BTC $79,044.39 -0.73%
ETH $2,484.23 +0.20%
BNB $741.23 -0.38%
XRP $1.39 -1.26%
SOL $104.16 -1.57%
TRX $0.3342 -0.27%
DOGE $0.0898 +1.33%
ADA $0.2191 +0.58%
BCH $261.30 +2.38%
LINK $12.96 +6.39%
HYPE $85.43 -3.92%
AAVE $131.71 -1.60%
SUI $0.8181 +3.82%
XLM $0.1904 +3.71%
ZEC $1,162.01 -0.25%
BTC $79,044.39 -0.73%
ETH $2,484.23 +0.20%
BNB $741.23 -0.38%
XRP $1.39 -1.26%
SOL $104.16 -1.57%
TRX $0.3342 -0.27%
DOGE $0.0898 +1.33%
ADA $0.2191 +0.58%
BCH $261.30 +2.38%
LINK $12.96 +6.39%
HYPE $85.43 -3.92%
AAVE $131.71 -1.60%
SUI $0.8181 +3.82%
XLM $0.1904 +3.71%
ZEC $1,162.01 -0.25%

tpu

All
Article
Flash

first_img Google disassembles retired servers to recycle DDR4 in response to memory shortages

Google's Senior Director of Supply Chain Infrastructure, Nikhil Cherian, revealed that to overcome memory bottlenecks, Google is developing software and hardware solutions and dismantling retired servers to recycle DDR4 components to establish an internal recycling supply chain. Google has designed special hardware adapters to connect the previous generation DDR4 to the new generation of AI servers while importing retired servers to remove their DDR4 modules for recycling.Cherian stated that the AI industry has rapidly shifted from being compute-constrained to memory-constrained, with high-performance memory accounting for about 75% of the bill of materials cost for a given AI server. A Goldman Sachs report indicated that memory prices will continue to rise in the third quarter, with personal computer DRAM prices expected to increase by 18% to 23% and server DRAM prices expected to rise by 13% to 18%. Trendforce data shows that in August, the spot market prices for DDR4 8GB and DDR5 8GB rose to $142 and $133, respectively.The two TPU ASICs launched by Google this year have been optimized for memory design, claiming to reduce memory demand to one-sixth of the original. The TPU8i chip features a dedicated layered memory design that relies on a high-speed DDR5 memory architecture to perform host-level tasks, with each chip equipped with 288GB of HBM3e high-bandwidth memory.

first_img Google claims that the cost of AI server memory has exceeded 75%, promoting a dual-track strategy for software and hardware

The SEMICON Taiwan 2026 Memory Summit took place on the 1st, where Nikhil Cherian, Senior Director of Supply Chain Infrastructure at Google Cloud under Alphabet, pointed out that with the popularity of multimodal and mixed expert architectures, AI computation has shifted from being power-limited to memory-limited, with high-performance memory accounting for over 75% of the cost of AI server hardware bill of materials. In the face of capacity, bandwidth, and power consumption bottlenecks, Google is breaking through the AI memory bottleneck through a dual-track strategy of hardware offloading for inference and training, and lossless quantization software algorithms.Google adopts an offloading strategy in hardware architecture, launching TPU 8i for low-latency inference and TPU 8t specialized for large-scale training. The TPU 8i is equipped with 288 GB of high-bandwidth memory, with SRAM capacity on the chip increased threefold to 384 MiB, placing dynamic conversation states and key-value caches on the chip itself to achieve zero chip-off latency. The TPU 8t forms a super-large computing cluster with 9600 chips, achieving a shared pool of HBM at a scale of 2 PB, eliminating chip-off data transfer bottlenecks, along with TPU Direct Storage technology.Google has developed the training-free TurboQuant lossless quantization algorithm, compressing the key-value cache of large models from 32 bits to 3 bits, reducing memory usage by six times without loss of accuracy, resulting in an eightfold acceleration in attention computation, and integrating old-generation DRAM technology to extend the lifecycle of components.

first_img Rising Sun Technology Holdings receives additional advanced packaging and testing orders from Google TPU and Intel

According to a report by the Economic Daily, ASE Technology Holding benefited from the increased capacity of Google TPU, significantly adding advanced packaging and testing orders. Intel is accelerating the promotion of its EMIB advanced packaging platform while simultaneously expanding its outsourcing orders to ASE Technology Holding, with the previously raised testing quotes showing benefits. Google continues to increase its AI capital expenditures to expand data centers, using TPU to support the Gemini large model, AI agents, and Google Cloud services. The market estimates that Google TPU will enter a large-scale production phase by 2028, with shipments expected to reach 12 to 15 million units. ASE Technology Holding has mastered 2.5D, 3D advanced packaging, and high-end testing technologies, and Google is also expanding its use of TSMC's advanced processes and CoWoS capacity.Intel is racing to advance its EMIB packaging platform, with major cloud companies like Meta and Google successively adopting the EMIB and EMIB-T platforms. ASE Technology Holding, positioned as a pure testing and packaging foundry, can directly procure organic embedded silicon bridge substrates to undertake backend wafer-level assembly and high-end testing services. Wu Tianyu, the Chief Operating Officer of ASE Technology Holding, believes that Intel's EMIB and TSMC's CoWoS are not zero-sum competitors. As the testing time for AI chips lengthens, the costs of equipment, materials, and labor are rising, and high-end packaging and testing capacity is becoming increasingly tight. Major testing and packaging factories have gradually raised prices based on products, capacity, and customer conditions, with increases of about 5% to 10%.

hot_img SemiAnalysis: Gemini has exited the frontier competition, and GCP is accelerating the sale of TPUs to third parties for profit

The research organization SemiAnalysis released an analysis indicating that Google DeepMind is no longer among the leading AI laboratories. A week prior, DeepMind co-founder Demis Hassabis stepped back from daily operations, and key members such as Google Chief Scientist Jeff Dean and Gemini co-lead Oriol Vinyals left to establish a new lab called Discovery Loop. The analysis suggests that the long-term struggle within Google over computing power allocation between Gemini and GCP has concluded with GCP emerging victorious.SemiAnalysis stated that Gemini 3.5 Pro has been canceled, and Gemini 3.6 Flash's performance is inferior to that of leading Chinese open-source models and Grok 4.5. Currently, Gemini has fallen to the 8th or 9th position in the large model rankings. Meanwhile, GCP is selling a large number of TPUs to competitors like Anthropic, having secured long-term leasing and sales contracts for hundreds of thousands of TPUs over the past nine months. The Tokenomics model estimates that Gemini's own ARR is about $12 billion, while GCP's third-party AI cloud service revenue is expected to exceed $73 billion by the end of 2027, with TPU system sales contributing an additional over $120 billion. GCP's latest quarterly growth rate is 82%, and it is expected to accelerate to over 100% by 2027 due to TPU system sales, contributing approximately $3 to Google's earnings per share.

$50 million increase in budget, zero debt milestone, and 100% output retention

According to BBX data, global listed companies made key progress in "health optimization" of their treasury structures yesterday:$50 million addition: Boyaa Interactive (0434.HK) announced yesterday that the board has approved an additional budget of $50 million for increasing its holdings in Bitcoin and Ethereum. So far, the group has invested over $250 million, aiming to establish itself as the largest digital asset treasury entity in Asia.Officially entering "debt-free holding": TeraWulf (NASDAQ: $WULF) announced yesterday that it has used cash flow generated from its high-performance computing business to pay off the last of its high-interest debt. As one of the most energy-efficient mining companies globally, the company reiterated that it has entered the "100% output retention" phase as of yesterday.Computing power as net reserves: HIVE Digital (TSX-V: $HIVE) disclosed yesterday that the profits from its AI data center in Sweden have fully covered global operating expenses for two consecutive months. This means that all Bitcoin produced yesterday is recorded as "net reserves," with no need for any market sell-off.$10 million procurement settled: Acurx Pharmaceuticals (NASDAQ: $ACXP) confirmed yesterday that its first $10 million strategic procurement of Bitcoin has been fully completed. As a biopharmaceutical company, ACXP plans to use this asset as a "value anchor" for its clinical research and development funding over the next five years.Automated treasury upgrade: Public.com launched an "automated treasury balancing" suite for small and medium-sized enterprises globally yesterday, supporting the automatic proportional conversion of idle corporate funds into BTC, with the first-day subscription scale exceeding $80 million.

$150 million mixed purchase plan with 92% output retention: Global corporate treasury moves towards "high-frequency assetization"

According to BBX data, yesterday global listed companies showed a trend of transitioning from "opportunistic buying" to "standardized automatic execution" in their crypto treasury allocations:$150 million allocation: Robinhood (NASDAQ: $HOOD) board approved a $150 million BTC/ETH mixed purchase plan today. The company clearly stated that it will leverage the liquidity advantage of its proprietary trading platform to complete the position within the next 30 days, aiming to optimize the long-term purchasing power of the company's idle cash.92% output retention: Iris Energy (NASDAQ: $IREN) disclosed its operational data for the first quarter of 2026, revealing that its Bitcoin output retention rate reached 92% yesterday, and announced plans to acquire an additional 800 BTC using renewable energy power premiums, aiming to align the company's asset structure more closely with "hard assets."40% net asset ratio: Meitu (1357.HK) financial report shows that the fair value of its held crypto assets has accounted for 40% of its total net assets. The company announced that it will continue to implement the "profit-to-coin" plan, continuously investing 20% of its annual business net profit into Bitcoin reserves.300 BTC energy arbitrage: Aker ASA (Seetee) disclosed that through its industrial energy arbitrage project in Norway, it converted surplus electricity into a holding of 300 BTC again yesterday. This giant is continuously diluting its fiat currency debt risk through a closed loop of "energy-computing-assets."$5 million standard configuration: Wolfspeed (NYSE: $WOLF) board passed a resolution to approve $5 million in Bitcoin as a long-term treasury reserve for the first time. This further validates that the semiconductor and high-tech manufacturing industries are beginning to view BTC as a "standard component" for financial health.
app_icon
ChainCatcher Building the Web3 world with innovations.