HW News - New NVIDIA GPUs (Yes, Really), Steam Deck Delay, Grand Theft GPU
Posted on November 15, 2021
This week, we have news regarding an unusual transaction between two x86 license holders: Intel and Centaur Technology. The transaction is reportedly worth $125M, and will see Intel take on some of Centaur Technology’s development talent. Though we suspect there’s more to the deal than that; we’ll just have to wait and see.
Additionally, we have a new AMD RDNA 2 GPU to discuss, one aimed at cloud offerings and virtualization. EVGA has also reported that the company suffered a massive loss thanks to a stolen shipment of RTX 30-series cards, and is warning that such cards will not be honored in terms of warranty service nor will they be eligible to be registered.
Elsewhere, we have news from Cadence on PCIe 6.0 IP and test silicon, new PSUs from Asus with 12-pin PCIe 5.0 connectors, and more. Find the news video and article below, as usual.
01:02 - GN News: New Office & Store Updates
This one is just for the video!
04:43 | Windows Power Plans & Core Assignment
Much was made of Alder Lake's onboard Intel Thread Director and how it would work with Windows 11 "to make the optimal scheduling decision for any workload or workflow." In an Intel presentation at Hot Chips in August, slides stated "background threads of any class [are] directed to energy efficient core[s]" and "priority threads are directed to performance cores." Regardless of whether a high-priority thread runs on a P core or gets pushed off to an E core, it gets a higher voltage/frequency than a low-priority thread would.
Intel's official recommendation for Alder Lake testing was to use Windows 11, which is designed around the Hybrid Architecture and Thread Director. From the Alder Lake Keynote: "Using OS hybrid optimization along with Thread Director feedback, background activities may be efficiently offloaded to E-cores, unlocking massive performance and responsiveness of P-cores." The reason we're quoting so frequently here is that the Windows 11 optimizations are described as a way to help multitasking, not improve performance in isolated workloads, which is part of why our Windows 10 scores were as good or better than Windows 11. Windows 11 does not magically make Cinebench scores higher.
One further wrench thrown in the works is that Windows 11 places higher priority on foreground apps. Quote: "We've done a lot of work in memory management to favor the app windows you have running in the foreground so that they're prioritized with more CPU and other system resources." That in itself isn't new, but Anandtech's coverage of the Alder Lake press briefing reported that with the Windows 11 Balanced power plan enabled, threads would be run on E-cores or P-cores based directly on the foreground window in a non-intelligent way, potentially overriding the Intel Thread Director's suggestions.
This sounds like an extension of the existing Microsoft Quality of Service policy that intentionally throttles background applications, but Microsoft made the E-versus-P core choice sound more subtle and intelligent prior to Alder Lake's launch: "Instead of just knowing whether an application is in the foreground or whether the application developer has set EcoQos, the system can make sure threads with higher performance requirements […] are assigned to performance cores (P-cores)." Intel expressed something similar: "Schedulers typically work with limited information, like whether an application is running in the foreground or background. Our performance monitoring unit can access much more from the hardware, including the instruction mix, the state of individual cores, and other relevant telemetry."
If the end result really is as simple as "front window get good core," it takes the wind out of those earlier claims. Regardless, we've been doing all of our CPU testing with the Windows High Performance power plan for years, ever since the last time the Balanced plan got screwy. Some quick and dirty testing with Cinebench single-threaded suggests that the workload does indeed move between E and P cores depending on the focused window, but further testing is needed.
11:13 | RTX 2060 12GB Inbound, New 30-Ti SKUs
This one is also just for the video.
Source: GN
13:52 | VIA Sells Part of Centaur Technology to Intel For $125M
In the world of x86 CPU designs, there’s primarily three players: Intel, AMD, and to a very small extent, Centaur Technology. While Intel and AMD are well known in the CPU landscape, Centaur Technology, a company that has spent most of its existence as a subsidiary of VIA Technologies, is certainly less known. And while Centaur Technology has been dabbling in x86 designs and IP since the early 1990s, its designs have never seen the kind of widespread adoption that Intel and AMD have. The closest is the Zhaoxin set of CPUs we reviewed.
All that is to say that news recently broke of a $125M deal struck between VIA and Intel to acquire...something regarding Centaur Technology. At this point, all we really know is that Intel will recruit some of Centaur’s engineering and development talent. And while it isn’t altogether surprising to see Intel make a consolidation move for Centaur -- Intel actually bought development assets from VIA back in 2015.
This isn’t a flat out sale of Centaur Technology, and so far as we know, VIA and Centaur will retain an x86 license, as well as the right to develop and sell x86 IP. There’s also no mention of assets, patents, or IP changing hands. Though, we have to imagine that $125M nets Intel something more than a recruiting opportunity, but neither company is prepared to disclose the details of the transaction. Intel has confirmed the deal to the press (such as AnandTech, who originally spotted the news) but provided no details. Interestingly, Centaur Technology’s website has been taken down, with a vague “under construction” message now appearing on the landing page.
Centaur Technology was born in 1995 under the banner of IDT, and was initially formed to develop x86 designs to compete with Intel’s offerings at the time. Centaur was snatched up by VIA in 1999, and has remained under VIA’s roof ever since. Centaur has a number of x86 designs, such as the VIA C3, C7, and VIA Nano, the latter of which would go on to be a fundamental part of the joint venture between VIA and Zhaoxin. Centaur’s most recent offering is the x86 CNS core, which the company framed as its new server-class core.
16:57 | AMD Announces Radeon Pro V620 GPU For Cloud Workloads and Game Streaming
In something of a surprise move, AMD rolled-out a new Radeon Pro V620 RDNA 2 GPU. AMD’s new V620 is aimed at virtualization, with AMD explicitly citing cloud gaming and “immersive AAA game experiences.” AMD also envisions the GPU for use in 3D workloads, visual compute, machine learning, as well as office productivity and development environments, such as DaaS (Desktop-As-A-Service) or WaaS (Workspace-As-A-Service) deployments.
The Radeon Pro V620 will ship with features both germane to RDNA 2 and of a GPU aimed at virtualization in the cloud. Such features include AMD’s Infinity Cache, hardware-based ray tracing, AMD FidelityFX, GPU partitioning, and support for both DirectX and Vulcan APIs. As is standard for GPU sharing, the card will also support SR-IOV-based sharing technology to protect data and traffic between users sharing PCIe devices.
Hardware wise, the Radeon Pro V620 offers 4,608 stream processors and 72 compute units. That’s in addition to 32GB of GDDR6 ECC memory operating at 16Gbps with a memory bandwidth of 512 GB/s over a 256-bit memory bus. The GPU has a board power of 300W and uses one PCIe x16 slot. AMD notes that the Radeon Pro V620 is available immediately, though this one is obviously only for data center customers and public cloud offerings.
19:16 | Cadence Rolls Out PCIe 6.0 Test Silicon and IP Built on TSMC’s N5
The PCI-SIG recently published the final draft (version 0.9) of the PCIe 6.0 specification, and Cadence is the first company out of the gate with test silicon and IP based on the emergent PCIe 6.0 interface. The new IP and test silicon from Cadence will allow Cadence customers to explore how to best integrate PCIe 6.0 into their products, as well as test for compatibility and interoperability.
Cadence is offering IP for a PCIe 6.0 DSP-based PHY and controller, as well as 5nm test chip silicon built on TSMC’s N5 node. Cadence notes that its PCIe 6.0 PHY test silicon delivered optimal electrical performance spanning all PCIe data rates, and that the PAM4/NRZ dual-mode transmitter maintained optimal signal integrity with low jitter. The test chip will allow customers to test signal integrity and jitter in their own PCIe 6.0 compatible products. The IP package offered by Cadence is also based on TSMC’s N5 node, and will allow vendors to add/test support for a number of PCIe 6.0 products, such as SSD controllers, ASICs, GPUs, etc.
21:30 | EVGA Reports Truckload of RTX 30-Series Cards Stolen
EVGA has reported that someone (or multiple someones) has stolen a truckload of the company’s RTX 30-series cards. The theft occurred somewhere in California, as the truck was en route to EVGA’s distribution center in Southern California, as per EVGA’s post on the company’s forums.
“PLEASE TAKE NOTICE that on October 29, 2021, a shipment of EVGA GeForce RTX 30-Series Graphics Cards was stolen from a truck en route from San Francisco to our Southern California distribution center.
These graphics cards are in high demand and each has an estimated retail value starting at $329.99 up to $1959.99 MSRP,” reads the post.
While EVGA omitted any details regarding specific models, the $330 to $1,960 price range pretty much covers the entire gamut for Nvidia’s RTX 30-series. Furthemore, EVGA’s post notes that warranty and claims will not be honored on these stolen cards, and the cards cannot be registered, unsurprisingly. EVGA no doubt is in possession of serial numbers and lot codes, so it shouldn’t be hard for them to identify a stolen card should it surface somewhere.
The global shortage and insatiable demand for GPUs (and other silicon) has paved the way for criminals and bot scripts to capitalize on any inventory that can be found, and in turn scalp the products for well above MSRP. That’s no doubt what’s happening here, and is yet another reason to not pay a scalper.
Source: https://forums.evga.com/Notice-of-Stolen-EVGA-GeForce-RTX-30Series-Graphics-Cards-m3490851.aspx
24:19 | Asus Thor II PSUs To Use 12-Pin PCIe 5.0 Connector, More Vendors Likely To Follow
Asus recently announced its new line of ROG Thor II PSUs, and interestingly enough, the new units come equipped with a 12-pin PCIe 5.0 connector -- one that looks alot like Nvidia’s 12-pin connector it pushed so hard for, that ultimately only ended up on its own Founders Edition models.
According to Asus, Nvidia's 12-pin connector is in fact PCIe 5.0 compliant. This coincides with recent news of a new 12VHPWR connector, first spotted by Igor from Igor’s Lab. The 12VHPWR connector is a 12-pin PCIe 5.0 connector, with an additional four signal/sense contacts at the bottom of the connector beneath the power pins. The connector being shown by Asus doesn’t have these four pins, so they appear to be optional.
The 12VHPWR shows 12 pins in the body of the connector itself, with a reported 9.2A per pin. Interestingly, Amphenol ICC already has the connector listed on its website, where the data sheet shows the connector supporting 9.5A per pin. According to Asus marketing, the new 12-pin cable is capable of delivering up to 600W, which again, jives with what we know about the 12VHPWR connector. That’s quite a bit more than what can be achieved with three 150W PCIe 8-pin connectors plus the 75W a PCIe slot can provide.
The new connector seems to be an effort to address two things born of each other: The trend of increasingly power hungry graphics cards, and multiple 8-pin headers affixed to a PCB. It isn’t uncommon to find current-gen cards with as many as three 8-pin headers, like on the Gigabyte RTX 3080 Xtreme that died testing Gigabyte PSUs. Nvidia seemed to be ahead of the curve on this one when it pushed for a 12-pin connector, which ultimately only ended up on its own FE RTX 30-series cards. However, the industry seems to be running out of options at this point -- there’s only so many 8-pin headers you can cram onto a board.
Source: https://www.tomshardware.com/news/nvidia-12-pin-pcie-power-connector-pcie-5-gpus
25:59 | Valve delays steam deck until [whenever]
Valve announced that the Steam Deck will now be delayed until February. The post from the Steam Deck store page reads more like Valve isn’t sure when delivery will actually happen, so don’t be surprised if the date is pushed back again. Valve cited “material shortages and delays” without getting specific about anything.
Reservation queues remain in the same order, and everyone gets shifted back to the new timeline. New estimated order availability time frames can be checked on the product page.
26:31 | NVIDIA Omniverse Avatar
On Tuesday, NVIDIA demo’d an animated toy avatar of Jensen Huang using its Omniverse animation systems. The video, which is available on NVIDIA’s YouTube channel, shows toy Jensen answering domain-specific questions to subject matter experts in real time. The point of the demonstration was to show that we’re getting closer to developing a JARVIS-like service and reference AI that can help answer questions for customer service, daily tasks, and research.
This iteration of the AI is from Project Tokkio, or “talking kiosk”, which combines elements of the NVIDIA technologies Omniverse, Maxine, Riva, and the Megatron-Turing NLG 530B large language model.
The Omniverse platform was used to build the surrounding scenes and create the facial and body animations based on the audio source. The Maxine SDK was used to make it photorealistic. NVIDIA Riva was added for speech AI, and the Megatron-Turing NLG 530B large language model was implemented for natural language understanding.
Ideally, we’ll have more realistic, helpful, and cuddly avatars, like a baby Yoda, to answer our medical research and astrophysics questions, but for now, toy Jensen will have to do.
28:32 | The Noctua Foam Stack!
Noctua recently announced a pair of compact air coolers, the NH-L9i-17xx and NH-L9i-17xx chromax.black, and a ground-breaking new technology, the FOAM STACK! The NA-FD1 is literally a stack of 7 foam spacers that can be combined in various height arrangements to create an air duct between a Noctua cooler and a case panel. In the installation guide video, Noctua claims a 6C CPU temperature improvement in two specific cases by adding the foam stack, sorry, fan duct. Perhaps we’ll get one of these in and test it to see if we can produce similar results.
The foam stack, sorry, fan duct is compatible with the two new coolers and their 4 predecessors the NH-L9i, NH-L9i chromax.black, NH-L9a-AM4, NH-L9a-AM4 chromax.black. The official specs of the coolers list them at the same 37mm height as the other NH-L9s when the fan is attached. Regarding performance, Noctua CEO, Roland Mossig stated, “We have managed to dissipate up to around 160W on the Core i9-12900K, pushing it to over 4.2GHz, and up to 125W on the Core i5-12600K running at 4.3GHz.”
Source: https://noctua.at/en/na-fd1
Host, Additional Reporting: Steve Burke
Reporting: Eric Hamilton
Additional Reporting: Patrick Lathan
Additional Reporting: Patrick Stone
Video: Keegan Gallick