Intel SSD 750 PCIe SSD Review: NVMe for the Client | Hacker News Reader

Intel SSD 750 PCIe SSD Review: NVMe for the Client

anandtech.com

71 pointsfraXis29 comments

The 1.2TB model uses 22 watts for writing at 1200MB/s (54MB per second per watt). I guess performance comes at a power cost.

As a comparison, the 1TB Samsung 850 Evo uses 5.7W maximum for writing at 520MB/s (91MB per second per watt).

The earlier versions had two performance modes, one with lower power consumption. Not sure if its MB/s or iops per watt you should measure.

In fact it's generally true that for a given architecture and silicon technology, power increases as the square of speed.

It's usually voltage as the square factor for CMOS. The equation I remember from VLSI is:

  Power = (Load Capacitance) * (Voltage)^2 * (Frequency)

In order to run at a faster speed you frequently have to increase your core voltage which causes a square increase.

Incidentally 1 farad * 1 volt ^ 2 * 1/s is 1 watt. So this formula seems accurate :)

Strange given Intel's superior manufacturing capabilities?

Intel/Micron NAND isn't superior to Samsung. Intel's 14 nm is great, but NAND is a totally different thing.

Does this remind anyone of the "Hard Card", from Plus Development? They were 3.5" hard drives mounted on an ISA card way back in the day. It's funny how technology is sometimes cyclical.

Yea, the IDE/ATA interface was originally based on the ISA bus.

Ha, indeed! I had a 20 MB Hard Card in my 12 MHz 286.

Both those numbers are but a quaint memory.

Ugh, I really wish the marketing people would pick different name for this evolution of SSD drives then NVMe. At least the NVM part stating: Non-Volatile Memory.

The current iteration of what we call NVM is just SSD on PCIe busses so they are not bound by the speed of SATA/SAS controllers.

Maybe it's just me but when I think NVM I think of Non-Volatile Memory something like RAM that doesn't go blank when you hit the power button. And like RAM can be mapped into (kernel or user) address space. And access to them go through processor caches (L2 or L3) strait into NVM.

The current SSD+ drives very much behave like the block devices of the old, just with a different host interface. Most importantly they can't be directly memory mapped without going through the page cache. That still leaves a lot of performance on the table (double copy, requiring RAM for page cache versus working set).

This grossly undersells and undersimplifies the value of NVMe, its wire protocol is fundamentally simplified and it features an unlimited number of queues of unlimited depth with individual interrupt (read: CPU core) routing per queue. SATA is nothing like that. Also, the term NVM has been in use since at least the 80s to refer to probably 50 different technologies by now.

It's not a marketing trick

Most importantly they can't be directly memory mapped without going through the page cache.

It's been ages since I last did anything with PCI MMIO, but it was definitely possible to map PCI devices' own memory into a process's address space using /dev/mem, and into kernel space as well. Is there no longer an equivalent concept for PCIe? Or was MMIO always attached to real RAM?

MMIO still exists but memory-mapping NAND flash isn't a good idea. Stalling the processor for ~50 us on a cache miss is likely to lead to horrible performance.

when I think NVM I think of Non-Volatile Memory something like RAM that doesn't go blank when you hit the power button

This is the "Persistent Memory" device class. See e.g. this article for a nice summary of the status quo: https://lwn.net/Articles/636096/

Why are PCIe SSD drives so much more expensive per GB compared to the SATA ones?

1TB SSD is now around $350-360 on Amazon.

Because their performance is much higher.

I'm pretty sure SATA SSD performance is bottlenecked by the SATA bandwidth. It's the same flash inside.

Perhaps, but the PCI card also contains it's own controller instead of using the existing SATA controller in the box. Not a complete justification for the added cost, but even if the flash is the same, getting it connected is more hardware than a SATA device needs.

Ultimately it doesn't matter why the performance is higher. Also, the Intel 750 has 18 channels while most SATA SSDs have 8-10 channels.

idle power consumption: 4 Watts - way too much for my taste... I can understand it needing 22 Watts during high speed writes but why this much during staying idle ?

for comparison I checked my external 2TB WD Passport 0820 harddisk - it uses 2.11 Watts when streaming HD video...

I agree, that's some high idle power consumption. If I had to hazard a guess, it's the controller aggressively maintaining the charge of the flash cells [1]. Samsung released a fix, not too long ago, for a bug with flash cells losing charge for their 840 and 840 Evo SSD line [2], so I imagine it's a delicate process to get right. Are the other NAND controllers less aggressive and/or more efficient with this process? I don't know. I assume it's the same flash cells as the SATA drives but with a PCIe-oriented controller, so I hardly expect them to lose charge any faster than the SATA counterparts.

[1] http://www.purestorage.com/resources/introduction-to-flash-m...

[2] http://www.anandtech.com/show/8617/samsung-releases-firmware...

That was due to the TLC NAND Samsung used, correct? More cost-effective but not as durable as MLC NAND (which is what most SSDs, including this new Intel drive, use).

Incidentally, while Samsung did release a firmware update for the 840 EVO to mitigate the problem, AFAIK they never fixed the original 840. (840 Pro owners need not worry, as it used MLC NAND and was not affected.)

Wow. Pretty me too product from Intel. They even priced and sized it to not clash with Samsung, because it has nothing to make you pick it over SM951.

Not to mention you can cram SM951 into a laptop.

Want!

That's an understatement. It's unbelievable to me how quickly we have moved from <100MB/sec drives being the norm to (relatively) inexpensive 1000+MB/sec drives with incredibly fast seek times.

Now if only the laptop manufacturers would include SSDs in more mainstream models. It's depressing walking into the average computer retailer and seeing practically every shelf stuffed with slow HDs and salesmen trying to pitch the huge importance of minor differences in processor speed.

It's depressing walking into the average computer retailer and seeing practically every shelf stuffed with slow HDs and salesmen trying to pitch the huge importance of minor differences in processor speed

As far as I can tell the computer market is bifurcating or has bifurcated into people who give a shit about their experience and will pay for better (in which case they tend to buy Apple, or some select PC brands like Lenovo (until the certificate fiasco) or the Dell XPS 13"), and people who buy solely on price. I can't imagine many sophisticated buyers going through a conventional retail store, and, furthermore, if price is the primary consideration then saying "It's a big drive and a cheap computer!" is better than trying to explain why an SSD is $50 more or whatever.

I can't imagine many sophisticated buyers going through a conventional retail store

Currently, but would it not be possible to fix them? How profitable are SSDs for example?

Doesn't seem like all that long ago that 1000+MBytes/sec was more than you'd get out of memcpy.

Go back 20 years, and you couldn't even read out of L1 cache that quickly...

Reply on news.ycombinator.com