Google releases new Cloud TPU & Machine Learning supercomputer in the cloud

Last year about this time Google released their 1st generation TPU chip to the world (see my TPU and HW vs. SW … post for more info).

This year they are releasing a new version of their hardware called the Cloud TPU chip and making it available in a cluster on their Google Cloud.  Cloud TPU is in Alpha testing now. As I understand it, access to the Cloud TPU will eventually be free to researchers who promise to freely publish their research and at a price for everyone else.

What’s different between TPU v1 and Cloud TPU v2

The differences between version 1 and 2 mostly seem to be tied to training Machine Learning Models.

TPU v1 didn’t have any real ability to train machine learning (ML) models. It was a relatively dumb (8 bit ALU) chip but if you had say a ML model already created to do something like understand speech, you could load that model into the TPU v1 board and have it be executed very fast. The TPU v1 chip board was also placed on a separate PCIe board (I think), connected to normal x86 CPUs  as sort of a CPU accelerator. The advantage of TPU v1 over GPUs or normal X86 CPUs was mostly in power consumption and speed of ML model execution.

Cloud TPU v2 looks to be a standalone multi-processor device, that’s connected to others via what looks like Ethernet connections. One thing that Google seems to be highlighting is the Cloud TPU’s floating point performance. A Cloud TPU device (board) is capable of 180 TeraFlops (trillion or 10^12 floating point operations per second). A 64 Cloud TPU device pod can theoretically execute 11.5 PetaFlops (10^15 FLops).

TPU v1 had no floating point capabilities whatsoever. So Cloud TPU is intended to speed up the training part of ML models which requires extensive floating point calculations. Presumably, they have also improved the ML model execution processing in Cloud TPU vs. TPU V1 as well. More information on their Cloud TPU chips is available here.

So how do you code a TPU?

Both TPU v1 and Cloud TPU are programmed by Google’s open source TensorFlow. TensorFlow is a set of software libraries to facilitate numerical computation via data flow graph programming.

Apparently with data flow programming you have many nodes and many more connections between them. When a connection is fired between nodes it transfers a multi-dimensional matrix (tensor) to the node. I guess the node takes this multidimensional array does some (floating point) calculations on this data and then determines which of its outgoing connections to fire and how to alter the tensor to send to across those connections.

Apparently, TensorFlow works with X86 servers, GPU chips, TPU v1 or Cloud TPU. Google TensorFlow 1.2.0 is now available. Google says that TensorFlow is in use in over 6000 open source projects. TensorFlow uses Python and 1.2.0 runs on Linux, Mac, & Windows. More information on TensorFlow can be found here.

So where can I get some Cloud TPUs

Google is releasing their new Cloud TPU in the TensorFlow Research Cloud (TFRC). The TFRC has 1000 Cloud TPU devices connected together which can be used by any organization to train machine learning algorithms and execute machine learning algorithms.

I signed up (here) to be an alpha tester. During the signup process the site asked me: what hardware (GPUs, CPUs) and platforms I was currently using to training my ML models; how long does my ML model take to train; how large a training (data) set do I use (ranging from 10GB to >1PB) as well as other ML model oriented questions. I guess there trying to understand what the market requirements are outside of Google’s own use.

Google’s been using more ML and other AI technologies in many of their products and this will no doubt accelerate with the introduction of the Cloud TPU. Making it available to others is an interesting play but this would be one way to amortize the cost of creating the chip. Another way would be to sell the Cloud TPU directly to businesses, government agencies, non government agencies, etc.

I have no real idea what I am going to do with alpha access to the TFRC but I was thinking maybe I could feed it all my blog posts and train a ML model to start writing blog post for me. If anyone has any other ideas, please let me know.


Photo credit(s): From Google’s website on the new Cloud TPU


Disaster recovery from VMware to AWS using Dell EMC Avamar & Data Domain

avI was at Dell EMC World2017 last week and although most of the news was on Dell’s new 14th generation server and Dell-EMC integration progress, Wednesday’s keynote was devoted to storage and non-server infrastructure news.

There was plenty of non-server news but one item that caught my attention was new functionality from Dell EMC Data Protection Division that used Avamar and Data Domain to provide disaster recovery for VMware VMs directly to AWS.

Data Domain (AWS) Cloud DR

Dell EMC Data Domain Cloud DR (DDCDR) is  a new capability that enables DD to backup to AWS S3 object storage and when needed restart the virtual machines within AWS.

DDCDR requires that a customer with Avamar backup and Data Domain (DD) storage install an OVA which deploys an “add-on” to their on-prem Avamar/DD system and install a lightweight VM (Cloud DR server) utility in their AWS domain.

Once the OVA is installed, it will read the changed data and will segment, encrypt, and compress the backup data and then send this and the backup metadata to AWS S3 objects. Avamar/DD policies can be established to control how many daily backup copies are to be saved to S3 object storage. There’s no need for Data Domain or Avamar to run in AWS.

When there’s a problem at the primary data center, an admin can click on a Avamar GUI button and have the Cloud DR server, uncompress, decrypt, rehydrate and restore the backup data into EBS volumes, translate the VMware VM image to an AMI image and then restarts the AMI on an AWS virtual server (EC2) with its data on EBS volume storage. The Cloud DR server will use the backup metadata to select the AWS EC2 instance with the proper CPU and RAM needed to run the application. Once this completes, the VM is running standalone, in an AWS EC2 instance. Presumably, you have to have EC2 and EBS storage volumes resources available under your AWS domain to be able to install the application and restore its data.

For simplicity purposes, the user can control almost all of the required functionality for DDCDR from the Avamar GUI alone. But in case of a site outage, the user can initiate the application DR from a portal supplied by the Cloud DR server utility.

There you have it, simplified, easy to use (AWS) Cloud DR for your VM applications all through Dell EMC Avamar, Data Domain storage and DDCDR. At the moment, it only works with AWS cloud but it’s likely to be available for other public clouds in the near future.


There was much more infrastructure news at Dell EMC World2017. I’ll discuss more details on their new storage offerings in my upcoming Storage Intelligence newsletter, due out the end of this month. If your interested in receiving your own copy of my newsletter, checkout the signup button in the upper right of this page.


[Edits were made for readability and technical accuracy after this post was published. Ed]

Crowdsourced vision for visually impaired

Read an article the other day in Christian Science Monitor (CSM) on the Be My Eyes App. The app is from and is available for the iPhone and Android smart phones.

Essentially there are two groups of people that use the app:

  • Visually helpful volunteers – these people signup for the app and when a visually impaired person needs help they provide visual aid by speaking to the person on the other end.
  • Visually impaired individuals – these people signup for the app and when they are having problems understanding what they are (or are not) looking at they can turn on their camera take video with their phone and it will be sent to a volunteer, they can then ask the volunteer for help in deciding what they are looking at.

So, the visually impaired ask questions about the scenes they are shooting with their phone camera and volunteers will provide an answer.

It’s easy to register as Sighted and I assume Blind. I downloaded the app, registered and tried a test call in minutes. You have to enable notifications, microphone access and camera access on your phone to use the app. The camera access is required to display the scene/video on your phone.

According to the app there are 492K sighted individuals, 34.1K blind individuals and they have been helped 214K times.

Sounds like an easy way to help the world.

There was no requests to identify a language to use, so it may only work for English speakers. And there was no way to disable/enable it for a period of time when you don’t want to be disturbed. But maybe you would just close the app.

But other than that it was simple to use and seemed effective.

Now if there were only an app that would provide the same service for the hearing impaired to supply captions or a “filtered” audio feed to ear buds.

The world need more apps like this…


AI’s Image recognition success feeds sound recognition improvements

I must do reCAPTCHA at least a dozen times a week for various websites I use. It’s become a real pain. And the fact that I know that what I am doing is helping some AI image recognition program do a better job of identifying street signs, mountains, or shop fronts doesn’t reduce my angst.

But that’s the thing with deep learning, machine learning, re-inforcement learning, etc. they all need massive amounts of annotated data that’s a correct interpretation of a scene in order to train properly.

Computers to the rescue

So, when I read a recent article in MIT News that Computers learn to recognize sounds by watching video, I was intrigued. What the researchers at MIT have done is use advanced image recognition to annotate film clips with the names of things that are making sounds on the film. They then fed this automatically annotated data into a sound identifying algorithm to improve its recognition capability.

They used this approach to train their sound recognition system to be  able to identify natural and artificial sounds like bird song, speaking in crowds, traffic sounds, etc.

They tested their newly automatically trained sound recognition against standard labeled sound sets and was able to categorize sound with a 92% accuracy for a 10 category data set and with a 74% accuracy with a 50 category dataset. Humans are able categorize these sounds with a 96% and 81% accuracy, respectively.

AI’s need for annotation

The problem with machine learning is that it needs a massive, properly annotated data set in order to learn properly. But getting annotated data takes too long or is too expensive to do for many things that we want AI for.

Using one AI tool to annotate data to train another AI tool is sort of bootstrapping AI technology. It’s acute trick but may have only limited application. I could only think of only a few more applications of similar technology:

  • Use chest strap or EKG technology to annotate audio clips of heart beat sounds at a wrist or other appendage to train a system to accurately determine pulse rates through sound alone.
  • Use wave monitoring technology to annotate pictures and audio clips of sea waves to train a system to accurately determine wave levels for better tsunami detection.
  • Use image recognition to annotate pictures of food and then use this train a system to recognize food smells (if they ever find a way to record smells).

But there may be many others. Just further refinement of what they have used could lead to finer grained people detection. For example, as (facial) image recognition gets better, it’s possible to annotate speaking film clips to train a sound recognition system to identify people from just hearing their speech. Intelligence applications for such technology are significant.

Nonetheless, I for one am happy that the next reCAPTCHA won’t be having me identify river sounds in a matrix of 9 sound clips.

But I fear there’s enough GreyBeards on Storage podcast recordings and Storage Field Day video clips already available to train a system to identify Ray’s and for sure, Howard’s voice anywhere on the planet…


Photo Credit(s): Wave by Matthew Potter; Waves crashing on Puget Sound by mikeskatieDay 16: Podcasting by Laura Blankenship

The fragility of public cloud IT

I have been reading AntiFragile again (by Nassim Taleb). And although he would probably disagree with my use of his concepts, it appears to me that IT is becoming more fragile, not less.

For example, recent outages at major public cloud providers display increased fragility for IT. Yet these problems, although almost national in scope, seldom deter individual organizations from their migration to the cloud.

Tragedy of the cloud commons

The issues are somewhat similar to the tragedy of the commons. When more and more entities use a common pool of resources, occasionally that common pool can become degraded. But because no-one really owns the common resources no one has any incentive to improve the situation.

Now the public cloud, although certainly a common pool of resources, is also most assuredly owned by corporations. So it’s not a true tragedy of the commons problem. Public cloud corporations have a real incentive to improve their services.

However, the fragility of IT in general, the web, and other electronic/data services all increases as they become more and more reliant on public cloud, common infrastructure. And I would propose this general IT fragility is really not owned by any one person, corporation or organization, let alone the public cloud providers.

Pre-cloud was less fragile, post-cloud more so

In the old days of last century, pre-cloud, if a human screwed up a CLI command the worst they could happen was to take out a corporation’s data services. Nowadays, post-cloud, if a similar human screws up a CLI command, the worst that can happen is that major portions of the internet services of a nation go down.

Strange Clouds by michaelroper (cc) (from Flickr)

Yes, over time, public cloud services have become better at not causing outages, but they aren’t going away. And if anything, better public cloud services just encourages more corporations to use them for more data services, causing any subsequent cloud outage to be more impactful, not less

The Internet was originally designed by DARPA to be more resilient to failures, outages and nuclear attack. But by centralizing IT infrastructure onto public cloud common infrastructure, we are reversing the web’s inherent fault tolerance and causing IT to be more susceptible to failures.

What can be done?

There are certainly things that can be done to improve the situation and make IT less fragile in the short and long run:

  1. Use the cloud for non-essential or temporary data services, that don’t hurt a corporation, organization or nation when outages occur.
  2. Build in fault-tolerance, automatic switchover for public cloud data services to other regions/clouds.
  3. Physically partition public cloud infrastructure into more regions and physically separate infrastructure segments within regions, such that any one admin has limited control over an amount of public cloud infrastructure.
  4. Divide an organizations or nations data services across public cloud infrastructures, across as many regions and segments as possible.
  5. Create a National Public IT Safety Board, not unlike the one for transportation, that does a formal post-mortem of every public cloud outage, proposes fixes, and enforces fix compliance.

The National Public IT Safety Board

The National Transportation Safety Board (NTSB) has worked well for air transportation. It relies on the cooperation of multiple equipment vendors, airlines, countries and other parties. It performs formal post mortems on any air transportation failure. It also enforces any fixes in processes, procedures, training and any other activities on equipment vendors, maintenance services, pilots, airlines and other entities that can impact public air transport safety. At the moment, air transport is probably the safest form of transportation available, and much of this is due to the NTSB

We need something similar for public (cloud) IT services. Yes most public cloud companies are doing this sort of work themselves in isolation, but we have a pressing need to accelerate this process across cloud vendors to improve public IT reliability even faster.

The public cloud is here to stay and if anything will become more encompassing, running more and more of the worlds IT. And as IoT, AI and automation becomes more pervasive, data processes that support these services, which will, no doubt run in the cloud, can impact public safety. Just think of what would happen in the future if an outage occurred in a major cloud provider running the backend for self-guided car algorithms during rush hour.

If the public cloud is to remain (at this point almost inevitable) then the safety and continuous functioning of this infrastructure becomes a public concern. As such, having a National Public IT Safety Board seems like the only way to have some entity own IT’s increased fragility due to  public cloud infrastructure consolidation.


In the meantime, as corporations, government and other entities contemplate migrating data services to the cloud, they should consider the broader impact they are having on the reliability of public IT. When public cloud outages occur, all organizations suffer from the reduced public perception of IT service reliability.

Photo Credits: Fragile by Bart Everson; Fragile Planet by Dave Ginsberg; Strange Clouds by Michael Roper

Hardware vs. software innovation – round 4

We, the industry and I, have had a long running debate on whether hardware innovation still makes sense anymore (see my Hardware vs. software innovation – rounds 1, 2, & 3 posts).

The news within the last week or so is that Dell-EMC cancelled their multi-million$, DSSD project, which was a new hardware innovation intensive, Tier 0 flash storage solution, offering 10 million of IO/sec at 100µsec response times to a rack of servers.

DSSD required specialized hardware and software in the client or host server, specialized cabling between the client and the DSSD storage device and specialized hardware and flash storage in the storage device.

What ultimately did DSSD in, was the emergence of NVMe protocols, NVMe SSDs and RoCE (RDMA over Converged Ethernet) NICs.

Last weeks post on Excelero (see my 4.5M IO/sec@227µsec … post) was just one example of what can be done with such “commodity” hardware. We just finished a GreyBeardsOnStorage podcast (GreyBeards podcast with Zivan Ori, CEO & Co-founder, E8 storage) with E8 Storage which is yet another approach to using NVMe-RoCE “commodity” hardware and providing amazing performance.

Both Excelero and E8 Storage offer over 4 million IO/sec with ~120 to ~230µsec response times to multiple racks of servers. All this with off the shelf, commodity hardware and lots of software magic.

Lessons for future hardware innovation

What can be learned from the DSSD to NVMe(SSDs & protocol)-RoCE technological transition for future hardware innovation:

  1. Closely track all commodity hardware innovations, especially ones that offer similar functionality and/or performance to what you are doing with your hardware.
  2. Intensely focus any specialized hardware innovation to a small subset of functionality that gives you the most bang, most benefits at minimum cost and avoid unnecessary changes to other hardware.
  3. Speedup hardware design-validation-prototype-production cycle as much as possible to get your solution to the market faster and try to outrun and get ahead of commodity hardware innovation for as long as possible.
  4. When (and not if) commodity hardware innovation emerges that provides  similar functionality/performance, abandon your hardware approach as quick as possible and adopt commodity hardware.

Of all the above, I believe the main problem is hardware innovation cycle times. Yes, hardware innovation costs too much (not discussed above) but I believe that these costs are a concern only if the product doesn’t succeed in the market.

When a storage (or any systems) company can startup and in 18-24 months produce a competitive product with only software development and aggressive hardware sourcing/validation/testing, having specialized hardware innovation that takes 18 months to start and another 1-2 years to get to GA ready is way too long.

What’s the solution?

I think FPGA’s have to be a part of any solution to making hardware innovation faster. With FPGA’s hardware innovation can occur in days weeks rather than months to years. Yes ASICs cost much less but cycle time is THE problem from my perspective.

I’d like to think that ASIC development cycle times of design, validation, prototype and production could also be reduced. But I don’t see how. Maybe AI can help to reduce time for design-validation. But independent FABs can only speed the prototype and production phases for new ASICs, so much.

ASIC failures also happen on a regular basis. There’s got to be a way to more quickly fix ASIC and other hardware errors. Yes some hardware fixes can be done in software but occasionally the fix requires hardware changes. A quicker hardware fix approach should help.

Finally, there must be an expectation that commodity hardware will catch up eventually, especially if the market is large enough. So an eventual changeover to commodity hardware should be baked in, from the start.


In the end, project failures like this happen. Hardware innovation needs to learn from them and move on. I commend Dell-EMC for making the hard decision to kill the project.

There will be a next time for specialized hardware innovation and it will be better. There are just too many problems that remain in the storage (and systems) industry and a select few of these can only be solved with specialized hardware.


Picture credit(s): Gravestones by Sherry NelsonMotherboard 1 by Gareth Palidwor; Copy of a DSSD slide photo taken from EMC presentation by Author (c) Dell-EMC

4.5M IO/sec@227µsec 4KB Read on 100GBE with 24 NVMe cards #SFD12

At Storage Field Day 12 (SFD12) this week we talked with Excelero, which is a startup out of Israel. They support a software defined block storage for Linux.

Excelero depends on NVMe SSDs in servers (hyper converged or as a storage system), 100GBE and RDMA NICs. (At the time I wrote this post, videos from the presentation were not available, but the TFD team assures me they will be up on their website soon).

I know, yet another software defined storage startup.

Well yesterday they demoed a single storage system that generated 2.5 M IO/sec random 4KB random writes or 4.5 M IO/Sec random 4KB reads. I didn’t record the random write average response time but it was less than 350µsec and the random read average response time was 227µsec. They only did these 30 second test runs a couple of times, but the IO performance was staggering.

But they used lots of hardware, right?

No. The target storage system used during their demo consisted of:

  • 1-Supermicro 2028U-TN24RT+, a 2U dual socket server with up to 24 NVMe 2.5″ drive slots;
  • 2-2 x 100Gbs Mellanox ConnectX-5 100Gbs Ethernet (R[DMA]-NICs); and
  • 24-Intel 2.5″ 400GB NVMe SSDs.

They also had a Dell Z9100-ON Switch  supporting 32 X 100Gbs QSFP28 ports and I think they were using 4 hosts but all this was not part of the storage target system.

I don’t recall the CPU processor used on the target but it was a relatively lowend, cheap ($300 or so) dual core, Intel standard CPU. I think they said the total target hardware cost $13K or so.

I priced out an equivalent system. 24 400GB 2.5″ NVMe Intel 750 SSDs would cost around $7.8K (Newegg); the 2 Mellanox ConnectX-5 cards $4K (Neutron USA); and the SuperMicro plus an Intel Cpu around $1.5K. So the total system is close to the ~$13K.

But it burned out the target CPU, didn’t it?

During the 4.5M IO/sec random read benchmark, the storage target CPU was at 0.3% busy and the highest consuming process on the target CPU was the Linux “Top” command used to display the PS status.

Excelero claims that the storage target system consumes absolutely no CPU processing to service an 4K read or write IO request. All of IO processing is done by hardware (the R(DMA)-NICs, the NVMe drives and PCIe bus) which bypasses the storage target CPU altogether.

We didn’t look at the host cpu utilization but driving 4.5M IO/sec would take a high level of CPU power even if their client software did most of this via RDMA messaging magic.

How is this possible?

Their client software running in the Linux host is roughly equivalent to an iSCSI initiator but talks a special RDMA protocol (patent pending by Excelero, RDDA protocol) that adds an IO request to the NVMe device submission queue and then rings the doorbell on the target system device and the SSD then takes it off the queue and executes it. In addition to the submission queue IO request they preprogram the PCIe MSI interrupt request message to somehow program (?) the target system R-NIC to send the read data/write status data back to the client host.

So there’s really no target CPU processing for any NVMe message handling or interrupt processing, it’s all done by the client SW and is handled between the NVMe drive and the target and client R-NICs.

The result is that the data is sent back to the requesting host automatically from the drive to the target R-NIC over the target’s PCIe bus and then from the target system to the client system via RDMA across 100GBE and the R-NICS and then from the client R-NIC to the client IO memory data buffer over the client’s PCIe bus.

Writes are a bit simpler as the 4KB write data can be encapsulated into the submission queue command for the write operation that’s sent to the NVMe device and the write IO status is relatively small amount of data that needs to be sent back to the client.

NVMe optimized for 4KB IO

Of course the NVMe protocol is set up to transfer up to 4KB of data with a (write command) submission queue element. And the PCIe MSI interrupt return message can be programmed to (I think) write a command in the R-NIC to cause the data transfer back for a read command directly into the client’s memory using RDMA with no CPU activity whatsoever in either operation. As long as your IO request is less than 4KB, this all works fine.

There is some minor CPU processing on the target to configure a LUN and set up the client to target connection. They essentially only support replicated RAID 10 protection across the NVMe SSDs.

They also showed another demo which used the same drive both across the 100Gbs Ethernet network and in local mode or direct as a local NVMe storage. The response times shown for both local and remote were within  5µsec of each other. This means that the overhead for going over the Ethernet link rather than going local cost you an additional 5µsec of response time.

Disaggregated vs. aggregated configuration

In addition to their standalone (disaggregated) storage target solution they also showed an (aggregated) Linux based, hyper converged client-target configuration with a smaller number of NVMe drives in them. This could be used in configurations where VMs operated and both client and target Excelero software was running on the same hardware.

Simply amazing

The product has no advanced data services. no high availability, snapshots, erasure coding, dedupe, compression replication, thin provisioning, etc. advanced data services are all lacking. But if I can clone a LUN at lets say 2.5M IO/sec I can get by with no snapshotting. And with hardware that’s this cheap I’m not sure I care about thin provisioning, dedupe and compression.  Remote site replication is never going to happen at these speeds. Ok HA is an important consideration but I think they can make that happen and they do support RAID 10 (data mirroring) so data mirroring is there for an NVMe device failure.

But if you want 4.5M 4K random reads or 2.5M 4K random writes on <$15K of hardware and happen to be running Linux, I think they have a solution for you. They showed some volume provisioning software but I was too overwhelmed trying to make sense of their performance to notice.

Yes it really screams for 4KB IO. But that covers a lot of IO activity these days. And if you can do Millions of them a second splitting up bigger IOs into 4K should not be a problem.

As far as I could tell they are selling Excelero software as a standalone product and offering it to OEMs. They already have a few customers using Excelero’s standalone software and will be announcing  OEMs soon.

I really want one for my Mac office environment, although what I’d do with a millions of IO/sec is another question.


Intel’s Optane (3D Xpoint) SSD specs in the wild

Read an article the other day in Ars Technica (Specs for 1st Intel 3DX SSD…) about a preview of the Intel Octane specs for their 375GB 3D Xpoint (3DX) flash card. The device is NVMe compliant, PCIe Gen3 add in card, that’s in a half height, half length, low profile form factor.

Intel’s Optane SSD vs. the competition

A couple of items from the Intel Optane spec sheet of interest to me as a storage guru:

  • 30 Drive writes per day/12.3 PBW (written) – 3DX, at launch, had advertised that it would have 1000 times the endurance of (2D-MLC?) NAND. Current flash cards (see Samsung SSD PRO NVMe 256GB Flash card specs) offer about 200TBW (for 256GB card) or 400TBW (for 512GB card). The Samsung PRO is based on 3D (V-)NAND, so its endurance is much better than  2D-MLC at these densities. That being said, the Octane drive is still ~40X the write endurance of the PRO 950. Not quite 1000 but certainly significantly better.
  • Sequential (bandwidth) performance (R/W) of 2400/2000 MB/sec – 3DX advertised 1000 times the performance of (2D-MLC,  non-NVMe?) NAND. Current 3D (V-)NAND cards (see Samsung SSD PRO above) above offers (R/W) 2200/900 MB/sec for an NVMe device. The Optane’s read bandwidth is a slight improvement but the write bandwidth is a 2.2X improvement over current competitive devices.
  • Random 4KB IOPs performance (R/W) of 550K/500K – Similar to the previous bulleted item, 3DX advertised 1000 times the performance of (2D-MLC,  non-NVMe?) NAND. Current 3D (V-)NAND cards like the Samsung SSD PRO offer Random 4KB IOPs performance  (R/W) of 270K/85K IOPS (@4 threads). Optane’s read random 4KB IOPs performance is 2X the PRO 950 but its write performance is ~5.9X better.
  • IO latency of <10 µsec. – 3DX advertised 10X better latency than the current (2D-MLC, non-NVMe) flash drives. According to storage review (Samsung 950 Pro M.2), the Samsung PRO 950 had a latency of ~22 µsec. Optane has at least 2X better latency than the current competition.
  • Density 375GB/HH-HL-LP – 3DX advertised 1000X the density of (then current DRAM). Today Micron offers a 4GiB DDR4/288 pin DIMM which is probably 1/2 the size of the HH flash drive. So maybe in the same space this could be 8GiB. This says that the Optane is about 100X denser than today’s DRAM.

Please note, when 3DX was launched, ~2 years ago, the then current NAND technology was 2D-MLC and NVMe was just a dream. So comparing launch claims against today’s current 3D-NAND, NVMe drives is not a fair comparison.

Nevertheless, the Optane SSD performs considerably better than current competitive NVMe drives and has significantly better endurance than current 3D (V-)NAND flash drives. All of which is a great step in the right direction.

What about DRAM replacement?

At launch, 3DX was also touted as a higher density, potential replacement for DRAM. But so far we haven’t seen any specs for what 3DX NVM looks like on a memory bus. It has much better density than DRAM, but we would need to see 3DX memory access times under 50ns to have a future as a DRAM replacement. Optane’s NVMe SSD at 10 µsec. is about 200X too slow, but then again it’s not a memory device configuration nor is it attached to a memory bus.


Photo Credit(s):  Intel Optane Spec sheet from Ars Technica Article,  DDR4 DRAM from Wikimedia user:Dsimic