I just want to know how much they paid. That figure alone won't indicate exactly how close their controller technology is to market, but it will indicate exactly how far ahead Micron believes it to be of the competition, including Intel's own house controllers.
Then again, Toshiba paid $65 million for OCZ, a steal in reality because even though the name isn't worth much, the Indilinx IP was. In my opinion the Barefoot 3 is one of the best SSD controllers ever made. It is incredibly consistent, reliable, and inexpensive. It's only real downside is power consumption\lack of DevSleep. I just don't understand why they can't add DevSleep to it when Toshiba engineers were able to tweak the latest revision to support TLC...
Consistent? Heck no. It had a pretty good average performance in steady state conditions, but the standard deviation was quite large. It was very inconsistent.
Compared to what? Barefoot is a very consistent controller, besting most controllers of its day (that being...2012) such as Sandforce, Samsung, Sandisk and Marvell. It was originally going to be OCZ's heavy IO (server) solution.
In fact, almost no controller other than Intel's were as consistent, with modern Samsung and Marvell controllers catching up. But that isn't to say it's fast. It's just that SSD performance (as in speed) is about as blown out of proportion as PPI in smartphones...at the end of the day, nobody is going to notice the difference between 4xxMB/sec and 5xxMB/sec.
Steady state and garbage collection are top notch for the price, IMHO. I also haven't seen a single modern OCZ drive fail, including a few ARC100's I have out in the field.
If you look at my comments...years ago...I swore I'd never own an OCZ product again after my Agility drive failed in literally a week. I eBay'd the replacement and went back to Intel. After OCZ released Barefoot I was intrigued, because I firmly believed Sandforce was the thorn in OCZ side. To date, I've had nothing but Sandforce-based SSD's fail, ranging from ADATA to even Intel. The controller (SF2281) is a ticking timebomb especially in machines were a power failure could happen or you might accidently fill the drive to capacity.
The aforementioned ARC100 scores slightly better on newegg, If i look at a random budget intel drive (intel 335) on newegg, the number of dead drives in reviews is around 2%, the ARC100 is above 10%.
You do realize you are comparing a controller from 2012 to controllers from 2014-2015 in that chart...which ironically also shows the Barefoot besting everyone in steady-state performance, showing the most optimized garbage collection of the bunch, while still maintaining consistency.
"In my opinion the Barefoot 3 is one of the best SSD controllers ever made. - Samus, 2015.
You never said "of its time", you said "ever made". You also said it had great consistency, of which it was terrible for consistency. Good average steady state performance is not great consistency if it peaks and troughs frequently. Granted, a high average throughput works well for a client workload, but you're mixing up the terminologies.
Stop being a ridiculous politician and taking my words out of context. I stand behind my statement that Barefoot 3 is one of the best controllers ever made, in the same way I can overwhelmingly justify stating the Intel PC29AS21 controller from the X25-M and SSD320 is THE best controller ever made and the Ford Model-T is the most important CAR ever made. Does that make them the best today? Hell no they're both antiques, but that is irrelevant in their importance in that they began a revolutionary movement away from the "old way" of doing things. Not surprisingly, the X25-M still has value today, much like a Barefoot 3 drive. Sure you can buy a new Crucial MX200 or Samsung EVO 850 with more performance, but where is the track record of these drives? Are they really that much better to justify the cost and potential reliability issues like their previous-gen counterparts?
Looking at those graphs it strikes me that the way anandtech currently plots them isn't very helpful.
Most people probably aren't interested in "performance consistency" per se, but reliably low latencies. And that means that ops / second is much less useful than seconds / op - afterall, you probably don't care about all those outliers in the 100000 iops range - a few *fast* ops aren't the problem - but the actually problematic range is all bunched up to a tiny part of the graph on the bottom. *zero* iops is the problem; and the difference between 2 and 20 iops in a hiccup matters a lot.
Also, standard deviation is the wrong measure here - again we don't care about ops that are too fast, nor even really ops that are slightly too slow; we care about the amount of outliers. standard deviation doesn't emphasize outliers enough; and it can make *faster* systems with identical maximum latency look worse (which is weird). I'd rather see maximum latency and a few percentiles (e.g. 99.9999, 99.99, 99).
On the topic of the noisy looking diagram for the vector: that wouldn't bother me. What does bother me are those dots on or near zero. That sounds like a system stall to me, and that's terrible. But these graphs don't quite make clear how terrible...
I'm pretty sure that reporting 99.99th percentile in our consistency tests wouldn't provide useful information for almost any real-world use. It's a problem if the whole drive stops delivering data for milliseconds at a time, but real-world workloads that have a queue depth of exactly one also have idle time for the drive instead of being a constant stream of only writes. We already report some latency percentiles for real-world access patterns as part of the ATSB tests.
The world gets along fine with occasional GC pauses in RAM, because most stuff that uses a garbage collected memory model has plenty of idle time to run the GC. If your storage requirements are hard realtime but with throughput requirements that stress a SSD, then the solution is to get a battery backup for your RAM, not search for the holy grail of garbage collectors to use with flash.
It's not realistic to automatically assume that Tidal's controllers are better than anything else on the market. They sure have some talented and experienced people behind the controller, but so do Intel, Samsung and every other SSD or controller company. The truth is that Micron really needs its in-house controller because being at the mercy of Marvell and SMI has handicapped the business from what it could be.
Remember that Toshiba acquired OCZ's assets after OCZ had filed for bankruptcy, so $35 million wasn't exactly market value. Besides, OCZ didn't have much choice because declining the deal would likely have been the end of the company and getting acquired by a NAND company was basically the only way OCZ could be competitive in the future (no more NAND sourcing issues).
Also, implementing DevSleep to Barefoot 3 would have required a new silicon design. OCZ tried everything they could with the current silicon, but power gating is one of those things that has to be put into the design from early on. A new silicon at this point wouldn't make much sense (and it would be costly), so OCZ is putting all resources on JetExpress that will feature DevSleep as well.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
12 Comments
Back to Article
Samus - Friday, October 9, 2015 - link
I just want to know how much they paid. That figure alone won't indicate exactly how close their controller technology is to market, but it will indicate exactly how far ahead Micron believes it to be of the competition, including Intel's own house controllers.Then again, Toshiba paid $65 million for OCZ, a steal in reality because even though the name isn't worth much, the Indilinx IP was. In my opinion the Barefoot 3 is one of the best SSD controllers ever made. It is incredibly consistent, reliable, and inexpensive. It's only real downside is power consumption\lack of DevSleep. I just don't understand why they can't add DevSleep to it when Toshiba engineers were able to tweak the latest revision to support TLC...
Gigaplex - Friday, October 9, 2015 - link
Consistent? Heck no. It had a pretty good average performance in steady state conditions, but the standard deviation was quite large. It was very inconsistent.Samus - Saturday, October 10, 2015 - link
Compared to what? Barefoot is a very consistent controller, besting most controllers of its day (that being...2012) such as Sandforce, Samsung, Sandisk and Marvell. It was originally going to be OCZ's heavy IO (server) solution.In fact, almost no controller other than Intel's were as consistent, with modern Samsung and Marvell controllers catching up. But that isn't to say it's fast. It's just that SSD performance (as in speed) is about as blown out of proportion as PPI in smartphones...at the end of the day, nobody is going to notice the difference between 4xxMB/sec and 5xxMB/sec.
Steady state and garbage collection are top notch for the price, IMHO. I also haven't seen a single modern OCZ drive fail, including a few ARC100's I have out in the field.
If you look at my comments...years ago...I swore I'd never own an OCZ product again after my Agility drive failed in literally a week. I eBay'd the replacement and went back to Intel. After OCZ released Barefoot I was intrigued, because I firmly believed Sandforce was the thorn in OCZ side. To date, I've had nothing but Sandforce-based SSD's fail, ranging from ADATA to even Intel. The controller (SF2281) is a ticking timebomb especially in machines were a power failure could happen or you might accidently fill the drive to capacity.
Darkstone - Saturday, October 10, 2015 - link
Are you sure? OCZ Octane drives are a streaming pile of crap on the reliability front, as is OCZ vector according to the number of 1-star reviews on newegg: http://www.newegg.com/Product/Product.aspx?Item=N8...The aforementioned ARC100 scores slightly better on newegg, If i look at a random budget intel drive (intel 335) on newegg, the number of dead drives in reviews is around 2%, the ARC100 is above 10%.
Gigaplex - Saturday, October 10, 2015 - link
Compared to lotshttp://www.anandtech.com/show/9009/ocz-vector-180-...
Samus - Saturday, October 10, 2015 - link
You do realize you are comparing a controller from 2012 to controllers from 2014-2015 in that chart...which ironically also shows the Barefoot besting everyone in steady-state performance, showing the most optimized garbage collection of the bunch, while still maintaining consistency.That is the consistency I am talking about.
Gigaplex - Sunday, October 11, 2015 - link
"In my opinion the Barefoot 3 is one of the best SSD controllers ever made. - Samus, 2015.You never said "of its time", you said "ever made". You also said it had great consistency, of which it was terrible for consistency. Good average steady state performance is not great consistency if it peaks and troughs frequently. Granted, a high average throughput works well for a client workload, but you're mixing up the terminologies.
Samus - Sunday, October 11, 2015 - link
Stop being a ridiculous politician and taking my words out of context. I stand behind my statement that Barefoot 3 is one of the best controllers ever made, in the same way I can overwhelmingly justify stating the Intel PC29AS21 controller from the X25-M and SSD320 is THE best controller ever made and the Ford Model-T is the most important CAR ever made. Does that make them the best today? Hell no they're both antiques, but that is irrelevant in their importance in that they began a revolutionary movement away from the "old way" of doing things. Not surprisingly, the X25-M still has value today, much like a Barefoot 3 drive. Sure you can buy a new Crucial MX200 or Samsung EVO 850 with more performance, but where is the track record of these drives? Are they really that much better to justify the cost and potential reliability issues like their previous-gen counterparts?emn13 - Sunday, October 11, 2015 - link
Looking at those graphs it strikes me that the way anandtech currently plots them isn't very helpful.Most people probably aren't interested in "performance consistency" per se, but reliably low latencies. And that means that ops / second is much less useful than seconds / op - afterall, you probably don't care about all those outliers in the 100000 iops range - a few *fast* ops aren't the problem - but the actually problematic range is all bunched up to a tiny part of the graph on the bottom. *zero* iops is the problem; and the difference between 2 and 20 iops in a hiccup matters a lot.
Also, standard deviation is the wrong measure here - again we don't care about ops that are too fast, nor even really ops that are slightly too slow; we care about the amount of outliers. standard deviation doesn't emphasize outliers enough; and it can make *faster* systems with identical maximum latency look worse (which is weird). I'd rather see maximum latency and a few percentiles (e.g. 99.9999, 99.99, 99).
On the topic of the noisy looking diagram for the vector: that wouldn't bother me. What does bother me are those dots on or near zero. That sounds like a system stall to me, and that's terrible. But these graphs don't quite make clear how terrible...
Billy Tallis - Sunday, October 11, 2015 - link
I'm pretty sure that reporting 99.99th percentile in our consistency tests wouldn't provide useful information for almost any real-world use. It's a problem if the whole drive stops delivering data for milliseconds at a time, but real-world workloads that have a queue depth of exactly one also have idle time for the drive instead of being a constant stream of only writes. We already report some latency percentiles for real-world access patterns as part of the ATSB tests.The world gets along fine with occasional GC pauses in RAM, because most stuff that uses a garbage collected memory model has plenty of idle time to run the GC. If your storage requirements are hard realtime but with throughput requirements that stress a SSD, then the solution is to get a battery backup for your RAM, not search for the holy grail of garbage collectors to use with flash.
Samus - Saturday, October 10, 2015 - link
http://www.anandtech.com/show/6363/ocz-vector-revi...Kristian Vättö - Saturday, October 10, 2015 - link
It's not realistic to automatically assume that Tidal's controllers are better than anything else on the market. They sure have some talented and experienced people behind the controller, but so do Intel, Samsung and every other SSD or controller company. The truth is that Micron really needs its in-house controller because being at the mercy of Marvell and SMI has handicapped the business from what it could be.Remember that Toshiba acquired OCZ's assets after OCZ had filed for bankruptcy, so $35 million wasn't exactly market value. Besides, OCZ didn't have much choice because declining the deal would likely have been the end of the company and getting acquired by a NAND company was basically the only way OCZ could be competitive in the future (no more NAND sourcing issues).
Also, implementing DevSleep to Barefoot 3 would have required a new silicon design. OCZ tried everything they could with the current silicon, but power gating is one of those things that has to be put into the design from early on. A new silicon at this point wouldn't make much sense (and it would be costly), so OCZ is putting all resources on JetExpress that will feature DevSleep as well.