mirror of
https://github.com/LCTT/TranslateProject.git
synced 2025-01-13 22:30:37 +08:00
translated
This commit is contained in:
geekpi
parent
e8ff77bac7
commit
c5871ebcf8
@ -1,70 +0,0 @@
|
||||
[#]: collector: (lujun9972)
|
||||
[#]: translator: (geekpi)
|
||||
[#]: reviewer: ( )
|
||||
[#]: publisher: ( )
|
||||
[#]: url: ( )
|
||||
[#]: subject: (Nvidia quietly unveils faster, lower power Tesla GPU accelerator)
|
||||
[#]: via: (https://www.networkworld.com/article/3482097/nvidia-quietly-unveils-faster-lower-power-tesla-gpu-accelerator.html)
|
||||
[#]: author: (Andy Patrizio https://www.networkworld.com/author/Andy-Patrizio/)
|
||||
|
||||
Nvidia quietly unveils faster, lower power Tesla GPU accelerator
|
||||
======
|
||||
Nvidia has upgraded its Volta line of Tesla GPU-accelerator cards to work faster using the same power as its old model.
|
||||
client
|
||||
|
||||
Nvidia was all over Supercomputing 19 last week, not surprisingly, and made a lot of news which we will get into later. But overlooked was perhaps the most interesting news of all: a new generation graphics-acceleration card that is faster and way more power efficient.
|
||||
|
||||
Multiple attendees and news sites spotted it at the show, and Nvidia confirmed to me that this is indeed a new card. Nvidia’s “Volta” generation of Tesla GPU-accelerator cards has been out since 2017, so an upgrade was well overdue.
|
||||
|
||||
[[Get regularly scheduled insights by signing up for Network World newsletters.]][1]
|
||||
|
||||
The V100S comes only in PCI Express 3 form factor for now but is expected to eventually support Nvidia’s SXM2 interface. SXM is a dual-slot card design by Nvidia that requires no connection to the power supply, unlike the PCIe cards. SXM2 allows the GPU to communicate either with each other or to the CPU through Nvidia’s NVLink, a high-bandwidth, energy-efficient interconnect that can transfer data up to ten times faster than PCIe.
|
||||
|
||||
[][2]
|
||||
|
||||
BrandPost Sponsored by HPE
|
||||
|
||||
[Take the Intelligent Route with Consumption-Based Storage][2]
|
||||
|
||||
Combine the agility and economics of HPE storage with HPE GreenLake and run your IT department with efficiency.
|
||||
|
||||
With this card, Nvidia is claiming 16.4 single-precision TFLOPS, 8.2 double-precision TFLOPS, and Tensor Core performance of up to 130 TFLOPS. That is only a 4-to-5 percent improvement over the V100 SXM2 design, but 16-to-17 percent faster than the PCIe V100 variant.
|
||||
|
||||
Memory capacity remains at 32GB but Nvidia added High Bandwidth Memory 2 (HBM2) to increase memory performance to 1,134GB/s, a 26 percent improvement over both PCIe and SXM2.
|
||||
|
||||
Now normally a performance boost would see a concurrent increase in power demand, but in this case, the power envelope for the PCIe card is 250 watts, same as the prior generation PCIe card. So this card delivers 16-to-17 percent more compute performance and 26 percent more memory bandwidth at the same power draw.
|
||||
|
||||
**Other News**
|
||||
|
||||
Nvidia made some other news at the conference:
|
||||
|
||||
* A new reference design and ecosystem support for its GPU-accelerated Arm-based reference servers for high-performance computing. The company says it has support from HPE/Cray, Marvell, Fujitsu, and Ampere, the startup led by former Intel executive Renee James looking to build Arm-based server processors.
|
||||
* These companies will use Nvidia's reference design, which consists of hardware and software components, to build their own GPU-accelerated servers for everything from hyperscale cloud providers to high-performance storage and exascale supercomputing. The design also comes with CUDA-X, a special version of Nvidia’s CUDA GPU development language for Arm processors.
|
||||
* Launch of Nvidia Magnum IO suite of software designed to help data scientists and AI and high-performance-computing researchers process massive amounts of data in minutes rather than hours. It is optimized to eliminate storage and I/O bottlenecks to deliver up to 20x faster data processing for multi-server, multi-GPU computing nodes.
|
||||
* Nvidia and DDN, developer of AI and multicloud data management, announced a bundling of DDN’s A3ITM data management system with Nvidia’s DGX SuperPOD systems with so customers can deploy HPC infrastructure with minimal complexity and reduced timelines. The SuperPODs would also come with the new NVIDIA Magnum IO software stack.
|
||||
* DDN said that SuperPOD was able to be deployed within hours and a single appliance could scale all to 80 nodes. Benchmarks over a variety of different deep-learning models showed that the DDN system could keep a DGXSuperPOD system fully saturated with data.
|
||||
|
||||
|
||||
|
||||
**Now see** [**10 of the world's fastest supercomputers**][3]
|
||||
|
||||
Join the Network World communities on [Facebook][4] and [LinkedIn][5] to comment on topics that are top of mind.
|
||||
|
||||
--------------------------------------------------------------------------------
|
||||
|
||||
via: https://www.networkworld.com/article/3482097/nvidia-quietly-unveils-faster-lower-power-tesla-gpu-accelerator.html
|
||||
|
||||
作者：[Andy Patrizio][a]
|
||||
选题：[lujun9972][b]
|
||||
译者：[译者ID](https://github.com/译者ID)
|
||||
校对：[校对者ID](https://github.com/校对者ID)
|
||||
|
||||
本文由 [LCTT](https://github.com/LCTT/TranslateProject) 原创编译，[Linux中国](https://linux.cn/) 荣誉推出
|
||||
|
||||
[a]: https://www.networkworld.com/author/Andy-Patrizio/
|
||||
[b]: https://github.com/lujun9972
|
||||
[1]: https://www.networkworld.com/newsletters/signup.html
|
||||
[2]: https://www.networkworld.com/article/3440100/take-the-intelligent-route-with-consumption-based-storage.html?utm_source=IDG&utm_medium=promotions&utm_campaign=HPE20773&utm_content=sidebar ( Take the Intelligent Route with Consumption-Based Storage)
|
||||
[3]: https://www.networkworld.com/article/3236875/embargo-10-of-the-worlds-fastest-supercomputers.html
|
||||
[4]: https://www.facebook.com/NetworkWorld/
|
||||
[5]: https://www.linkedin.com/company/network-world
|
||||
@ -0,0 +1,56 @@
|
||||
[#]: collector: (lujun9972)
|
||||
[#]: translator: (geekpi)
|
||||
[#]: reviewer: ( )
|
||||
[#]: publisher: ( )
|
||||
[#]: url: ( )
|
||||
[#]: subject: (Nvidia quietly unveils faster, lower power Tesla GPU accelerator)
|
||||
[#]: via: (https://www.networkworld.com/article/3482097/nvidia-quietly-unveils-faster-lower-power-tesla-gpu-accelerator.html)
|
||||
[#]: author: (Andy Patrizio https://www.networkworld.com/author/Andy-Patrizio/)
|
||||
|
||||
Nvidia 悄悄推出更快、更低功耗的 Tesla GPU 加速器
|
||||
======
|
||||
Nvidia 升级了其 Volta 系列的 Tesla GPU 加速卡，使其能够以旧型号的相同功率更快地工作。
|
||||
|
||||
Nvidia 上周举行了 Supercomputing 19 大会，不出意外的是公布了很多新闻，这些我们将稍后提到。但被忽略的一条或许是其中最有趣的：一张更快、功耗更低的新一代图形加速卡。
|
||||
|
||||
多名与会者与多个新闻站点发现了这点，Nvidia 向我证实这确实是一张新卡。Nvidia 的 “Volta” 这代 Tesla GPU 加速卡在 2017 年就已淘汰，因此升级工作应该早已过期。
|
||||
|
||||
[[Get regularly scheduled insights by signing up for Network World newsletters.]][1]
|
||||
|
||||
V100S 目前仅提供 PCI Express 3 接口，但有望最终支持 Nvidia 的 SXM2 接口。SXM 是 Nvidia 的双插槽卡设计，与 PCIe 卡不同，它不需要连接电源。SXM2 允许 GPU 通过 Nvidia 的 NVLink（一种高带宽，节能互连）相互之间或与 CPU 进行通信，其数据传输速度比 PCIe 快十倍。
|
||||
|
||||
借助此卡，Nvidia 声称拥有单精度 16.4 TFLOPS，双精度 8.2 TFLOPS 并且 Tensor Core 性能高达 130 TFLOPS。这仅比 V100 SXM2 设计提高了 4％ 至 5％，但比 PCIe V100 变体提高了 16％ 至 17％。
|
||||
|
||||
内存容量保持在 32 GB，但 Nvidia 添加了 High Bandwidth Memory 2（HBM2），以将内存性能提高到 1,134 GB/s，这比 PCIe 和 SXM2 都提高了 26％。
|
||||
|
||||
通常情况下，性能提升将同时导致功率增加，但在这里，PCIe 卡的总体功率为 250 瓦，与上一代 PCIe 卡相同。因此，在相同功耗下，该卡可额外提供 16-17％ 的计算性能，并增加 26％ 的内存带宽。
|
||||
|
||||
**其他新闻**
|
||||
|
||||
Nvidia 在会上还发布了其他新闻：
|
||||
|
||||
* 其 GPU 加速的基于 Arm 的高性能计算参考服务器的新参考设计和生态系统支持。该公司表示，它得到了 HPE/Cray、Marvell、富士通和 Ampere 的支持，Ampere 是 Intel 前高管勒尼·詹姆斯（Renee James）领导的一家初创公司，它希望建立基于 Arm 的服务器处理器。
|
||||
* 这些公司将使用 Nvidia 的参考设计（包括硬件和软件组件）来使用 GPU 构建从超大规模云提供商到高性能存储和百亿亿次超级计算等。该设计还带来了 CUDA-X，这是 Nvidia 用于 Arm 处理器的 CUDA GPU 的特殊版本开发语言。
|
||||
* 推出 Nvidia Magnum IO 套件，旨在帮助数据科学家和 AI 以及高性能计算研究人员在几分钟而不是几小时内处理大量数据。它经过优化，消除了存储和 I/O 瓶颈，可为多服务器、多 GPU 计算节点提供高达 20 倍的数据处理速度。
|
||||
* Nvidia 和 DDN （AI 以及多云数据管理开发商）宣布将 DDN 的 A3ITM 数据管理系统与 Nvidia 的 DGX SuperPOD 系统捆绑在一起，以便客户能够以最小的复杂性和更短的时限部署 HPC 基础架构。SuperPOD 还带有新的 NVIDIA Magnum IO 软件栈。
|
||||
* DDN 表示，SuperPOD 能够在数小时内部署，并且单个设备可扩展至 80 个节点。不同的深度学习模型的基准测试表明，DDN 系统可以使 DGXSuperPOD 系统完全保持数据饱和。
|
||||
|
||||
|
||||
在 [Facebook][4] 和 [LinkedIn][5] 加入 Network World 社区评论热门主题。
|
||||
|
||||
--------------------------------------------------------------------------------
|
||||
|
||||
via: https://www.networkworld.com/article/3482097/nvidia-quietly-unveils-faster-lower-power-tesla-gpu-accelerator.html
|
||||
|
||||
作者：[Andy Patrizio][a]
|
||||
选题：[lujun9972][b]
|
||||
译者：[geekpi](https://github.com/geekpi)
|
||||
校对：[校对者ID](https://github.com/校对者ID)
|
||||
|
||||
本文由 [LCTT](https://github.com/LCTT/TranslateProject) 原创编译，[Linux中国](https://linux.cn/) 荣誉推出
|
||||
|
||||
[a]: https://www.networkworld.com/author/Andy-Patrizio/
|
||||
[b]: https://github.com/lujun9972
|
||||
[1]: https://www.networkworld.com/newsletters/signup.html
|
||||
[4]: https://www.facebook.com/NetworkWorld/
|
||||
[5]: https://www.linkedin.com/company/network-world
|
||||
Loading…
Reference in New Issue
Block a user