Industry stub

HW News - 2.5nm Transistor, AMD Graphics Brain Drain, RTX 2060

Posted on December 19, 2018

Despite EOY slow-downs in the news cycle, we still spotted several major industry topics and engineering advancements worthy of recap. Aside from Intel's recent announcements, the most noteworthy news items came out of MIT for engineering efforts on 2.5nm-wide transistors, out of Intel for acquiring more AMD talent, and out of the rumor mill for the RTX 2060, which is mostly confirmed at this point.

As always, show notes are below the embedded video:


Engineers Construct 2.5nm Transistor

Researchers from MIT, collaborating with the University of Colorado, have managed to build a 3D transistor coming in at only 2.5nm, using a fabrication process that works at an atomic level. The researchers believe that in order to keep up with Moore’s Law, which some have declared dead, semiconductors will have to be manipulated atom by atom.

In order to do this, the researchers use 3D transistors (FinFETs), that stand vertically like fins, and are thousands of times thinner than a strand of human hair. To create the smaller FinFETs, they use a modified thermal ALE (Atomic Level Etching) technique. The result is FinFETs as narrow as 2.5nm, and theoretically more efficient than commercial counterparts.

MIT’s website does a great job of describing the process, writing:

“Microfabrication involves deposition (growing film on a substrate) and etching (engraving patterns on the surface). To form transistors, the substrate surface gets exposed to light through photomasks with the shape and structure of the transistor. All material exposed to light can be etched away with chemicals, while material hidden behind the photomask remains. The state-of-the-art techniques for microfabrication are known as atomic layer deposition (ALD) and atomic layer etching (ALE). In ALD, two chemicals are deposited onto the substrate surface and react with one another in a vacuum reactor to form a film of desired thickness, one atomic layer at a time. Traditional ALE techniques use plasma with highly energetic ions that strip away individual atoms on the material’s surface. But these cause surface damage. These methods also expose material to air, where oxidization causes additional defects that hinder performance.”

The researchers believe this method could be easily integrated into existing fabrication processes, with real world impact in the form of denser chips, higher performance, and increased yields.

To learn more about the intricacies of the process, check out the report from MIT linked below.    


Intel: Sunny Cove, B365, Gen11 & Xe Graphics

Intel held its Architecture Day 2018, where they made several announcements including the new Sunny Cove microarchitecture headed for both Xeon and Core chips. Additionally, they made announcements for their new Xe Graphics brand (no Arctic Sound news just yet) and hybrid x86 CPUs. Although not related to Intel’s Architecture day, Intel also quietly released the B365 chipset.

Intel pulled the wraps off Sunny Cove, something of a clean slate architecture for the company. For years, Intel maintained a tick-tock cadence, where a new architecture was tied to a new, smaller process node. Intel hit a massive roadblock with the shrink to 10nm, being stuck with 14nm and a Skylakeesque architecture since 2014. This, coupled with AMD’s resurgent competitiveness, meant that Intel needed to rethink their approach to architecture design.

Thus, Sunny Cove is the first of three new architectures in Intel’s roadmap that are designed to be portable between nodes, and will be Intel’s first raw IPC increase since Skylake. Sunny Cove will debut on the 10nm node and is slated for the first half of 2019, although it isn’t clear which chips the architecture will accompany. Interestingly, Anandtech speculates that Sunny Cove cores paired with Gen 11 graphics will debut as Ice Lake.

Accompanying Sunny Cove is Intel’s new Gen11 graphics engine. With Gen11, Intel increased execution units from 24 to 64, with compute performance seeing an increase pushing it past 1 TFLOPS. The new Gen11 design will also support tile-based rendering and will offer an Intel version of adaptive sync. Intel’s adaptive sync technology will not require any external hardware and will be compatible with FreeSync monitors.

Beyond Gen11 is Intel’s new Xe Graphics brand, which will take them into discrete GPU territory. Details are scarce, and Intel more teased than anything else. However, it seems that with Xe, Intel is aiming for scalability, hinting they intend to design GPUs ranging from integrated solutions all the way to mid-range and enthusiast discrete GPUs, even targeting data center. These designs will all likely be underpinned by the same architecture. Intel also reaffirmed their plan to bring discrete GPUs to market by 2020.

Lastly, Intel debuted a new “hybrid x86” processor, built with their new 3D chip stacking process, Foveros. Foveros will allow various dies and silicon to be stacked on the same package, in conjunction with EMIB (Embedded Multi-Die Interconnect Bridge). This will allow chips to be built with a mix and match approach combining CPUs, GPUs, and different I/O elements with a high speed interconnect.

Intel demonstrated a working sample for an unnamed customer, whereby they combined one Sunny Cove core and four Atom cores in a 10nm design, coupled with a separate 22nm I/O chip connected through TVS (Through Silicon Via). Intel announced they’re planning to roll out an entire product line with Foveros designed chips.

As an aside, Intel also quietly released the B365 chipset. Like the H310C chipset, the B365 is a 22nm variant of the 14nm B360, and uses a Kaby Lake PCH. This is no doubt another move to free up fab space for 14nm silicon production, as Intel continues to struggle with meeting demand for 14nm CPUs.                


GN deep dive with David Kanter:

Rumor: AMD Vega II and New Vega 10 and 20 GPUs

New reports of AMD filing a trademark for a new Vega logo suggest that the company isn’t done with Vega just yet. Looking at the logo, one could interpret it two ways: as either Vega II (2), or as the roman numeral VII, denoting a 7nm shrink.

Further fueling the Vega refresh rumor mill is the discovery of new device IDs in a Linux driver. AMD released patches to RadeonSI Mesa and AMDKFD/AMDGPU kernel drivers, where a total of 7 new device IDs were spotted by Phoronix -- one new Vega 20 ID, and six Vega 10 IDs.

While logos and device IDs can only lead us to speculation, it’s clear AMD is planning something with Vega, presumably to exist alongside Navi. Hopefully we’ll know more around CES 2019.


Rumor: RTX 2060 Alleged Host of 1920 Cores, TU106

Videocardz has allegedly “confirmed” the existence of a Gigabyte RTX 2060 card, launching early next year, according to the source. The leak asserts that RTX 2060 will boast 1920 Cuda cores and the TU106 GPU silicon. The card is also rumored to have 6GB of GDDR6 memory, unsurprisingly.

Comparatively, the current crop of GTX 1060s run 1280 Cuda cores, 6GB of GDDR5, or GDDR5X after a recent refresh. Still no word on Tensor cores or real time raytracing for 2060 cards. We, along with many others in the tech space, have speculated the RTX 2060 won’t offer that technology. Moreover, it’s entirely possible that none of this is accurate to begin with.

Again, we’re just around the corner from CES 2019, where a lot of these rumors will be confirmed or put to rest. Probably, anyway.     


AMD Brain Drain Into Intel

Intel has poached yet another source of talend from AMD’s group, this time in the form of one Damien Triolet. Triolet, some may remember, was the former Senior Editor at French publication Triolet specialized in GPU reviews, which is ultimately what attracted AMD to bring him aboard as their Technical Marketing Manager.

As of November 2018, it seems Triolet has jumped ship to Intel, and is now in a position of  Technical Marketing within Intel’s Gaming and Graphics department. Intel recently snatched up Ryan Shrout, formerly of PC Perspective.


NVIDIA Shares Slide to Half Value, Investor Wants Out

We recently reported on Nvidia’s earnings, and how they suffering from a self induced “crypto hangover”. As it would seem, that hangover isn’t subsiding anytime soon. Nvidia put a lot of eggs in the cryptocurrency basket, failing to predict when that bubble would burst (as did AMD, for that matter) and has been left with excessive amounts of GPUs in the channel, particularly GTX 1060s.

As such, they had to temper their Q4 earning expectations with a dose of this reality, lowering it by more than $1 billion. Subsequently, their stock market shares started taking a nosedive. Well, that nosedive seems to have not found a bottom end yet, as Nvidia’s shares recently hit roughly half of their peak value -- from $289.36 to $149.

Worse yet, one of Nvidia’s biggest shareholders, Softbank Group, wants out. According to a report from Bloomberg, Softbank is considering selling out sometime next year, after Nvidia’s less than stellar stock market performance. Softbank gained a $3 billion stake in Nvidia in 2017, and if it opts to sell, the investor could stand to net a $3 billion profit.


That’s it for our HW News this time. To catch up on past HW News episodes, head over to our hardware section. Be sure to check out our recent deep dive videos with David Kanter and Scott Wasson. Also, have a look at our newest mouse pad over at the GN store, or pick up our newest shirt.

Editorial: Eric Hamilton
Host: Steve Burke
Video: Andrew Coleman, Josh Svoboda