Optane 101: Memory or storage? Yes.

By now, you’ve seen the word “Optane” bandied about on VentureBeat (such as here and here) and probably countless other places — and for good reason. Intel, the maker of all Optane products, is heavily promoting the results of its decade-long R&D investment in this new memory/storage hybrid. But what exactly is Optane, and what is it good for? (Hint: Massively memory-hungry applications like analytics and AI). If you’re not feeling up to speed, don’t worry. We’ll have you covered on all the basics in the next few minutes.

The bottom line

Optane is a new Intel technology that blurs the traditional lines between DRAM memory and NAND flash storage.

Optane DC solid state drives provide super-fast data caching and agile system expansion.

Capacity up to 512GB per persistent memory module; configurable for persistent or volatile operation; ideal for applications that emphasize high capacity and low latency over raw throughput

Strong contender for data centers; future for clients. Costs and advantages are case-specific, impacted by DRAM prices. Early user experience still emerging.

Now, let’s dive into some more detail.

Media vs. memory vs. storage

First, understand that Intel Optane is neither DRAM nor NAND flash memory. It’s a new set of technologies based on what Intel calls 3D XPoint media, which was co-developed with Micron. (We’re going to stumble around here with words like media, memory, and storage, but will prefer “media.”) 3D XPoint works like NAND in that it’s non-volatile, meaning data doesn’t disappear if the system or components lose power.

However, 3D XPoint has significantly lower latency than NAND. That lets it perform much more like DRAM in some situations, especially with high volumes of small files, such as online transaction processing (OLTP). In addition, 3D XPoint features orders-of-magnitude higher endurance than NAND, which makes it very attractive in data center applications involving massive amounts of data writing.

When combined with Intel firmware and drivers, 3D XPoint gets branded as simply “Optane.”

So, is Optane memory or storage? The answer depends on where you put it in a system and how it gets configured.

Optane memory

Consider Intel Optane Memory, the first product delivered to market with 3D XPoint media. Available in 16GB or 32GB models, Optane memory products are essentially tiny PCIe NVMe SSDs built on the M.2 form factor. They serve as a fast cache for storage. Frequently loaded files get stashed on Optane memory, alleviating the need to find those files on NAND SSDs or hard drives, which will entail much higher latency. Optane memory is targeted at PCs, but therein lies the rub. Most PCs don’t pull that much file traffic from storage and don’t need that sort of caching performance, And because, unlike NAND, 3D XPoint doesn’t require an erase cycle when writing to media, Optane is strong on write performance. Still, most client applications don’t have that much high-volume, small-size writing to do.

Optane SSDs: Client and data center 

Next came Intel Optane SSDs and Data Center (DC) SSDs. Today, the Intel Optane SSD 8 Series ships in 58GB to 118GB capacities, also using the M.2 form factor. The 9 Series reaches from 480GB to 1.5TB but employs the M.2, U.2, and Add In Card (AIC) form factors. Again, Intel bills these as client SSDs, and they certainly have good roles to play under certain conditions. But NAND SSDs remain the go-to for clients across most desktop-class applications, especially when price and throughput performance (as opposed to latency) are being balanced.

Things change once we step into the data center. The SKUs don’t look that different from their client counterparts — capacities from 100GB to 1.5TB across U.2, M.2, and half-height, half-length (HHHL) AIC form factors — except in two regards: price and endurance. Yes, the Intel Optane SSD DC P4800X (750GB) costs roughly double the Intel Optane SSD 905P (960GB). But look at its endurance advantage: 41 petabytes written (PBW) versus 17.52 PBW. In other words, on average, you can exhaust more than two consumer Optane storage drives — and pay for IT to replace them — in the time it takes to wear out one DC Optane drive.

Optane DC Persistent Memory

Lastly, Intel Optane DC Persistent Memory modules (DCPMM) place 3D XPoint media on DDR4 form factor memory sticks. (Note: There’s no DDR4 media on the module, but DCPMMs do insert into the DDR4 DIMM sockets on compatible server motherboards.) Again, Optane media is slower than most DDR4, but not much slower in many cases. Why use it, then? Because Optane DCPMMs come in capacities up to 512GB – much higher than DDR4 modules, which top out at 64GB each. Thus, if you have applications and workloads that prioritize capacity over speed, a common situation for in-memory databases and servers with high virtual machine density, Optane DCPMMs may be a strong fit.

The value proposition for DCPMM was stronger in early 2018 and early 2019, when DRAM prices were higher. This allowed DPCMMs to win resoundingly on capacity and per-gigabyte price. As DRAM prices have plummeted, though, the two are at near-parity, which is why you now hear Intel talking more about the capacity benefits in application-specific settings. As Optane gradually proves itself in enterprises, expect to see Intel lower DCPMM prices to push the technology into the mainstream.

As for total performance, DCPMM use case stories and trials are just emerging from the earliest enterprise adopters. Intel has yet to publish clear data showing “Optane DCPMMs show X% performance advantage over DRAM-only configurations in Y and Z environments.” This is partially because server configurations, which often employ virtualization and cross-system load sharing, can be very tricky to typify. But it’s also because the technology is so new that it hasn’t been widely tested. For now, the theory is that large DCPMM pools, while slower than DRAM-only pools, will reduce the need for disk I/O swaps. That will accelerate total performance above and beyond the levels reduced by adopting a somewhat slower media.

Early Intel Tiger Lake Benchmarks Point to Beefy Cache, Gen12 Graphics

Several Tiger Lake Y (TGL-Y) and Tiger Lake U (TGL-U) entries have started to appeared in the Geekbench 4 and SiSoftware databases. Tiger Lake is slated to debut next year to succeed Ice Lake (ICL). 

Starting with the Geekbench 4 submission, the Tiger Lake Y sample checked in with a quad-core, eight-thread configuration and a 1.19 GHz base clock. This chip appears to be identical to the one that was spotted back in July. This time, however, we have more information on the cache configuration.

According to Geekbench 4’s listing, the quad-core Tiger Lake Y chip in question has 1.25MB of L2 cache per core and up to 12MB of L3 cache, which should equate to 3MB per core. In comparison, Ice Lake features 512KB of L1I cache and 2MB of L3 cache per core. However, Tiger Lake Y’s L1D and L1 cache design seems to remain unchanged. The sample shows 48KB of L1D cache and 32KB of L1I cache per core, the same as Ice Lake.

It’s already known that Tiger Lake processors will employ Intel’s Gen12 (Generation 12) graphics processing unit. The latest SiSoftware submissions give us a sneak peek of what we may be able to expect from Gen12 in terms of specifications.

The submissions both point to Gen12 sporting up to 96 Execution Units (EUs), which adds up to 768 shading units. As you may recall, Gen11 maxes out at 64 EUs so, we’re looking at a 50% increase.

The iGPU (integrated GPU) inside Tiger Lake Y appears to operate at 1.1 GHz, with the Tiger Lake U’s iGPU working at 1 GHz. This isn’t surprising, as Intel’s Y-series is confined to a lower TDP (thermal design power) than the U-series and doesn’t have as much room to breathe.

In other news, Linux publication Phoronix today uncovered Intel’s latest update to add 16-bit atomic operation support for its Tiger Lake processors. It’s safe to assume that Intel is working diligently on Tiger Lake for a 2020 release.

Intel Comet Lake-S 10-Core CPU Benchmarks Surface

Fresh benchmarks for Intel’s 10th-Generation Comet Lake-S (CML-S) processors have started to appeared in the Geekbench 4 database. The results reveal the chips’ core counts, L2 and L3 cache, as well as the preliminary base and boost clocks. But as always, it’s wise to take these with a some grains of salt, since they are submissions of unreleased hardware. 

As a quick refresher, Comet Lake-S will replace Intel’s current Coffee Lake Refresh lineup. The upcoming Comet Lake-S chips continue to hail from Intel’s 14nm process node. However, they’re likely to be based on an improved 14nm+++ process. Comet Lake-S is also expected to bring higher core counts and more cache.

Starting with what we assume is the flagship part, the unidentified Comet Lake-S processor will apparently come rocking 10 cores and 20 threads. It seemingly has 640KB of L1 cache, 2.5MB of L2 cache and 20MB of L3 cache. Geekbench 4 lists the processor with a 1.51 GHz base clock and 3.19 GHz boost clock.

The other unknown Comet Lake-S part is said to sport six cores and 12 threads, plus 384KB of L1 cache, 1.5MB of L2 cache and 12MB of L3 cache. Geekbench 4 identified the chip as having a 1.99 GHz base clock and 2.89 GHz boost clock.

According to Geekbench 4’s report, both Comet Lake-S processors are equipped with Intel’s UHD Graphics 630, which is the same iGPU (integrated graphics processing unit) that debuted with the chipmaker’s Coffee Lake family. It’s perplexing that Geekbench 4 only detects 23 EUs (execution units) for both Comet Lake-S parts when the UHD Graphics 630 is known to offer up to 24 EUs. When it comes to speeds, the 10-core chip’s iGPU is clocked at 1.2 GHz, while the six-core is confined to 1.15 GHz.

AMD has already landed the first blow with its third-generation Ryzen desktop processors that maxes out at 16 cores with the AMD Ryzen 9 3950X. Obviously, Intel would be at a disadvantage core-wise if Comet Lake-S arrives topping out at just 10 cores, as rumored. So, It’ll be interesting to see how Intel ends up responding. 

ASUS Brings Wi-Fi 6 to Desktops with PCE-AX3000 Card

ASUS has introduced one of the industry’s first Wi-Fi 6 cards for desktops. The ASUS PCE-AX3000 dual band PCIe 3.0 x1 adapter promises to provide up to 2.4 Gbps data transfer rates over 160 MHz channels when used with an appropriate router.

The ASUS PCE-AX3000 card is essentially an adapter carrying an M.2-2230 Wi-Fi 6 card for laptops. The actual radio that powers the card is unknown, however it supports all key features of the Wi-Fi 6 specification, including 2.4 GHz and 5 GHz bands, 80 MHz and 160 MHz channels, MU-MIMO, OFDMA (Orthogonal Frequency-Division Multiple Access) for allowing different devices to be served by one channel, and WPA3. Besides Wi-Fi 6, the card also supports Bluetooth 5.0 technology allowing desktops to connect to various mobile and peripheral devices.

The PCE-AX3000 Dual Band PCI-E Wi-Fi 6 has two large antennas to ensure strong signal and fast connectivity in various situations. The adapter should be compatible with all existing desktops that use Windows 10 (64-bit) or Linux.

Numerous modern laptops feature Wi-Fi 6, providing higher performance in crowded networks, but not everything is that rosy with desktops as Wi-Fi 6 adapters for desktops are practically nonexistent. Luckily, with products like the ASUS PCE-AX3000, 2.4 Gbps connectivity will now be available for desktop computers too.

Centaur Creates First x86 SoC with Integrated AI Co-Processor

People typically think of x86 processors as coming from Intel and AMD, but there is a third architectural license holder: VIA. This week, an Austin, Texas-based subsidiary of the Taipei-headquartered company announced it’s demonstrating an x86 processor that comes with an integrated artificial intelligence (AI) co-processor.

VIA’s Centaur Technology is a small CPU design company. The unnamed processor it’s developing is built on the 16nm fabrication process and manufactured by TSMC. It’s a complete system-on-chip (SoC) with eight cores, 16MB of L3cache and am AI co-processor. In total, it has a die size of 195 square millimeters, which isn’t all that big. For comparison, a Ryzen 5 chip with one CCD and one IOD has a die measuring 199 square millimeters.

Centaur’s new chip isn’t meant to land in consumer PCs. Rather, the end goal is to land in enterprise systems aimed at deep learning and other industrial applications. 

Centaur is developing this processor to tackle the challenge of x86 processors needing external inference acceleration (such as a GPU with Nvidia’s Tensor cores). It wants to integrate this feature into one chip and, consequently, reduce power consumption for deep learning tasks.

Next to its eight x86 cores, 16MB of L3 cache and 20 TOPS AI co-processor, Centaur’s chip comes with a total of 44 PCIe lanes and four DDR4 memory channels. Therefore, if a user wanted to further improve a supporting system’s inference performance they’d also be able to add GPUs into the mix. 

Currently, the reference platform runs at 2.5 GHz. It also comes with the AVX-512 instruction set, which, thus far, has only been implemented in very select few processors.

The 16MB of memory that the AI co-processors also have access to enable them to communicate at up to 20 TBps, which has led to the lowest latency for image classification within just 330 microseconds in the MLPerf benchmark.

“We set out to design an AI co-processor with 50 times the inference performance of a general-purpose CPU. We achieved that goal. Now we are working to enhance the hardware for both high-performance and low-power systems, and we are disclosing some of our technology details to encourage feedback from potential customers and technology partners,” Glenn Hendry, Centaur’s Chief Architect of the AI co-processor, said in a statement.

Centaur’s x86 CPU with its AI co-processor isn’t ready for prime time yet but will be demonstrated at ISC East on November 20 and 21, with technical details to be published on December 2. 

AMD might unveil a new GPU with ray tracing support at CES

It’s only a matter of time before AMD rolls out a GPU architecture with hardware-based real-time ray tracing support (though not in time for any Black Friday deals), to take on Nvidia’s growing army of GeForce RTX graphics cards. But might we see something in that regard at CES in January, 2020? According to the latest rumor mill chatter, the answer is yes.

Take this with a dose of skepticism, but according to Chiphell forum member Wjm47196, AMD will be previewing a second-generation Radeon DNA (RDNA) Navi GPU architecture at the annual convention in Las Vegas, as spotted by WCCFTech.

I can see your eyes rolling in the back of the head, but there are couple of things that make this rumor at least plausible. One is Wjm47196’s track record. The same user had posted accurate information about AMD’s first generation Navi architecture before it launched. Same goes for some other GPU launches.

Secondly, we know AMD will implement ray tracing support into a future Navi GPU architecture, because both Sony’s PlayStation 5 and Microsoft’s next Xbox (Project Scarlett) will feature one, based on custom silicon.

A CES announcement also doesn’t mean the GPU would launch soon after. The past several years, AMD has announced products and disclosed roadmaps in December in January for products that eventually launched in June or July. So an initial preview even in January, six months before availability, would be perfectly in line with AMD’s previous GPUs.

Assuming nothing in AMD’s roadmap has changed, the next high-end version of Navi (Navi 20) will be built on an enhanced 7nm manufacturing process (7nm+). As these things go, we can assume it will have faster clocks and better power efficiency, probably more GPU cores, and of course ray tracing.

Beyond that, there is not much in the way of official information. There are also plenty of questions to go around, though. For example, will Navi 20 be competitive with Nvidia’s top cards, or slot in a peg or two below whatever Nvidia’s fastest cards will be at the time? How much will they cost? And will Intel have a discrete GPU for gaming when Navi 20 actually launches?

Strap in, 2020 could be a wild ride.

Demand for HDD Storage Booming: 240 EB Shipped in Q3 2019

Demand for storage is stronger than ever as, in addition to growing user-generated data, machine-generated data now requires a formidable amount of storage space and will need even more in the future as the number of connected devices increases. Since the bulk of data continues to be stored on hard drives, it is not surprising that the third quarter was particularly successful for HDD makers, which despite the dent that flash sales are making, still set a record in terms of total shipped capacity.

Record 240 EB Shipped

The combined capacity of all of the hard drives sold by Seagate, Toshiba, and Western Digital in the third quarter totaled a whopping 240 exabytes (EB), of which 140 EB were nearline HDDs, according to TrendFocus. To put the number into perspective: the combined capacity of all HDDs and SSDs shipped last year was 912 EB and, today we are talking only about hard drives in one quarter. In fact, even when compared to Q2’s 207.5 EB, this is still 15% growth.

Market observers are attributing the significant rise in shipping HDD capacity to increasing shipments of enterprise-class nearline 14 TB, 15 TB, and 16 TB hard drives, which by now have been qualified by a substantial number of companies. For example, the average capacity of Seagate’s enterprise HDD (mission critical + nearline drives) was 6.3 TB in Q3 2019, up from 5.2 TB in the same period last year. By contrast, the average capacity of Seagate’s client hard drives was at 1.2 TB (unchanged from Q3 2018), as the bulk of such drives are intended for notebooks and their capacities range from 500 GB to 1 TB. In terms of units shipped, client and consumer electronics HDDs by far outsell enterprise and nearline drives, so, the average capacity of a Seagate HDD is 2.9 TB, up from 2.5 TB a year ago.

Unit Shipments Down

When it comes to total unit shipments however, not everything is quite as rosy, based on data from TrendFocus. The three manufacturers sold 83 million HDDs in the third quarter, up from 78.6 million in Q2 and 78 million in Q1, but down from around 88 million in the same quarter of 2018. In fact HDD sales have been in decline for a long time. Nidec, a Japanese motor manufacturer who is responsible for around 85% of all HDD spindle motors, said earlier this year that unit sales of hard drives had declined by around 43% from 2010 to 2018, going from around 650 million units in 2010 to 375 million units in 2018. Based on Nidec’s forecast, HDD shipments will drop to 309 million drives in 2019, which will further drop to 290 million units in 2020.

Market Shares

When it comes to market shares in Q3 2019, Seagate was leading the pack with a 40.2% share, Western Digital followed with 35.4%, whereas Toshiba trailed with 24.4% of HDD unit shipments.

For readers interested to learn more about Seagate’s and Western Digital’s HDD businesses today, we included some supplementary information below.

Corsair’s New iCUE QL PWM Fans Pack 34 Addressable RGB LEDS Each

It appears that RGB has been in the market for so long that companies are starting to feel the need to up the ante, which is exactly what Corsair is doing with its new iCUE QL RGB PWM fans announced today. For the time being, the company is only announcing them in 120 and 140mm variants.

What’s special about these RGB-enabled fans is that they have not one, two or even three lighting zones. No, they have four lighting zones spanning over a grand total of 34 individually addressable RGB LEDs per fan. I don’t think my entire PC even has that many RGB LEDs in the entire build!

The lighting zones are located at the outer and inner rings of the fans on the front and exhaust sides. Corsair’s goal here was to make accompanying systems “look great from all angles,” its announcement said. The inner ring of the intake side of the fan also passes some light into the lightly translucent fan blades for added effect.

Thankfully, all this RGB craze seemingly hasn’t gone at the total expense of the PWM fans’ performance. Corsair is still trying to ensure these units aren’t all too loud with built-in vibration dampers, PWM control and silent fan hubs. The 120mm units will run at up to 1,500 RPM while pushing up to 41.8 CFM of air, and the 140mm units are said to push up to 50.2CFM at 1,250 RPM. Corsair rates both at 18-26dBA through their RPM range, and although they aren’t expected to perform at industry-leading silence levels, these are fairly respectable numbers (assuming Corsair’s testing conditions were fair). They’ll go down to 525 and 550 RPM for the QL120 and QL140, respectively, so based on this information I wouldn’t concern yourself with excessive noise levels.

Corsair will sell the units in single-packs and in three-packs (120mm) or two-packs (140mm) with an included Corsair iCue Node CORE hub that allows you to connect up to six fans. They come with a two-year warranty, and the fans are rated to operate with a mean time between failure of 40,000 hours.

Corsair’s color-crazy fans are currently available at Amazon starting at $44.99 for 120mm or $49.99 for 140mm. The 120mm three-pack is $139.99, and the 140mm two-pack is $119.99, but “temporarily out of stock.”

Grab a Radeon RX 5700 graphics card for just $290 right now

AMD’s Radeon RX 5700 is a great graphics card, if you can’t spend a little extra money for the 5700 XT. The card has dropped to $290 a few times, and now one model from XFX has returned to that price. For comparison, most other RX 5700 cards start at $330.

The model on sale uses AMD’s reference design, meaning there is no factory overclock, no custom cooler, and so on. Don’t worry, though—our review of the reference 5700 found that it still out-performed the RTX 2060 and AMD’s own Vega 64. It’s a great graphics card for 1440p and 1080p gaming, though you may have to turn settings a bit on the latter (at least in super-demanding games) to maintain silky-smooth 60 fps.

The exact hardware specifications include a core clock of 1465 MHz, a boost clock of 1725 MHz, 8GB of GDDR6 memory, and 2304 Stream Processors. For video output, you get one HDMI 2.0b and three DisplayPort 1.4.

The SK Hynix Gold S31 SATA SSD Review: Hynix 3D NAND Finally Shows Up

SK Hynix has been in the NAND and SSD business for a long time, but we haven’t had the opportunity to review a drive with SK Hynix NAND in years. In most respects, SK Hynix fared the worst with the transition to 3D NAND, and their 3D NAND has not been used in very many SSDs aside from their own models. SK Hynix 3D NAND has been considerably more popular in mobile applications like smartphones and memory cards, and their client OEM SSDs are widespread but not sampled for review. This year, SK Hynix decided to start competing directly in the retail SSD market by introducing the SK Hynix Gold S31 SATA SSDs. (Outside North America, SK Hynix has been selling some SSDs under subsidiary brands.)

The Gold S31 showcases SK Hynix’s vertical integration with the NAND, DRAM, controller and firmware all produced in-house. Samsung is the only other company that regularly produces SSDs with this degree of vertical integration; Micron uses third-party controllers for most of their SSDs, and the other NAND manufacturers have in-house controllers but don’t make their own DRAM.

The NAND in the Gold S31 SSDs is SK Hynix’s fourth generation 3D NAND, a 72-layer design that is competing against the 64L and to some extent 92/96L generations from other manufacturers. All three capacities of the S31 are made using 512Gbit TLC dies.

SK Hynix didn’t share much information about the SSD controller that the S31 uses, but it’s codenamed “Quartz” and is a fourth generation design. Based on the chip markings of “SH87830CC” it would appear to be a descendant of the LM87800 controller that was part of Hynix’s 2012 acquisition of Link A Media Devices (LAMD).

The Gold S31 uses LPDDR3 for its DRAM cache. Aside from Samsung, most SSD vendors use DDR3L or DDR4 instead of any generation of LPDDR. Since Hynix is one of the major DRAM manufacturers, it isn’t too surprising to see them go a bit fancier here.

Combined, these components are the backbone of a mainstream SATA SSD. SK Hynix has attached the typical 5-year warranty and ~0.3 drive writes per day endurance rating, and priced it to compete against other mainstream SATA drives.

The Gold S31 is only available in the 2.5″/7mm SATA form factor, but a peek inside at the PCB layout shows how easy it is now to fit a reasonable amount of storage on a much smaller M.2 card. The PCB takes up only a fraction of the interior of a 2.5″ drive case, and the board is held in place by a plastic insert. The PCB is double-sided to accommodate two NAND flash memory packages, and on the primary side of the PCB are one package each for the controller and NAND.

SK Hynix sampled us the full range of capacities for the Gold S31. Their competitors haven’t sent us any current-generation 250GB-class mainstream SATA drives, so we’re only comparing the smallest S31 against the DRAMless Toshiba TR200 and against larger drives. To see how the 250GB S31 matches up against older drives of similar capacity, head over to our Bench database.