Oct 15, 2016

The Week In Storage And CPUs: Transistors Hit 1nm, PCs Continue To Plummet, IBM Takes Cap Off CAPI

This week in storage and CPUs found news that researchers are developing the first 1nm transistors, which is a promising development as Moore's law grinds to a halt. Speaking of grinding to a halt, the PC market continued its frightening fall. The industry is pounding the war drums of open interfaces, and IBM and all its friends joined forces to combat Intel's aspirations of an Intel-branded datacenter. We'll cover those after the recap. There's a lot of cud this week, so let's get right to chewing.

The goodness of Kaby Lake is almost upon us, so it's no surprise to see the motherboard vendors lining up to support Intel's latest. Asus started by slapping the "Kaby" tag on 87 of its 100-series motherboards last week, and ASRock weighed in with its own salvo of upgrades this week. Not to be forgotten, Apollo Lake SoCs are also about to wade into the marketplace, so Asus released its J3455M-E motherboard, but in a case of oneupmanship, ASRock announced five new Apollo Lake motherboards with passive cooling. 

WD armed itself with SanDisk for a mere $16 billion when it gobbled it up during a fire sale. Now WD has NAND fabs and a war chest bristling with flash patents (5,000+), so it is free to assault the SSD market. WD made its somewhat triumphant return to SSD with its new Blue series, but it falls below our pricing expectations. NAND is non-volatile, but the SSD market surely is; prices fluctuate rapidly, so we expect WD to adjust its pricing soon. Of course, the high price may just be WD dipping its toes into the SSD waters to see what the market will stomach, or perhaps the NAND shortage is already rearing its ugly head. We'll know in a few weeks.

Intel famously acquired Altera for the tidy sum of $16.5 billion (which seems to be the going price for everything these days), and the first fruits of its acquisition come to market in the form of the Stratix 10 FPGA. The new chip barrels in with the 14nm+ tri-gate process and up to 1TB/s of memory bandwidth from on-package HBM2 in tow. If that doesn't tickle your fancy, perhaps the 5.5 million logic elements or the impressive 80 GFlops-per-watt will (good eye, bit_user). Intel also threw in a 64-bit quad-core ARM Cortex-A53 processor just to get us talking, but I wouldn't be surprised to see that change to an Atom in the future.

A new day dawns, and with it, a new interconnect. The Gen-Z consortium consists of industry heavyweights such as AMD, ARM, Dell EMC, IBM, HPE, Micron, Samsung, Seagate, SK hynix, Western Digital, and others, and the group is promoting a new bus interconnect for next-generation storage class memory. Unlike other interconnects, the Gen-Z spec uses memory semantics (load/stores - put/gets) to speed up the fabric. Gen-Z is perhaps most notable for its lack of Intel. The industry is fearful that Intel will utilize a proprietary interface for 3D XPoint and other emerging memories, so the broad consortium aims to strike back at the Intel empire with a free interconnect to keep prices in check.

Patriot's Hellfire M.2 NVMe SSDs have a great name that evokes images of missiles, bombs, and fire. These are always the first things on my mind when I'm looking to upgrade my computing experience, so I'm all ears. Chris Ramseyer shot the Hellfire at our test regimen and found it to offer much better performance with the new Phison E7 firmware. It also has an attractive price point. The SSD market is in a state of flux as a new truckload of products come to market, so it might be wise to wait for more products in the next month before making your next purchase. 

Researchers Bust Out Lube And Carbon Nanotubes To Design 1nm Transistors (A Little Lube Fixes Everything)

Carbon nanotubes are the cure for all that ails you; at least, that's what researchers think. We've been awash in carbon nanotube news as of late, as researchers recently created nanotube-based transistors that outperform silicon, and Fujitsu weighed in with an announcement that it is bringing carbon nanotube-powered NRAM memory to market next year. Of course, all of this nanotube-iness began back in 1952, and we have yet to experience the revolution, so take the news with a grain of salt (nanotubes?).

Researchers at the University of California, Berkeley, added to the recent news with the announcement that they have created a 1nm transistor using carbon nanotubes and molybdenum disulfide (MoS2--a common engine lubricant) in lieu of silicon. The development of a 1nm transistor is important because most industry leaders predict that transistors, and thus the seemingly-immortal Moore's Law, will hit a brick wall at 5nm. Semiconductor fabs have already encountered significant quantum tunneling challenges as they reduce the size of the silicon transistors below 14nm, which is one of the key contributors to the recent slowdown in the chip-shrinking cadence.

Transistors have a source, drain, and gate. In the image above, the current passes from the source to the drain, which the gate controls. The gate switches when voltage is applied. The electrons that flow through the MoS2 channel are heavier because it has more resistance than silicon, thus allowing their flow to be controlled with shorter gate lengths. Researchers can roll MoS2 into 0.65nm-thick sheets while retaining the ability to store energy in an electric field, but there is no suitable way to forge a 1nm gate using standard lithography techniques. The researchers pulled out the trusty hollow carbon nanotubes for the proof of concept. However, we should keep in mind that this is merely a proof of concept, and the new technology has no plans, at least as of yet, for mass production.

The researchers published a paper in Science that one can access with a subscription. The abstract notes:

Demonstrating a new technology and actually overcoming the economic barriers that stand between all promising technologies and mass production are always two vastly different things. The new development does show the ability to overcome the limitations of Moore's Law with new materials and techniques, but it will be some time before we see a true production-worthy silicon alternative emerge.

IBM Takes The Cap Off CAPI

CAPI (Coherent Accelerator Processor Interface) is perhaps the most exciting accelerator interface that you've never heard of, largely because the promising interface is relegated to the high end of the server market. Accelerators are becoming more popular as the rapid generational increases in computing power are waning. These slip-in computational power boosters attack a range of workloads such as compression, encryption, networking, and storage applications. Accelerators wield FPGAs and ASICs, and even flash storage, to great effect; but the nefarious PCIe bus stands in the way of unlocking the true potential of the speedy add-ons.

IBM developed the CAPI interface to tackle the bandwidth problem. CAPI utilizes standard PCIe connections but leverages a custom protocol that provides up to 50% more performance (25Gbits per second) than a typical PCIe 4.0 slot (16Gbits per second). IBM claims the combination of accelerators and CAPI offers up to 10X the workload performance in big data applications, largely due to increased bandwidth and reduced data movement (it's coherent). Using the existing PCIe infrastructure with an optimized interface, much like Nvidia does with its proprietary NVLINK (which IBM is also a vocal proponent of), affords a faster pathway with lower latency that can help to address the most challenging of workloads, such as machine learning, with customized high-performance accelerators.

IBM started down the path of open source-ry through its OpenPOWER Foundation, which began with a mere five members but has ballooned to 260. The initial effort involved opening up the "target" side of the interface, which was necessary to spur development of third-party accelerators. Now that a robust ecosystem of cards has appeared, the company has decided to open up the other end (host) of the CAPI interface. 

IBM still stands behind PCIe and NVLINK, but it has brought together a broad industry conflagration, which includes AMD, Dell, EMC, Google, Mellanox, Micron, and Xilinx, to support the new OpenCAPI Consortium. The purpose of opening up the spec is to allow other vendors, and processor architectures beyond IBM's own Power products, to leverage the speedy interface. Xilinx announced that it is whipping up FPGAs to support the new interface, and Mellanox has plans to infuse its networking gear with the new interface, as well.

Google and Rackspace have a new server codenamed Zaius that will leverage POWER9 processors and the OpenCAPI interface. IBM also plans to introduce its OpenCAPI-powered servers in the second half of 2017.

The formation of OpenCAPI, and the heavyweights behind the initiative, signals yet another broad consortium of industry players that are lining up behind open interfaces, as opposed to the proprietary interfaces that hamstring development and promote stratospheric pricing models (Intel, they are eyeballing you).

Most notably, ARM and AMD are in the OpenCAPI consortium, so look for the interface to spread quickly amongst the Intel competition. We asked, and yes, Intel is free to join the consortium. If enough of the industry adopts the interface and moves forward while leaving Intel behind, Intel might not have a choice.

PC Market Continues To Slump--Just Not Quite As Slumpy

The poor PC market has been under duress for an extended period of time, so it really is no surprise that the decline continues. Gartner released its preliminary quarterly PC Tale of Doom, otherwise known as the PC Quarterly Statistics Worldwide By Region report, which indicates that PC shipments have experienced a record-setting second year of declines.

Worldwide PC shipments weighed in at 68.9 million units in 3Q16, which represents a 5.7% decline over the previous quarter. Gartner attributes much of the loss to the proliferation of computing devices, which includes phones and tablets. The company predicts that the average consumer wields at least three computing devices, which may help to reduce the need to upgrade their notebook and desktop PCs. It's notable that both the smartphone and tablet markets are also in decline, so the proliferation of devices may have a knock-on effect across the board.

The U.S. market has improved to the tune of 15.3%, but the global picture helped to drag down overall PC shipments. The report also noted that the top six vendors claimed a record-setting 78% of the market share, which might indicate that smaller vendors are suffering as the industry continues to consolidate. Surprisingly, Intel recently increased its 3Q16 revenue projections due to increased chip sales, so perhaps there is a glimmer of hope on the horizon.

The current malaise in the computing industry likely bears some of the blame, as well. We’re stuck in a never-ending cycle of incremental upgrades that don't really bring any new exciting technologies to bear, so there is nothing to spur a heavy refresh cycle. The industry envisioned a sharp increase due to the Windows 10 upgrade, but it appears that never materialized, likely because consumers could simply upgrade their existing operating system for free.

No comments:

Post a Comment