By: aaron spink (aaronspink.delete@this.notearthlink.net), April 18, 2019 11:42 am
Room: Moderated Discussions
anon (spam.delete.delete.delete@this.this.this.spam.com) on April 18, 2019 11:24 am wrote:
> Are we looking at the same specs? It says 3500 MB/s for both for me.
> Either way that's limited more by the controller, the number of chips and the interface rather
> than the NAND. They could surely build a controller that would squeeze more out of the MLC, but
> what use is a controller that can get 5000-6000 MB/s read speeds with a PCIe 3.0 x4 interface?
>
EVO Plus has the same read bandwidth but higher write bandwidth.
> We're still talking about TLC here, not QLC. Take a look at those and then tell
> me if that's still acceptable. It's debatable but I'm leaning towards no.
>
QLC doesn't change the bandwidth per chip for all intents and purposes (there are slightly higher latencies but for the purposes of these type of workloads it is in the noise). QLC can reach the same practical read bandwidths that TLC can. QLC's issues are write latency and endurance, both of which don't actually really effect sequential performance.
> Correct and I don't think QLC will cut it. What does it offer compared with MLC/TLC
> large enough to cache a few games + an HDD which will definitely be cheaper?
For $80 you'll probably be able to get ~3-4 TB of QLC. HDD have a realistic floor of ~$30. If you want to do flash + HDD using TLC, you are looking at ~1TB of flash + 3-4TB of disk. 3-4TB of QLC is going to cream that in the type of workloads under consideration. The point of QLC is that in read dominate workloads, nothing else can really compete with it from either a cost AND performance perspective. That is why datacenters are seriously looking at switching from HDD to QLC/OLC for nearline in the near future.
> Are we looking at the same specs? It says 3500 MB/s for both for me.
> Either way that's limited more by the controller, the number of chips and the interface rather
> than the NAND. They could surely build a controller that would squeeze more out of the MLC, but
> what use is a controller that can get 5000-6000 MB/s read speeds with a PCIe 3.0 x4 interface?
>
EVO Plus has the same read bandwidth but higher write bandwidth.
> We're still talking about TLC here, not QLC. Take a look at those and then tell
> me if that's still acceptable. It's debatable but I'm leaning towards no.
>
QLC doesn't change the bandwidth per chip for all intents and purposes (there are slightly higher latencies but for the purposes of these type of workloads it is in the noise). QLC can reach the same practical read bandwidths that TLC can. QLC's issues are write latency and endurance, both of which don't actually really effect sequential performance.
> Correct and I don't think QLC will cut it. What does it offer compared with MLC/TLC
> large enough to cache a few games + an HDD which will definitely be cheaper?
For $80 you'll probably be able to get ~3-4 TB of QLC. HDD have a realistic floor of ~$30. If you want to do flash + HDD using TLC, you are looking at ~1TB of flash + 3-4TB of disk. 3-4TB of QLC is going to cream that in the type of workloads under consideration. The point of QLC is that in read dominate workloads, nothing else can really compete with it from either a cost AND performance perspective. That is why datacenters are seriously looking at switching from HDD to QLC/OLC for nearline in the near future.
