On June 30th, the Netherlands introduced new export restrictions on advanced semiconductor manufacturing equipment. Despite facing export controls from the US, Japan, and the Netherlands, TrendForce anticipates the market share of Chinese foundries in terms of 12-inch wafer production capacity will likely increase from 24% in 2022 to an estimated 26% in 2026. Moreover, if the exports of 40/28 nm equipment eventually receive approval, there's a chance that this market share could expand even further, possibly reaching 28% by 2026. This growth potential should not be dismissed.

Several manufacturing processes including photolithography, deposition, and epitaxy will be subject to these recent export restrictions. Beginning September 1st, the export of all controlled items will require formal authorization. TrendForce reports that Chinese foundries have been primarily developing mature processes like 55 nm, 40 nm, and 28 nm. Furthermore, demand for deposition equipment can be largely met by local Chinese vendors, meaning concerns regarding expansion and development are minimal. The main limiting factor, however, remains the equipment used in photolithography.

Leaked photos of a cinder block-sized NVIDIA flagship graphics card cooler appeared online over a week ago, with speculation pointing to it originating from an extremely powerful RTX-40 series GPU—perhaps a theoretical GeForce RTX 4090 Ti or something codenamed TITAN ADA. The pictured prototype outsizes several existing reference designs—its substantial bulk could be enough to tame the fully unlocked potential of Team Green's already large AD102.

Last week's photos have been traced back to the source—as reported by Wccftech, it seems that a seller on the Chinese Taobao Goofish platform is attempting to flog the unit for roughly $122,750 (888,888 RMB). The seller/site member "Hayaka" is apparently open to accepting offers from the highest bidder, but the prospective buyer will not be getting their hands on any working hardware—the listing is for the cooler alone. No GPU or PCB is included according to the provided information, so the winner will be procuring a very expensive (albeit highly unique) mantelpiece.

ASUS Chinese GM—Tony Yu—has shown off a graphics card concept on Bilibili that has a rather unusual feature, a slot for an M.2 NVMe SSD. The card is based on NVIDIA's GeForce RT 4060 Ti GPU and although not all details are clear at this point in time, but ASUS is taking advantage of the unused PCIe lanes on the card, since the AD106 GPU only uses eight PCIe lanes, the PCIe connector on the card has space for a further eight lanes. In theory ASUS could have added a pair of SSDs, since there are a total of eight lanes available, but as this was just a proof of concept, they seemingly stuck with a single SSD.

It's unclear if ASUS relies on bifurcation or if the company has added some kind of bridge chip, but bifurcation makes more sense, as a bridge chip would add a lot more cost. The neat thing with the NVMe drive being on the GPU, is that it also connects to the heatsink of the graphics card, which means the cooling should be rather good. However, for this to work properly, the SSD would have to be mounted back to front compared to how it would be mounted on a motherboard. Based on the test results, the SSD runs at a cool 42 degrees C, even when the GPU is being stress tested. It's likely that this product will not make it to markets outside of China, if it's ever launched into retail.

A group of researchers in China have used a new approach to AI to create a full RISC-V processor from scratch. The team set out to answer the question of whether an AI could design an entire processor on its own without human intervention. While AI design tools do already exist and are used for complex circuit design and validation today, they are generally limited in use and scope. The key improvements shown in this approach over traditional or AI assisted logic design are the automated capabilities, as well as its speed. The traditional assistive tools for designing circuits still require many hours of manual programming and validation to design a functional circuit. Even for such a simple processor as the one created by the AI, the team claims the design would have taken 1000x as long to be done by humans. The AI was trained by observing specific inputs and outputs of existing CPU designs, with the paper summarizing the approach as such:

(...) a new AI approach, which generates large-scale Boolean function with almost 100% validation accuracy (e.g., > 99.99999999999% as Intel) from only external input-output examples rather than formal programs written by the human. This approach generates the Boolean function represented by a graph structure called Binary Speculation Diagram (BSD), with a theoretical accuracy lower bound by using the Monte Carlo based expansion, and the distance of Boolean functions is used to tackle the intractability.

Intel China has taken to Weibo and Bilibili in a new effort to explain how things will pan out for 14th generation CPU lineups, with emphasis on its new branding and naming conventions for 2023 and beyond. These announcements contain the company's first official acknowledgement of Raptor Lake-U, Raptor Lake-S and Raptor Lake-HX SKUs getting a refresh. Team Blue's branding scheme is set to become even more convoluted with Meteor Lake premiering with "Core Ultra" instead of the old "i" labelling system. The China office's product rundowns indicate that the Raptor Lake Refresh series will be split into Core # and Core i# families, which complicates matters further.

It seems that Raptor Lake-U & H Refresh (mobile) and the entire Meteor Lake lineup will be assuming the new Core # branding scheme, but the latter series will be more performant—hence the adding of Ultra (e.g Ultra Core 5/7/9), so customers can tell the difference between product lines lumped into the same generation! Desktop Raptor Lake-S Refresh and high-end mobile Raptor Lake-HX Refresh processors will retain the company's traditional "Core i" naming convention, this will eventually be retired with the 14th gen family.

Yeston introduced its first Game Ace RTX 40-series custom graphics card yesterday, with a product launch occurring soon after—only for the Chinese market it seems, so prospective Western buyers will have to carry out a few extra steps to acquire one. The GeForce RTX 4060 Ti Game Ace is yet another example of the company producing a white PCB-toting graphics card, albeit in the lower mid-range tier. Yeston has released white boards in the past—most notably for its high-end Sakura Sugar series RTX 4080 and RX 7900 XT models.

The Weibo announcement outlines that the RTX 4060 Ti Game Ace will costs around $432 - 460 (3099 - 3299 RMB), and it offers a simple 2-color lighting system: "the combination of pure white and light blue is fresh and soft, and with the blue LED, it will bring you a little coolness in this hot summer." Yeston's press release was lacking in terms of technical details, but the photos show a very basic looking card that lacks a dual-BIOS switch.

The top technology companies in China have been ordering a lot of NVIDIA enterprise-grade GPUs, even though U.S. sanctions have prevented the shipment of A100 and H100 models (plus AMD's MI250 Instinct accelerator) to the nation in recent times. ByteDance - best known for developing TikTok - managed to grab plenty of Ampere enterprise units prior to last Autumn's cutoff period, and has continued to purchase Team Green's H800 GPU, which is a cut-down version of the H100 flagship. Smaller outfits are relying on less direct sources to acquire HBC GPUs—according to a Reuters investigative article, international trade restrictions have created a thriving black market for "top-end NVIDIA AI chips."

Their reporters carried out some on-site sleuthing: "Visiting the famed Huaqiangbei electronics area in the southern Chinese city of Shenzhen is a good bet - in particular, the SEG Plaza skyscraper whose first 10 floors are crammed with shops selling everything from camera parts to drones. The chips are not advertised but asking discreetly works...They don't come cheap. Two vendors there, who spoke with Reuters in person on condition of anonymity, said they could provide small numbers of A100 artificial intelligence chips made by the U.S. chip designer, pricing them at $20,000 a piece - double the usual price."

TikTok developer ByteDance, and other major Chinese tech firms including Tencent, Alibaba and Baidu are reported (by local media) to be snapping up lots of NVIDIA HPC GPUs, with even more orders placed this year. ByteDance is alleged to have spent enough on new products in 2023 to match the expenditure of the entire Chinese tech market on similar NVIDIA purchases for FY2022. According to news publication Jitwei, ByteDance has placed orders totaling $1 billion so far this year with Team Green—the report suggests that a mix of A100 and H800 GPU shipments have been sent to the company's mainland data centers.

The older Ampere-based A100 units were likely ordered prior to trade sanctions enforced on China post-August 2022, with further wiggle room allowed—meaning that shipments continued until September. The H800 GPU is a cut-down variant of 2022's flagship "Hopper" H100 model, designed specifically for the Chinese enterprise market—with reduced performance in order to meet export restriction standards. The H800 costs around $10,000 (average sale price per accelerator) according to Tom's Hardware, so it must offer some level of potency at that price. ByteDance has ordered roughly 100,000 units—with an unspecified split between H800 and A100 stock. Despite the development of competing HPC products within China, it seems that the nation's top-flight technology companies are heading directly to NVIDIA to acquire the best-of-the-best and highly mature AI processing hardware.

Zotac has debuted a new flagship NVIDIA model for the Chinese market only - its latest GeForce RTX 4090 24 GB graphics card now arrives as a Prime Gamer Force (PGF) Overclocked edition, which is reported to sit in a tier above the (more familiar) AMP Extreme series. A listing on JD.com has it priced at 17,599 RMB (just under $2500), so this a serious bit of kit. Zotac PGF cards are normally tricked out with the company's best cooling solutions, which allows for greater clock speeds and power—this new model is the first example of an Ada Lovelace GPU getting togged up with Prime Gamer Force livery. The Zotac RTX 4090 PGF OC's cooler design appears to share the same overall shape from the AMP Extreme AIRO variant, but it has been updated with new shroud and backplate designs, four (software controlled) RGB lighting areas and an understated white/black aesthetic.

The RTX 4090 PGF OC edition is cooled by three 11 cm fans and a vapor chamber (akin to the AMP Extreme AIRO's solution). The card is one of the largest available on the market with dimensions of 381 mm (L) x 154 mm (W) x 74 mm (D)—beating the equivalent Gigabyte AORUS Master and its Zotac AMP sibling (by 2.5 cm length-wise). ITHome published their (mandarin language) review of the China market-exclusive flagship Zotac card earlier today. The RTX 4090 PGF appears to have a single 16-pin power connector and a 24+4 phase VRM design, with a maximum TDP rating of 525 W. An on-card BIOS switch can enable a 2625 MHz OC mode, the default clock speed is set at 2580 MHz—fairly typical for a high-end GeForce RTX 4090 model. The PGF card outdoes its AMP Extreme AIRO sibling by a small margin in overclocked mode—it is 45 MHz faster (2625 MHz vs. 2580 MHz).

The Moore Threads MTT S80 gaming-oriented graphics card has been tested mostly by Chinese hardware publications, but Japan's PC Watch has managed to get hold of a sample unit configured with 16 GB GDDR6 (14 Gbps) for evaluation purposes and soon published their findings in a "HotHot REVIEW!" The MTT S80 GPU appears to be based on PowerVR architecture (developed by Imagination Technologies), but official Moore Threads literature boasts that their own Chunxaio design is behind all proceedings with 4096 "MUSA" cores. The GPU's clock speed is set at 1.8 GHz, and maximum compute performance has been measured at 14.2 TFLOPS. A 256-bit memory bus grants a bandwidth transfer rate of 448 GB/s. PC Watch notes that the card's support for PCIe Gen 5 x 16 (offering up to 128 GB/s bandwidth) is quite surprising, given the early nature of this connection standard.

Moore Threads has claimed in the past that their cards support Direct X, but PC Watch has discovered that the S80 does not work with DX12, and their tests also demonstrated significant compatibility issues under DX11—with plenty of system freezes and error messages logged. The reviewer(s) had to downshift in some cases to DX9 game environments, in order to gather reliable/stable data. TPU's GPU-Z utility is shown to have no registration information for the S80, and it cannot read the GPU's clock. PC Watch compared their sample unit to an NVIDIA GeForce GTX 1050 Ti graphics card—the entry level 2016-era GPU managed to best the newer competition in terms of in-game performance and power efficiency.

MetaX a new Chinese GPU developer had physical products on display at the recent International Social Public Safety Products Expo—ITHome noted that the company's first ever offering is the N100 GPGPU, specifically designed for artificial intelligence and video processing tasks. MetaX's Yang Jian, a company co-founder, said that his team's primary focus is on AI and data center applications—presumably due to recent demand presenting a lucrative prospect. China's access to Western-developed hardware has been heavily restricted, so native silicon developers have received heavy investment from the State. A MetaX MXN single-slot low-profile card sporting their N100 GPGPU and onboard HBM2E memory is said to offer compute performance of 160 INT8 TOPS and 80 FP16 TFLOPS.

MetaX is also targeting an entry into the gaming GPU market—ITHome reckons that non-enterprise cards are due in 2025. The company's website outlines an MXG-series for graphics rendering, but the information presented indicates that these products are targeting cloud gaming and data center sectors. The company likely requires more time and experience to develop complicated software that will be necessary for the operation of desktop graphics solutions. A similar Chinese chip startup - Biren Technology - has been working on BR100 and BR104 GPGPUs, in a bid to take on NVIDIA's downgraded Hopper H100 accelerators for the region. Biren's engineering team faced challenging circumstances when sanctions blocked their access to TSMC's foundry, plus the departure of their chief architect, who was alleged to be working on an unreleased range of mainstream/gaming graphics cards.

The United States government has imposed sanctions on companies exporting their goods to China with the aim of limiting the country's technological advancements. This forced many companies to reduce their shipments of the latest technologies; however, according to the latest information from The Wall Street Journal, the Biden administration will allow companies to keep expanding their production capacities in China. As the source notes, quoting statements from government officials, the top semiconductor makers such as Samsung, SK Hynix, and TSMC, all of which have a chip production facility in China, will be allowed to expand the production capacity without any US backlash.

Of course, this does not contradict the plan of a US export-control policy, which the administration plans to continue. Alan Estevez, undersecretary of commerce for industry and security, noted last week in the industry gathering that the US plans to continue these restrictions for another year. Reportedly, all manufacturers of wafer fab equipment (WFE) from the US must acquire an export license from the Department of Commerce before exporting any tools for making either logic of memory chip indented for customers in China. Chipmakers Samsung, SK Hynix, and TSMC all received their licenses to export from October 2022 to October 2023. However, the US government now allows these companies to continue upgrading their Chinese plans beyond the renewed license expiry date of October 2024.

According to the latest report from Reuters, a former Samsung executive was arrested by the South Korean authorities yesterday, being accused of stealing company secrets to build a similar chip production facility in China. The former executive had worked for SK Hynix before joining Samsung, where he was involved in the Samsung Electronics division responsible for semiconductor factories. According to the report, the person planned to build a competing factory 1.5 km from a Samsung chip manufacturing facility in Xian, China. The suspect, who was not identified publically, has a collective of 28 years of experience with the South Korean chip makers.

Interestingly, the suspect also caused financial harm to the company, which the Suwon District Prosecutors' Office estimates to be around 300 billion won ($233 million). Prosecutors have announced the indictment of six additional individuals suspected of involvement in the case, including an employee of an inspection company who is charged with allegedly disclosing the architectural blueprint of Samsung's semiconductor plant. A police official commented, "We will sternly deal with any leakage of our technology abroad and strongly respond to illegal leak of domestic companies' core technologies in semiconductor, automobile and shipbuilding sectors among other."

Zephyr has produced an NVIDIA GeForce custom graphics card that sports a very unique pink printed circuit board—bright and pastel colors have featured on cooling solutions in the past, but this new-ish product presents the first example of a PCB with a tinge of blush. Renowned hardware tipster harukaze5719 broke from his normal delivery of very cold and macho tech on social media, and shared his discovery of Zephyr's GeForce RTX 3060 Ti compact ITX card.

International buyers will be disappointed to learn that the pink Ampere card is a China market exclusive, with the company only offering a limited number of products on JD.com. VideoCardz notes that the card's specifications are not at all special, despite its interesting compact form factor and brightly toned cooling solution design. It is a non-overclocked model based on the older RTX 3060 Ti GDDR6 GPU variant with a standard 1750 MHz boost clock, 8 GB VRAM configuration, and a single 8-pin power connector.

A set of reports published by Niko Partners has calculated that the games market in China generated $45.5 billion in revenue last year. Native games companies (including Tencent and NetEase) accounted for 47% of worldwide mobile games revenue, along with a 39% revenue share of international PC games sales. Domestic figures show that 66% of China's gaming revenue comes from mobile, 31% from PC and 3% from console titles. The company's analysts posit that China persists as the largest global games market, with upward trends pushing growth to a predicted $57 billion of revenue by the year 2027. The gaming population is expected to expand to around 730 million participants by then.

Lisa Hanson, CEO and founder of Niko Partners included her own observations of the region's dominance: "Chinese game companies are growing internationally, and they are making bold investments at higher rates than ever. PC games revenue generated overseas by Chinese owned companies rose by 22% in 2022 and is expected to grow by a 13.8% CAGR through 2027 - which is higher than the domestic growth rate by a significant margin... China's market can be tough for domestic and foreign companies, but the country remains the number one market globally for games revenue and the number of gamers, and cannot be ignored.

Newegg has been running various discount promotions on Ryzen 7000-series desktop processors over the past month, but their latest offer finally affects one of AMD's lower end models. The Ryzen 5 7600X CPU is on offer over at the American e-tailer - by using a time limited promo code (MDSCS2342) when checking out you can get the price down to $209 - resulting in a saving of about $31 (from the normal $240). The Ryzen 5 7600 (non-X) CPU is not part of this Memorial Weekend promotion - it is currently listed at $223, with no active Newegg promo code granting a generous saving.

Both of these Zen 4 (Raphael) Ryzen 5 processors were awarded TPU's Editor's Choice accolade - despite receiving criticism for "high platform costs," so it is nice to see that the 7600X is available for just north of $200 in the United States - this saving can offset the added premium of an AM5 motherboard, DDR5 RAM and necessary CPU cooling solution (no Wraith unit is included in the slimline box). Customers in China are fortunate to get an even better online deal - the Ryzen 5 7600X and 7600 are available for way below MSRP via JD.com, with both CPUs priced at an all-time low of $183 (1299 RMB) each.

PowerLeader, a Chinese computer manufacturer, recently introduced its Powerstar lineup of processors, the first being the PSTAR P3-01105 CPU. Initially, the PSTAR SKU closely resembled the Intel Core i3-10105 10th Generation Comet Lake CPU from Intel. Using a "storm core" microarchitecture, this CPU claims to beat any domestic x86 CPU on performance. While the similarity with the Core i3-10105 CPU was close, the clock speed of 3.7 GHz, core configuration, and a slightly different name were significant indicators of similarity between the two. Thanks to the recent Geekbench 5 benchmark run submission, we now know that the Powerstar CPU is just a rebadged Intel Comet Lake processor. PowerLeader expects the Powerstar units to reach 1.5 million shipments this year, and presumably, more models will be rebadged for an announcement later.

Originally, PowerLeader wanted to ship these designs in pre-assembled systems, so the pricing and availability of standalone PowesStar chips are unknown.

There were rumblings late last week about China's competition regulatory bodies giving Microsoft the go ahead for its proposed takeover of the Activision, Blizzard & King games publishing group. The crowd-sourced content site Seeking Alpha was the first online outlet to break the news last Friday - the author had learned from capital market firm - Dealreporter - that China's State Administration for Market Regulation had granted unconditional approval for the $68.7 billion bid, following the conclusion of a "Phase 3" investigation. The latest judgement arrived only a few days after the EU Commission's approval of the deal.

Microsoft has chosen to make an official announcement about this verdict - and has today released statements to several gaming news outlets, including GamesIndustry.biz and Eurogamer: "China's unconditional clearance of our acquisition of Activision Blizzard follows clearance decisions from jurisdictions such as the European Union and Japan, bringing the total to 37 countries representing more than two billion people. The acquisition combined with our recent commitments to the European Commission will empower consumers worldwide to play more games on more devices." Microsoft's next challenge sits with the US government's Federal Trade Commission (FTC) - a final deliberation is due this August.

Increasing its HPL score from 1.02 Eflop/s in November 2022 to an impressive 1.194 Eflop/s on this list, Frontier was able to improve upon its score after a stagnation between June 2022 and November 2022. Considering exascale was only a goal to aspire to just a few years ago, a roughly 17% increase here is an enormous success. Additionally, Frontier earned a score of 9.95 Eflop/s on the HLP-MxP benchmark, which measures performance for mixed-precision calculation. This is also an increase over the 7.94 EFlop/s that the system achieved on the previous list and nearly 10 times more powerful than the machine's HPL score. Frontier is based on the HPE Cray EX235a architecture and utilizes AMD EPYC 64C 2 GHz processors. It also has 8,699,904 cores and an incredible energy efficiency rating of 52.59 Gflops/watt. It also relies on gigabit ethernet for data transfer.

The Chinese Government over the weekend announced restrictions on sales of Micron Technology memory products to Chinese government agencies and infrastructure operators, including the country's telecommunications network infrastructure and data-center service providers. The country's cyberspace regulator conducted a cybersecurity review of Micron Technology products and found that they fail a "network security review."

"The review found that Micron's products have serious network security risks, which pose significant security risks to China's critical information infrastructure supply chain, affecting China's national security," the Cyberspace Administration of China (CAC) said in a statement to Reuters. This move by China is seen as a geopolitical retaliation to the U.S. Government restricting the sales of memory products by YMTC to American PC and ICT companies, a decision that caused the Chinese memory maker to lose major customers such as Apple. This is, however, a partial ban on the Idaho-based memory maker, since China hasn't yet restricted Chinese PC and ICT companies from sourcing Micron Technology products the way Washington banned YMTC.

Chinese company Loongson, specializing in creating processors for usage in mainland China, has been steadily working on enabling its next-generation Loongson 3A6000 CPUs. Aiming to provide the performance level of Intel Willow Cove and AMD Zen 3, these new CPUs will use Loongson's custom LoongArch Instruction Set Architecture (ISA) with a new set of 64-bit superscalar LA664 cores. Today, thanks to the report from Phoronix, we find out that Loongson has submitted some Linux patches that enable the upcoming 3A6000 CPUs to work with Linux-based operating systems at launch. Interestingly, as the new CPU generation gets closer to launch, more Linux kernel patches begin to surface.

Today's kernel patches focus on supporting the hardware page table walker (PTW). As PTW can handle all fast paths of TLBI/TLBL/TLBS/TLBM exceptions by hardware, software only needs to handle slow paths such as page faults. Additionally, in the past, LoongArch utilized "dbar 0" as a complete barrier for all operations. However, this full completion barrier severely impacted performance. As a result, Loongson-3A6000 and subsequent processors have introduced various alternative hints. Loongson plans to ship samples to select customers in the first half of 2023, so we could see more information surfacing soon.

Following the news that Oppo is likely to pull out of the French market, it now appears that the company has closed its chip design unit named Zeku. According to the South China Morning Post, workers at Zeku got less than a day's notice that the company was shutting down, which was on Thursday last week. Employees weren't even allowed back into the office to collect their personal belongings. As recent as two weeks ago, Zeku was still looking for new hires, which makes the abrupt closure even more peculiar, especially as Zeku is said to have employed upwards of 3,000 people.

Considering that Oppo is ranked as the fourth largest smartphone maker in terms of globally shipped units in Q1, the way the company has handled the closure of its chip design unit reeks of desperation. It should be noted that Zeku wasn't designing chips to power Oppo's phones, but presumably Zeku was behind the MariSilicon X, which was Oppo's own imaging co-processor. In all fairness, Oppo is far from the only company to have closed down business units this year, but the company is also said to be suffering in terms of shipments of smartphones, as the company saw a drop of 22 percent in shipments last year, to 103 million units. This appears to be part of the reason behind the shutdown of Zeku, but Oppo hasn't provided much in terms of details behind the closure.

The bipartisan Protecting Circuit Boards and Substrates Act of 2023 introduced by Representatives Blake Moore (R-UT-1) and Anna Eshoo (D-CA-16) finishes the job the CHIPS Act began by incentivizing investment in the domestic printed circuit board (PCB) industry. This bill is a necessary follow-on to the CHIPS Act: without a trusted, reliable domestic source of PCBs and substrates, computer chips don't connect to end use electronic devices.

Domestic PCB production shrunk over the past 20 years, falling from 30% to barely 4% of the world's supply. Ninety percent of the world's supply now comes from Asia…56% in China alone.

TrendForce's latest research indicates that, as production cuts to DRAM and NAND Flash have not kept pace with weakening demand, the ASP of some products is expected to decline further in 2Q23. DRAM prices are projected to fall 13~18%; NAND Flash is expected to fall between 8~13%.

TrendForce reports that the significant drop in DRAM prices was mostly attributed to high inventory levels of DDR4 and LPDDR5 as PC DRAM, server DRAM, and mobile DRAM collectively account for over 85% of DRAM consumption. Meanwhile, the market share for DDR5 remains relatively low.

The recent growth in demand for training Large Language Models (LLMs) like Generative Pre-trained Transformer (GPT) has sparked the interest of many companies to invest in GPU solutions that are used to train these models. However, countries like China have struggled with US sanctions, and NVIDIA has to create custom models that meet US export regulations. Carrying two GPUs, H800 and A800, they represent cut-down versions of the original H100 and A100, respectively. We reported about H800; however, it remained as mysterious as A800 that we are talking about today. Thanks to MyDrivers, we have information that the A800 GPU performance is within 70% of the regular A100.

The regular A100 GPU manages 9.7 TeraFLOPs of FP64, 19.5 TeraFLOPS of FP64 Tensor, and up to 624 BF16/FP16 TeraFLOPS with sparsity. A rough napkin math would suggest that 70% performance of the original (a 30% cut) would equal 6.8 TeraFLOPs of FP64 precision, 13.7 TeraFLOPs of FP64 Tensor, and 437 BF16/FP16 TeraFLOPs with sparsity. MyDrivers notes that A800 can be had for 100,000 Yuan, translating to about 14,462 USD at the time of writing. This is not the most capable GPU that Chinese companies can acquire, as H800 exists. However, we don't have any information about its performance for now.