Samsung announce SM961 PM961

[Adelan1]

Coming with improved specs :

http://www.hardwareluxx.com/index.p...amsung-announces-new-sm961-pm961-m2-ssds.html

 

[ShintaiDK]

Hopefully they shrinked the controller so the heat is reduced.

 

[Insert_Nickname]

The read performance almost fully saturates a PCIe 3.0 x4 link...?

That didn't take long.

 

[Hellhammer]

The read performance almost fully saturates a PCIe 3.0 x4 link...?

That didn't take long.

PCIe only has about 80% efficiency, so 3.2GB/s is pretty much the maximum that 3.0 x4 can reach.

 

[meloz]

The read performance almost fully saturates a PCIe 3.0 x4 link...?

That didn't take long.

I think that is more to do with very high queue depths used in these type of benchmark than any realistic workloads. I can see the value of high queue depth benchmarking for random read/write performance, but for sequential performance it is blatantly used as a tool for marketing by all companies.

In real world no one will run against the 3,000 MB/s bottleneck. Specially since most servers are bottlenecked by 10 gigabit ethernet (1250 MB/s) anyway. Or am I missing something obvious?

No coverage on Anandtech at all? Site keeps getting worse. Very little information about this new Polaris controller, too. Nevertheless I found this part amusing:

The Polaris controller has five cores and can access eight memory channels.

So this controller can access more memory channels than your typical Xeon. Desktop CPUs are worse, mostly stuck with 2 channels. Perhaps we will see some sort of HMB2 type solution in future Samsung controllers?!

 

[hojnikb]

I wish manufacturers made B-key ssds aswell. Some laptops come with this (usually because they have sata m.2 ssds in them) but standard does allow for 2 lane pcie aswell, so it should be there.

Obviously speed is reduced, but still a lot.

No adapters to convert from M-key to B-key exist either, at least not to my knowledge.

 

[hojnikb]

PCIe only has about 80% efficiency, so 3.2GB/s is pretty much the maximum that 3.0 x4 can reach.

Isn't efficiency improved greatly with 3.0 (thanks to 128b/130b encoding). If so, x4 should pull just a bit under 4GB/s.

 

[Hellhammer]

I think that is more to do with very high queue depths used in these type of benchmark than any realistic workloads. I can see the value of high queue depth benchmarking for random read/write performance, but for sequential performance it is blatantly used as a tool for marketing by all companies.

Sequential performance can usually hit its maximum data rate at fairly low QD (1-4) since IOs tend to be larger and can be broken down into smaller pieces to utilize full parallelism that the architecture supports.

In real world no one will run against the 3,000 MB/s bottleneck. Specially since most servers are bottlenecked by 10 gigabit ethernet (1250 MB/s) anyway. Or am I missing something obvious?

These are client drives, not enterprise. In client usage, you are right that an SSD, especially PCIe one, is rarely going to be a bottleneck, but there are IO intensive scenarios where a faster SSD is going to increase performance.

No coverage on Anandtech at all? Site keeps getting worse.

This was not a formal announcement, it even says so in the linked article. It looks like these were simply showcased at a trade show.

So this controller can access more memory channels than your typical Xeon. Desktop CPUs are worse, mostly stuck with 2 channels. Perhaps we will see some sort of HMB2 type solution in future Samsung controllers?!

NAND channels are quite different from DRAM. A single ONFI channel is good for 400-800MB/s nowadays, whereas DRAM bandwidth is measured in gigabytes per second.

Isn't efficiency improved greatly with 3.0 (thanks to 128b/130b encoding). If so, x4 should pull just a bit under 4GB/s.

There are other overheads and inefficiencies in addition to the encoding scheme, so in reality PCIe 3.0 x4 is good for about 3.2GB/s. Ryan confirmed this with a CUDA bandwidth test, which should max out the interface. I've heard that there are some BIOS settings that can be played around with to increase PCIe bandwidth, but by default it won't go above ~800MB/s per lane.

 

[Insert_Nickname]

SThere are other overheads and inefficiencies in addition to the encoding scheme, so in reality PCIe 3.0 x4 is good for about 3.2GB/s. Ryan confirmed this with a CUDA bandwidth test, which should max out the interface. I've heard that there are some BIOS settings that can be played around with to increase PCIe bandwidth, but by default it won't go above ~800MB/s per lane.

Was referring to both drives. Anyway >3000MB/s sequential is amazing. 450K IOPS on a client drive isn't half bad either. (that's my understatement of the month)

Wonder when we'll see a retail version of the Polaris controller.

In real world no one will run against the 3,000 MB/s bottleneck. Specially since most servers are bottlenecked by 10 gigabit ethernet (1250 MB/s) anyway. Or am I missing something obvious?

Not immediately obvious, no.

But it wasn't that long ago I was complaining about Skylake's PCH only having a DMI 3.0 link (PCIe 3.0 x4, with a few extra features) to the CPU. This thing can saturate that link all by itself. With Kaby Lake's PCH adding 10Gbit USB 3.1 support and 10Gbit Ethernet finally seeming to come within reasonable pricing, how long before that link becomes a bottleneck? The Haswell platform is already limited by the DMI 2.0 PCH link for high performance SSDs (and its only 3 years old), unless you tap into the PCIe 3.0 lanes provided by the CPU. Which cuts into the lanes available for the graphics card.

Of course, you're unlikely to ever bump into that limit in practice, but I've never believed in being interface limited. This simply should not be a factor in 2016(...!).

There is also the race-to-idle factor. The quicker the SSD can get data to the CPU, the quicker it can power down.

 

[VirtualLarry]

Speaking of PCI-E SSDs, and bottlenecks:

I have a (pair) of Samsung SM951 PCI-E 3.0 x4 AHCI M.2 SSDs, that I purchased off of Newegg's ebay store.

They are plugged into a (pair) of ASRock Z170 Pro4S ATX mobos, running identical Avexir DDR4-2400 8GB kits, with G4400 CPUs. Both are overclocked, one to 4.29, one to 4.455 (130BLCK and 135BLCK).

Both are currently running Win7 64-bit SP1 + patches. Running MS AHCI drivers for the PCI-E SSD.

For some reason, I get nearly 2000/630MB/sec Seq. read/write, and around 240MB/sec 4K32 read and write.

There's a YouTube video, where they show what they claim is the same drive (same capacity too), and they get nearly double the 4K32 read and write speeds.

I've been told that my random read/write is slow, for some reason. I can't quite figure it out.

I've seen benchmarks of PNY's new CS2211 SSDs, and their 240GB and 480GB sizes put out better 4K32 read/write benchmarks using CDM, than my PCI-E SSD does, and they are SATA6G!

 

[VirtualLarry]

Hey, I just had a sudden thought. I have a SATA6G SSD plugged into the mobo, as well as a BD-RE XL drive (LG 14X).

I think that I left SATA 0 empty, because of the M.2 slot. But I think I used SATA 1. Could that be an issue?

Does the PCI-E connections to the chipset, overlap BOTH SATA 0 and 1's chipset lines? If I populated SATA 1 with a SATA6G SSD, could that be causing only half of the chipset lines to the PCI-E 3.0 x4 M.2 slot to be active, thus cutting my bandwidth in half?

I suppose it would be easy enough to test, just unplug my SATA6G SSD temporarily.

Would be curious if true.

Edit: Well, there goes my theory. I have two identical machines, only, the secondary machine doesn't have the SATA6G SSD connected to the mobo. And it benchmarks about the same as the primary machine, for the PCI-E 3.0 x4 AHCI M.2 SSD.

 

[Soulkeeper]

wow samsung is just sweaping the floor with the competition.
Still no answer from crucial/micron
marvell has had the nvme controller ready for a year+ if i'm not mistaken ...

 

[Brian Stirling]

I wonder how much the 1TB version is going to cost -- the 512GB version of the 950 Pro is going for about $330 now so I'd guess the 1TB should be in the $660 range though likely more at first.


Brian

 

[Shmee]

Will these be OEM parts? So no warranty?

 

[Adelan1]

I think SM961 is the successor to SM951 , which was an OEM drive.

I guess the consumer version 960 Pro and 960 Evo is out in Q3.

 

[Arachnotronic]

lol Polaris controller. AMD buyout confirmed!!! /s

 

[ShintaiDK]

Will these be OEM parts? So no warranty?

You still get 2 years warranty In the EU. Assuming you can find one. Other places also got consumer laws to protect the customers.

However you will have to go via the point of purchase on OEM parts. Tho that's how we usually do for any part.

 

[coercitiv]

You still get 2 years warranty In the EU. Assuming you can find one. Other places also got consumer laws to protect the customers.

SM951 comes with a 3 year warranty from my usual vendor (EU). The new model might actually improve on that, since it's V-NAND based.

 

[wpcoe]

Hope it's okay to revive a three-month old thread. I've just read elsewhere that the SM961 will be shipping soon.

Are the higher-capacity (512GB or 1TB) SM961s dual-sided, i.e. are there memory chips on both sides of the PCB?

I've read online that some motherboards, including the one I've got on order which has the M.2 slot on the underside of the motherboard, sometimes have problems with chips on the rear of the PCB making contact with components on the motherboard. In fact, I've read that the 1TB Samsung 950 Pro is held up until they have higher-density memory chips to be able to put all the chip on a single side, for that very reason.

 

[ClockHound]

I know we have to pay more for more speed. It's the storage industry's real Moore's Law. But, jeez, we have to pay more to get less hardware too? Same flash, no case = 30-50% premium. For a speed improvement that only affects 1% of consumer use cases.

Worried that when m.2 becomes popular on the desktop (which I dread, since it's purpose is in snappy lappys), is that we'll have to pay even more for a blinged-up heatsink version. ;-)

 

[trungma]

Coming with improved specs :
http://www.hardwareluxx.com/index.p...amsung-announces-new-sm961-pm961-m2-ssds.html

Crap. Just bought a SM951 NVME 256 GB for 110 bucks.

 

[VirtualLarry]

Crap. Just bought a SM951 NVME 256 GB for 110 bucks.

Don't feel too bad. I paid $100 for a 128GB SM951 AHCI model from Newegg's ebay store last Oct./Nov.-ish. Still, I think it was worth it.

I specifically got the AHCI model (thankfully) because I was installing Win7 64-bit on Skylake, which comes with it's own set of issues, and I wouldn't have been able to install a boot-time NVMe driver off of a USB stick at install time.

 

[Insert_Nickname]

I specifically got the AHCI model (thankfully) because I was installing Win7 64-bit on Skylake, which comes with it's own set of issues, and I wouldn't have been able to install a boot-time NVMe driver off of a USB stick at install time.

Just for the record, it is possible to boot 7 from an NVMe drive. There is some work required however.

https://support.microsoft.com/en-gb/kb/2990941

Don't forget to add Skylake xHCI drivers while you're at it.

 

[Edrick]

Just purchased a SM961 512GB drive for $260 shipped. That less than the 950 Pro is selling for. I will post a comparison between that drive and my 950 Pro once it arrives.

 

[Adelan1]

I picked up this screenshots from another forum :