Hardware stub

HW News - AMD Multi-Chip GPU, Noctua Passive Cooler NH-P1, Xbox Super Resolution, & More

Posted on June 16, 2021

This week's news talks about the Beast Canyon NUC, Noctua's NH-P1 cooler, Xbox consoles getting AMD FidelityFX Super Resolution and other topics, like a potential SiFive buyout by Intel.

At GN, we’ve been busy reviewing Nvidia’s latest additions to its RTX 30-series. If you haven’t already, check out our RTX 3080 Ti review and RTX 3070 Ti review.

01:01 | GN PSU Reviews & New Shirts

We'll leave this to the video! Mostly recapping our recent PSU review.

04:11 | AMD Confirms Multi-Chip Design on HPC GPU

Last year, AMD launched its flagship HPC accelerator in the form of the Instinct MI100. Most notably, that GPU marked the arrival of AMD’s compute-focused CDNA architecture. Most of this year, rumors have been abound about its successor, and have especially pointed to an MCM design.

In a recent Linux patch, the patch notes confirm the suspicions of an MCM approach, which isn’t surprising. AMD, Nvidia, and Intel have all been working to bring MCM designs into their GPUs, and it appears AMD may be first to market here. To be clear, and because we’ve seen confusion, AMD has had dual-GPUs in the past, but that isn’t what we’re referring to here. A dual-GPU refers to two monolithic GPU dies sharing the same PCB, whereas an MCM approach integrates multiple dies onto the same package and substrate.

The Linux patch details a primary and secondary die, each having slightly different functions in order to make the GPU appear monolithic to the host system. “On aldebaran, only primary die fetches valid power data. Show power/energy values as 0 on secondary die. Also, power limit should not be set through secondary die,” according to the patch notes. 

AMD has a number of patents on GPU chiplets, wherein they describe using crosslinks and active bridges to interconnect chiplets in order to mimic a monolithic die to a host system, and theoretically overcome the traditional problems of workload scaling across multiple GPUs. AMD’s MI200 cards, alongside its Trento Epyc CPUs, are set to power the upcoming Frontier supercomputer that’s supposed to be delivered later this year.    

Source: https://lists.freedesktop.org/archives/amd-gfx/2021-June/065161.html

07:27 | AMD Expands on 3D V-Cache

As impressive as AMD’s 3D V-Cache is, it’s still behind the remarkable advancements of the gaming chair industry: Those already have 4 dimensions, while AMD is still stuck on 3. 

AMD published a new video in its “Bring Up” series, where the company talked a little more about 3D V-Cache advancements that were shown briefly during the Computex keynote. AMD noted that the active silicon is 20 microns of depth, with L3 attaching straight to that. It re-emphasized an earlier point that the attachment is done without bumps, adhesive, or solder. AMD stated:

“The moment you press them face to face, they’re bonded together just from the attractive forces of the copper atoms alone, which effectively means that there’s zero gap between these two dies.”

AMD reminded everyone that shorter distances reduce electrical noise and heat and that most vertical stacking solutions use microbumps. AMD noted that it’s using Through-Silicon Vias, or TSVs, which are filled and bonded with copper. The end result is more conductivity and a smaller footprint than traditional microbumps or C4. Not the explosive kind. 

The video is actually pretty good, except when they said “special sauce.”

AMD noted that the changes cut physical distance of cache-to-core to “1/1000th” of what it was when side by side. That means signals arrive 1000x faster and allegedly generate 1/1000th as much heat.

Source: https://www.youtube.com/watch?v=Uh3WobaaP70

10:58 | Intel Beast Canyon Leak

Intel’s “Next Unit of Computing” (or NUCs) have long been one of the best-tested Intel product lines in our reviews. So when an infamous Twitter-leaker like @9550pro tweets a picture of a new one, it gets our attention. The new NUC appears to be a “Compute Element” similar to the Ghost Canyon NUC that we looked at last year. The Ghost Canyon NUC was interesting for being a PCIe socketable computer, in essence, and for its overall unique physical hardware design. It was housed inside a small enclosure, mounted on a small board next to a GPU. The picture @9550pro tweeted looks very similar, with some small changes to the cooler of the compute element itself. The compute element is labelled i9-11900KB, indicating that it corresponds to the desktop cpu the 11900K. The i9-11900K was Intel’s competitor successor to the 10900K.

For a bit of background, back in 2018 we got our hands on the Hades Canyon NUC and pushed it further than Intel ever intended with custom liquid cooling. Last year, we looked at the Ghost Canyon NUC and earlier this year we looked at the Phantom Canyon NUC. Assuming there’s truth to this rumor, we look forward to testing the Beast Canyon when it comes out.

Source: https://twitter.com/9550pro/status/1402245251687870466

13:14 | Intel Reportedly Eying SiFive Buyout

A report from Bloomberg claims that Intel has put up $2 billion to acquire the chip designer SiFive. For the unaware, SiFive is perhaps the preeminent silicon designer for RISC-V, and interest in the firm -- in addition to the RISC-V ISA in general -- has skyrocketed since Nvidia has moved in to buy Arm. As of this writing, neither Intel nor SiFive have officially commented on the story, and Bloomberg is citing unnamed sources. 

In addition to courting interest from Intel, Bloomberg states that SiFive is also fielding offers from other parties, as well as investment opportunities. On that last point, there’s also a reasonable chance that with new investment capital, SiFive may choose to remain independent.

SiFive recently announced that it would be working with Intel’s new foundry services business, and while the details of that collaboration still haven’t been made clear, it’s widely assumed that it would involve making RISC-V designs available to IFS customers. IFS is already slated to license Intel x86 IP, and potentially Arm designs as well. It wouldn’t be surprising for Intel to want to offer another ISA option to its catalogue, especially given the heightened interest in RISC-V, and open ISAs in general.    

Source: https://www.bloomberg.com/news/articles/2021-06-10/chipmaker-sifive-is-said-to-draw-intel-takeover-interest

14:35 | Xbox Series X|S Will Get AMD Super Resolution 

In addition to Intel being interested in AMD’s newest Super Resolution technology, Microsoft has also expressed interest. In fact, Microsoft has already confirmed that the feature will land on the Xbox Series X|S consoles in the near future.

"At Xbox, we’re excited by the potential of AMD’s FidelityFX Super Resolution technology as another great method for developers to increase framerates and resolution. We will have more to share on this soon," Microsoft told IGN.

This isn’t altogether a surprising development, as Microsoft and AMD recently announced that Microsoft would be building AMD’s FidelityFX tools into Xbox development kits. FidelityFX tools for Xbox makes sense, especially as Microsoft is looking for more parity between its consoles and PC titles.      

Source: https://www.ign.com/articles/xbox-series-x-s-amd-fidelity-fx-super-resolution-support-confirmed

15:29 | Colorful & Nvidia Open GPU History Museum

Colorful has announced that it has partnered with Nvidia to open the first ever GPU history museum. Colorful announced its new museum following the news that the company had relocated to Shenzhen New Generation Industrial Park, in China. Unfortunately, given that it’s largely headed-up by Colorful, the museum is lacking significant chunks of history. GPU history is rich and complex, and if you’re looking for another museum that tackles it in a bit more depth, we’d recommend the Computer History Museum in Mountain View, California.

That said, the Shenzhen museum itself is interesting. The museum attempts to showcase the early origins of the GPU, dating back to the 1980s, and illustrates how it has evolved into the current generation of video cards. Presumably, Colorful will continue to update the exhibits as new cards and microarchitectures emerge.

Colorful highlights how GPUs became mainstream in the early ‘90s with the VooDoo add-in card from 3Dfx. There’s also Nvidia’s first graphics product in the form of the NV1, and the later introduction of the GeForce brand when Nvidia debuted the GeForce 256. 

There’s also some more rare and obscure GPUs in the mix as well, such as ATI’s (pre-AMD) Rage Fury MAXX. The Rage Fury MAXX was the enthusiast culmination of ATI’s Rage 128 Pro chip. This card packaged two Rage 128 Pro chips on a single PCB, and was ATI’s first dual-GPU card. 

The museum also houses GPUs from companies such as IBM, 3Dlabs, Intel, S3, Trident, 3Dfx, and more. There’s also a section dedicated to the history of Colorful, pointing out how the company was founded in 1995 and entered the video card market in 1999 with its own VooDoo 2 graphics card.    

Source: https://www.techspot.com/news/90028-colorful-nvidia-opening-world-first-gpu-museum.html

17:35 | Noctua’s Fanless CPU Tower Is Alive

Noctua first teased its fanless CPU cooler prototype way back at Computex 2019 -- which we covered here -- with no official product name and a loose release window targeting 1H '2020. Due to manufacturing challenges, that release got pushed back considerably. In our Computex 2019 coverage, we detailed some of the challenges Noctua was facing with mass producing its passive CPU cooler, such as Noctua’s initial goal of moving up to a 1.5mm fin thickness.

Recently, the highly anticipated cooler showed signs of life last month, after Noctua responded to a tweet and stated the cooler was coming “very soon.” Even more recently, an early -- and since removed -- listing on Newegg showed the cooler, replete with a product name and price. According to the Newegg listing, Noctua’s passive cooler will be known as the Noctua NH-P1, and will retail for $100. At Computex 2019, we noted that Noctua was hoping to target a $100 (or less) price tag for its passive CPU cooler.

The Newegg listing didn’t mention a whole lot of details. Notably absent were product dimensions and fin thickness. The listing did mention the use of Noctua’s SecuFirm2+ mounting system, with the expected compatibility for most modern CPU sockets (LGA115x, LGA1200 AM4, etc.) and the use of Noctua’s NT-H2 TIM. 

This listing also recommended using the NH-P1 with CPUs that have a “low to moderate heat dissipation.” Listed examples included Intel’s i7-9700K and AMD’s R7 3700X. In response to the early listing, Noctua told PCMag that “This was not intended, but the official launch will be very soon!”  

Source: We have one, but also: https://www.reddit.com/r/hardware/comments/ntz0g8/newegg_sales_page_for_noctua_nhp1_passive_cpu/

19:30 | SK Hynix Denies Claims of Excessive DRAM Manufacturing Flaws

In a weird turn of events, and seemingly out of nowhere, SK Hynix has denied claims that it shipped more than 240,000 wafers containing defective memory. While SK Hynix does admit that some of its recent DRAM products have suffered defects, it also states that rumors of the wafers’ demise have been greatly exaggerated. 

“We can confirm that a defect was found in a few DRAM products,” SK Hynix told The Register in a statement. “We're currently talking to a limited number of customers affected by this to address the issue. While it’s too early to estimate the potential losses, we don't think they would be that significant as the defect is within the range of typical quality issue check,” SK Hynix said. 

It’s hard to trace just where the claim came from, but according to Yonhap News, the rumor has been circulating since at least Monday, June 7th. Furthermore, it seems SK Hynix has opened a police investigation into the matter, on grounds of libel and false accusations.   

Source: https://www.theregister.com/2021/06/09/sk_hynix_dram_defects/

20:51 | SK Hynix Offers Small Update on HBM3

Also in SK Hynix news, there’s a very brief update on the company’s HBM3 development. According to an updated HBM2e product page, SK Hynix now lists HBM3 as “under development.”

“SK hynix leads the HBM market with ambitions for even faster HBM solutions: Our HBM3, under development, will be capable of processing more than 665GB of data per second at 5.2Gbps in I/O speed,” reads the product description for SK Hynix’s upcoming HBM3.    

That’s up from SK Hynix's HBM2E, which tops out at 460 GB/s of bandwidth. Obviously, this has big bandwidth implications for GPUs and devices that use multiple stacks of HBM, as the aggregate bandwidth goes up with every stack. We won’t bother with any speculative math here, as JEDEC has yet to finalize the HBM3 standard, and finalized specs and products could certainly change.

SK Hynix didn’t disclose any release date for its HBM3, so we’ll have to keep waiting. 

Source: https://product.skhynix.com/products/dram/hbm/hbm2e.go

https://www.tomshardware.com/news/hbm3-to-top-665-gbps-bandwidth-per-chip-sk-hynix-says

21:37 | Intel Updates Speculative Execution Terminology

Intel is trotting out what it’s calling a “unified terminology” for speculative execution attacks and vulnerabilities. In its documentation, Intel cites the need for more precise terminology, given the rise of phrases such as “Spectre-like” and “Meltdown-like.” Additionally, Intel notes that certain terms are often used interchangeably -- such as side-channel and covert-channel. To give Intel credit, it is correct in that a lot of language around these vulnerabilities and attacks have become somewhat vague and ambiguous. That said, we're not security researchers, so we can’t fully say as to whether or not Intel’s approach here is the best. 

Moving forward, Intel is now referring to speculative execution side channel vulnerabilities as transient execution attacks, and will also refer to specific types of transient execution attacks. Intel notes that this refined terminology will be what the company adopts and uses for all of its future guidance. It should be noted that while Intel’s use of the term is new, the term itself isn’t. In many security research circles and research papers, transient execution attack(s) seems to be the preferred term.

Intel goes onto describe a number of different transient execution attacks: 

“A transient execution attack is characterized by the relationship between the disclosure gadget and the victim/adversary domains: a transient execution attack is domain-bypass if the disclosure gadget is in the adversary’s domain, cross-domain if the disclosure gadget is in the victim’s domain, or in-domain if the disclosure gadget is in an adversary-controlled sandbox within the victim’s domain,” Intel explains. 

Intel states that the use of this terminology will help both security researchers and software developers.     

Source: https://software.intel.com/content/www/us/en/develop/articles/software-security-guidance/best-practices/refined-speculative-execution-terminology.html

23:40 | Tesla Using AMD GPUs for “PS5-Level Gaming”

Whilst unveiling the Model S Plaid, which slots in between the Model S Flannel and the Model S Tartan in Tesla’s line-up, Musk took the opportunity to talk about the “PS5-level performance” of the entertainment system in the new Tesla. This was demonstrated with someone inside the Model S Plaid playing games on the onboard entertainment system. Tesla chose Cyberpunk 2077 to show off the power of the system. Clearly, this is all Musk’s master plan attempt to cash-in on Cyberpunk’s long absence from the Playstation store, since no one can buy consoles or GPUs aggrieved Playstation owners can get their kicks in Night City on the new Tesla. Cyberpunk 2077 is a great example for this, because driving and cars in that game are so well-designed. We’ve included some footage of exclusive leaks of the real Tesla Plaid’s self-driving capabilities that GN received from an unidentified leaker. Musk didn’t mention whether or not spontaneous T-posing would be a problem in the new Teslas.

Musk didn’t go into details on the hardware itself, but during the Computex keynote, Dr. Lisa Su revealed that AMD RDNA2 would appear in new Tesla Model S and Model X cars. Given that AMD designed the silicon powering the Playstation 5 (and Xbox Series Whatever) Musk’s choice of performance comparison is probably pretty accurate.

Source: 23:16 in Tesla Event - https://youtu.be/kUmkbzQ-BS0?t=1396

25:13 | Seagate Launches Cyberpunk Drive

Seagate is launching a limited edition Cyberpunk 2077 themed SSD. These 1TB SDDs feature a massive, yellow aluminum heatsink, and production will be limited to a groan-worthy 2077 drives. We’re generally fans of products that try to break free of red-and-black-ARGB-GAMER aesthetic that most companies use; if you don’t believe me, see our reviews of Yeston’s GPUs and & the CutePet Case. This SSD however looks a bit like a tanning booth with its hollowed out heatsink and eerie blue-ish light. 

More importantly, we’re skeptical about a product branded with another product that was infamous for how buggy it is -- if somehow you’ve forgotten this or didn’t know, the editors have kindly been playing footage of just the bugs we encountered in testing throughout this item. Luckily though, Seagate includes data recovery services with the device, which they say protects you against “water damage or natural disaster.”

Source: https://www.seagate.com/products/gaming-drives/pc-gaming/firecuda-520-cyberpunk-2077-ssd/

Host, Copy Editing: Steve Burke
Editorial: Eric Hamilton
Video, Writing: Keegan Gallick