Supercomputer-on-a-chip goes live: single PCIe card packs more than 6k cores
InspireSemi introduces Thunderbird I Accelerated Computing chip with 1,536 custom RISC-V CPU cores for scientific and data processing. Energy-efficient, scalable to 360,000 cores, suitable for high-performance computing tasks. CEO praises team's work.
Read original article
InspireSemi has unveiled the Thunderbird I Accelerated Computing chip, featuring 1,536 custom 64-bit RISC-V CPU cores designed for scientific computing and data processing. The chip integrates a high-speed mesh network fabric for efficient communication among cores, crucial for synchronized operations. The upcoming product release will include a PCIe add-in card hosting over 6,000 interconnected CPU cores, suitable for high-performance computing tasks like climate science and medical research. InspireSemi emphasizes the chip's energy efficiency, initially intended for blockchain applications. The company plans to start customer deliveries in the fourth quarter but has not disclosed pricing details yet. Thunderbird I aims to provide a more environmentally friendly alternative to traditional data center GPUs. CEO Ron Van Dell commended the engineering team for completing the chip's design and initiating production with partners like TSMC. The chip's scalability to over 360,000 cores positions it as a powerful solution for various compute-intensive applications, including AI, machine learning, and graph analytics.
Related
TSMC experimenting with rectangular wafers vs. round for more chips per wafer
TSMC is developing an advanced chip packaging method to address AI-driven demand for computing power. Intel and Samsung are also exploring similar approaches to boost semiconductor capabilities amid the AI boom.
Intel's Gaudi 3 will cost half the price of Nvidia's H100
Intel's Gaudi 3 AI processor is priced at $15,650, half of Nvidia's H100. Intel aims to compete in the AI market dominated by Nvidia, facing challenges from cloud providers' custom AI processors.
First 128TB SSDs will launch in the coming months
Phison's Pascari brand plans to release 128TB SSDs, competing with Samsung, Solidigm, and Kioxia. These SSDs target high-performance computing, AI, and data centers, with larger models expected soon. The X200 PCIe Gen5 Enterprise SSDs with CoXProcessor CPU architecture aim to meet the rising demand for storage solutions amidst increasing data volumes and generative AI integration, addressing businesses' data management challenges effectively.
Testing AMD's Giant MI300X
AMD introduces Radeon Instinct MI300X to challenge NVIDIA in GPU compute market. MI300X features chiplet setup, Infinity Cache, CDNA 3 architecture, competitive performance against NVIDIA's H100, and excels in local memory bandwidth tests.
Intel CPU with Optical OCI Chiplet Demoed with 4Tbps of Bandwidth and 100M Reach
Intel unveiled a new CPU with an Optical Compute Interconnect chiplet, offering 4Tbps bandwidth and 100m reach. The chiplet merges Photonic and Electronic Integrated Circuits, featuring built-in lasers for improved server communication, especially in AI servers. Intel plans to scale the technology up to 32Tbps, focusing on power savings and enhanced bandwidth. This advancement aligns with the industry's move towards co-packaged optics for better efficiency and performance, marking a significant shift in server architecture towards power efficiency and high-speed data transfer.
8 commentsFrom the (air quotes) article...
> The chip's architecture integrates a high-speed mesh network fabric that provides substantial bandwidth and minimal latency communication among cores, important for applications that rely on synchronized operations across multiple threads. This efficient network integration manages interactions within the chip's core array and memory systems, ensuring optimal performance without the common bottlenecks.
If anyone can tell me what that means, as it relates to the reality we live in, I'm all ears.
Also from the 'article' as it were...
>InspireSemi also stresses Thunderbird I’s energy efficiency, a carryover from its initial design for energy-sensitive blockchain computing applications.
It's an AI Pump and Dump. It all makes sense now.
> Raw performance: 24 TFLOPS (FP64)
> Energy efficiency: 50 GFLOPS/Watt (FP64)
[1] (If you try to select body text, the page reloads.)
An RTX 4090 does 1.3 TFLOPS (FP64) at TDP 450 W [2], so 3 GFLOPS/W. If so, this is substantially better, and built on a core that can use standard tooling. [edited]
They don't say how much money you can cram onto the board, or what the memory bandwidth is; just that it's a lot...
1. https://inspiresemi.com/#solutions
2. https://www.techpowerup.com/gpu-specs/geforce-rtx-4090.c3889
> Ron Van Dell, CEO of InspireSemi
The Dell link goes to their Dell tag page, where I doubt you'd find anything about Ron Van.
A little bit of creativity in naming can go a long way.
Related
TSMC experimenting with rectangular wafers vs. round for more chips per wafer
TSMC is developing an advanced chip packaging method to address AI-driven demand for computing power. Intel and Samsung are also exploring similar approaches to boost semiconductor capabilities amid the AI boom.
Intel's Gaudi 3 will cost half the price of Nvidia's H100
Intel's Gaudi 3 AI processor is priced at $15,650, half of Nvidia's H100. Intel aims to compete in the AI market dominated by Nvidia, facing challenges from cloud providers' custom AI processors.
First 128TB SSDs will launch in the coming months
Phison's Pascari brand plans to release 128TB SSDs, competing with Samsung, Solidigm, and Kioxia. These SSDs target high-performance computing, AI, and data centers, with larger models expected soon. The X200 PCIe Gen5 Enterprise SSDs with CoXProcessor CPU architecture aim to meet the rising demand for storage solutions amidst increasing data volumes and generative AI integration, addressing businesses' data management challenges effectively.
Testing AMD's Giant MI300X
AMD introduces Radeon Instinct MI300X to challenge NVIDIA in GPU compute market. MI300X features chiplet setup, Infinity Cache, CDNA 3 architecture, competitive performance against NVIDIA's H100, and excels in local memory bandwidth tests.
Intel CPU with Optical OCI Chiplet Demoed with 4Tbps of Bandwidth and 100M Reach
Intel unveiled a new CPU with an Optical Compute Interconnect chiplet, offering 4Tbps bandwidth and 100m reach. The chiplet merges Photonic and Electronic Integrated Circuits, featuring built-in lasers for improved server communication, especially in AI servers. Intel plans to scale the technology up to 32Tbps, focusing on power savings and enhanced bandwidth. This advancement aligns with the industry's move towards co-packaged optics for better efficiency and performance, marking a significant shift in server architecture towards power efficiency and high-speed data transfer.