performance still, but due to wire bandwidth and protocol overhead there is a hard cap on randoms that will be always be lower than sequential even with infinitely fast controllers and NAND.

To clarify though, my perspective is from 4KQD32 numbers that are at 300 MB/s on many top tier drives. The sub 50 MB/sec random 4K QD 0 scores have tremendous opportunity to improve in terms of wire bandwidth, but likely won't see that much improvement either given that programming and accessing flash memory takes significant time (from the perspective of the controller), and lack of queuing does not allow the controller to pipeline and interleave flash commands. Having to open/close a command to the NAND flash one at a time without being able to aggregate multiple data requests is very inefficient. There is also significant host controller waste. As far as I know, a 4K random bench with QD=0 has the OS perform a single 4k operation, and waits for completion from end point to end point from the SATA controller drivers to the NAND flash and back, before issuing the next command.

 

[exdeath]

Won't Thunderbolt surpass Sata 12 ?

Yeah, 20 gbit > 12 gbit, but Thunderbolt is like USB sharing bandwidth amongst many devices including your uncompressed video signal to your display.

I wouldn't want my theoretical 2,500 MB/sec single channel SSD to be sharing bandwidth with a uncompressed 2560x1600x4x60 display signal in the universal application Thunderbolt is envisioned for.

Dedicated 20 gbit ports and controllers for SSDs would be welcome.