HMB vs DRAM Cache on Gen4 and Gen5 drives?

[ImServs]

Looking at buying a 4TB NVMe drive and was wondering if this should still be a consideration.

In the past, DRAM served a bigger purpose, and made a huge difference on performance, but with the speed increases on read/write as well as the growth of error correction, is spending $50-$100 more for DRAM worth it (on higher capacity drives obviously)? Is there any reason to argue HMB is significantly worse than DRAM cache with all the improvements with SSDs?

[NobleGamer]

14 minutes ago, ImServs said:

DRAM served a bigger purpose, and made a huge difference on performance

You talk about like it doesn't, yet it still does within it's the same pcie version - Its just a matter of whether how you use it can benefit from where DRAM tends to make a difference: Lower latency, sustained reads & write throughput, and perhaps random reads/writes throughput. Consider that HMB size varies by manufacturer, but it's generally 60-100MB. I don't believe it uses all your available RAM or anything. So cache will deplete VERY quick compared to gigs of DRAM.

 

A pcie 4 DRAM-less can have better advertised throughput than pcie 3 with DRAM, but what they don't advertise is whether it's better in those areas I mentioned above. You'll leave to research benchmarks that you trust on that. I personally use pcppartpicker.com since their random and sustained read/write tests seem to have solid methodology.

[Eigenvektor]

1 hour ago, ImServs said:

Is there any reason to argue HMB is significantly worse than DRAM cache with all the improvements with SSDs?

It uses some of your RAM, rather than something included on the SSD. So not only will you have to sacrifice some of your RAM, latency will always be larger than a dedicated resource that sits much closer to where it is needed. For most uses cases the raw speed of a drive is also far less important than its IOPS.

 

~edit: I should also add that the speed of Gen5 drives is more or less irrelevant on desktop. You'll not notice any difference in day to day operation, unless that consists of copying terabytes of data between two such drives. Even going from Gen3 to Gen4 make very little difference in most cases.

[kokosnh]

1 hour ago, ImServs said:

Is there any reason to argue HMB is significantly worse than DRAM cache with all the improvements with SSDs?

It is, and always was…

and what improvements do you mean?  

was there any? What do you mean? 

do you mean the feature of HMB in NVMe 1.2 vs 1.3 vs 1.4 vs 2.0 spec? 

 

DRAM is used to cache a FTL, so a table of data allocation in all NAND memory cells of the SSD. 

if the ssd don’t have a dram, it has to read the FTL from NAND, hence random 4K, and latency penalty. 

 

you usually need around 1GB of RAM, to store FTL, of 1TB data space in NAND. 

 to cache all of 4TB nand space, you would need a 4GB od your RAM. 

And nobody would allow that…
 

HBM usually have only around 32-64MB so it can cache 32-64GB on NAND space in the partial copy of FTL. 
it’s enough to fool benchmarks, but not enough for sustained workload, or reading old data ( latency penalty, you have to read FTL from NAND to read data from NAND, you read nand twice, to get to the data ).

 

do this matter, will depends on your workload, if the storage is the bottleneck, then yes, but it rarely is.

 

All my SATA, and M.2 SSD have DRAM, but that’s just me.