NVIDIA H100 80GB PCIE 5.0 GPU 高速显卡 计算卡 加速器

NVIDIA H100 80GB PCIE 5.0 GPU 高速显卡 计算卡 加速器

NVIDIA H100 80GB PCIE 5.0 GPU 高速显卡 计算卡 加速器

GPU 最初是纯粹用于图形渲染的设备，但其高度并行的特性也使其对某些计算任务具有吸引力。随着过去几十年 GPU 计算场景的发展，Nvidia 进行了大量投资来占领计算市场。其中一部分涉及认识到计算任务与图形任务有不同的需求，并分散其 GPU 系列以更好地瞄准每个市场。

H100 是 Nvidia 面向计算的 GPU 系列的最新成员。它采用Hopper架构，并建立在一个巨大的814 mm2芯片上，使用台积电的4N工艺和800亿个晶体管。这个巨大的芯片实现了 144 个流式多处理器 (SM)、60 MB 的 L2 缓存和 12 个 512 位 HBM 内存控制器。我们正在 Lambda Cloud 上测试 H100 的 PCIe 版本，该版本支持 114 个 SM、50 MB 的 L2 缓存和 10 个 HBM2 内存控制器。该卡最多可消耗 350 W 的功率。

Nvidia 还提供了 SXM 外形 H100，其功耗高达 700W，并启用了 132 个 SM。SXM H100 还使用 HBM3 内存，提供额外的带宽来满足这些额外的着色器的需要。

NVIDIA H100 80GB PCIE 5.0 GPU 高速显卡 计算卡 加速器

Gpus were originally intended purely as devices for graphics rendering, but their highly parallel nature also makes them attractive for certain computing tasks. As the GPU computing scene has evolved over the past few decades, Nvidia has made significant investments to capture the computing market. Part of this involves recognizing that computing tasks have different needs than graphics tasks, and diversifying its GPU family to better target each market.

The H100 is the latest addition to Nvidia's line of Gpus for computing. It uses the Hopper architecture and is built on a huge 814 mm2 chip using TSMC's 4N process and 80 billion transistors. This huge chip implements 144 streaming multiprocessors (SM), 60 MB of L2 cache, and 12 512-bit HBM memory controllers. We are testing the PCIe version of the H100 on Lambda Cloud, which supports 114 SM, 50 MB of L2 cache, and 10 HBM2 memory controllers. The card can consume up to 350 W of power.

Nvidia also offers the SXM form factor H100, which consumes up to 700W and enables 132 SMS. The SXM H100 also uses HBM3 memory, providing additional bandwidth to meet the needs of these additional shaders.