Another big computing power AI chip released! It's 3 times faster than B200 and has just raised 2.4 billion dollars.

On February 24, Intel announced that it had provided U.S.AI chipUnicorn SambaNova's overUS$350 million (about RMB 2.4 billion)Series E strategic financing injection for an undisclosed amount.

Meanwhile, SambaNova introduces its fifth-generation AI chipSN50No. It's called “The only chip that delivers the speed and throughput needed for intelligent body AI”, with top speeds of comparable chip5 timesThe single-model parameter scale supported by multi-core interconnections is up to10 trillionThe context length is up to10 million tokens.

According to SambaNova's disclosure, compared to the Blackwell B200 GPU, the SN50's maximum speed is its5 times, the throughput of an intelligent's reasoning is its3 timesabove, which is well represented on a range of models such as Meta's Llama 3.3 70B.

According to foreign media reports, prior to this, Intel had discussed the idea of taking a loan for aboutUS$1.6 billion (about RMB 11 billion)acquisition of SambaNova, but negotiations eventually broke down. Neither party has ever responded to the matter, however.

Founded in 2017 by a number of Stanford professors, SambaNova is valued at $5 billion (roughly Rs. 34.4 billion) after a 2021 funding round. Its chairman is Intel CEO Lip-Bu Tan. Intel has invested in SambaNova several times.

In its latest announcement, SambaNova and Intel announced a multi-year strategic partnership program designed to deliver high-performance, cost-effective AI inference solutions to create the next-generation heterogeneous AI datacenter, integrating Intel Xeon processors, Intel GPUs, Intel networking and storage, and SambaNova systems to unlock the multi-billion dollar inference market opportunity.

After the news was announced, Intel's shares rose more than 5% on Tuesday EST.

First, 5 times the computing power, 4 times the network bandwidth, can support 10 trillion parameters of large models

The SN50 chip is based on SambaNova's Reconfigurable Data Flow Unit (RDU) architecture, with ultra-low latency for real-time response, and is capable of supportingseveral thousandconcurrent AI sessions and reduces the cost per token through higher hardware utilization.

The arithmetic power per chip is increased to the fourth-generation SN40L's5 timesThe network bandwidth was increased to4x.

The SambaRack SN50 will16 piecesTogether, the SN50 chipset can run up to10 trillion parameters,10 million tokensof oversized models.

The interconnected SambaRacks are available through theTB per secondinterconnect speeds of up to256 blockschips, thereby reducing the first token generation time and supporting larger batch sizes, so models with higher throughput and responsiveness can be deployed.

SambaRack's power consumption averages only20kWthat can run in existing air-cooled data centers. This brings total cost of ownership (TCO) benefits to inference service providers running models such as gpt-oss, with energy savings of the B200 GPU's8 times.

SN50 will be available onSecond half of 2026Shipments begin. SoftBank Group will be the first customer to deploy SN50 in a next-generation AI data center in Japan.

Two, based on data flow, three-layer memory architecture, faster and more energy efficient when running large models

The SambaNova team believes that intelligences require intelligent, predictive, and resilient infrastructure, and that to achieve viable intelligences, hardware must be able to instantly adapt to sudden workloads and switch between expert models without latency.

GPUs excel at AI model training, but AI inference is a data movement and memory optimization challenge that requires a different architectural approach.

To perform AI inference, the GPU must make multiple redundant calls to off-chip memory, and each memory call increases processing latency and consumes energy, which can lead to high power consumption issues.

The standard approach to deploying multiple models on the GPU is to load the models into High Bandwidth Memory (HBM), but GPU HBM resources are scarce and costly.

When a workload requires an unloaded model, the system must unload the current model and acquire a new one, a process typically measured in seconds. Even with vLLM's level 1 sleep mode, waking up a small model takes 0.1 to 0.8 seconds.

For the large inference models required by the intelligentsia, this wake-up time causes a delay of 3 to 6 seconds. For AI intelligences executing a 10-step reasoning process involving 5 different models, these delays can accumulate up to 30 seconds, rendering the real-time intelligences workflow unusable.

SambaNova's RDU is designed to solve this problem.

RDU maps the computational graph of a given AI model to the most efficient path for data transfer on the processor. This approach eliminates redundant memory calls and can significantly reduce latency and power consumption.

Unlike GPUs that are limited to the HBM capacity of a single card, the SN50 RDU utilizes a uniqueThree-Layer Memory ArchitectureThe combination of theOn-chip SRAM, massive HBM and ultra-high-speed SRAM.

This hierarchical structure allows the chip to carry the largest model, supporting themillisecondRun and switch multiple models in time.

In addition, with SN50, input tokens can be cached in memory, thus reducing pre-population processing time and the first token fetch time (TTFT) of a request.

Combined with these advantages, its memory architecture is ideally suited as a cache for intelligences that can process tasks more efficiently.

SambaNova has also introduced configurable model packs on SambaStack powered by RDUs that are faster to switch to than traditional GPU architectures and inference frameworks such as vLLM.

Third, join forces with Intel to accelerate the launch of the AI cloud platform

As part of a multi-year strategic partnership, Intel plans to make a strategic investment in SambaNova to accelerate the launch of an AI cloud platform based on Intel technology.

The collaboration is expected to cover three key areas:

(1) AI Cloud Extension: SambaNova is expanding its vertically integrated AI cloud platform built on Intel Xeon processors and optimized for large language models and multimodal models. The platform will deliver low-latency, high-throughput AI services supported by reference architectures, deployment blueprints, and partnerships with system integrators and software vendors.

(2) Integration of AI infrastructure: Combines SambaNova's systems with Intel's CPU, gas pedal, and networking technologies to provide scalable, production-ready reasoning capabilities for inference, code generation, multimodal applications, and intelligent body workflows.

(3) Marketing Execution: Co-selling and co-marketing through Intel's global enterprise, cloud and partner channels to accelerate the adoption of the AI ecosystem.

According to Intel's disclosure, the partnership complements Intel's existing datacenter GPU investments and does not change its competitive strategy in AI, and Intel will continue to increase its investments in GPU intellectual property, architectures, products, software, and systems, as well as strengthen its strategic roadmap for AI from the edge to the cloud.

Conclusion: AI reasoning market sees new combinations

With AI reasoning booming, tech companies are looking for AI hardware infrastructure solutions that are faster and more efficient, and less costly for enterprise-level AI deployments.

Many cloud-based AI chip creators are focusing the selling point of their own products on high energy efficiency, both in response to the trend of high demand for AI reasoning and to avoid the sharp edge of NVIDIA's strength in AI training.

Through the partnership, SambaNova can leverage Intel's global reach to scale its AI processors, and Intel enhances its overall strength in AI reasoning. This will provide a competitive option for the increasingly diverse AI inference market.

Article source: Wisdom