[av_heading heading=’On May 14, 2020, NVIDIA Founder and CEO Jensen Huang presented a new NVIDIA Ampere accelerator architecture during the GTC conference to address the most demanding tasks in the areas of HPC (High Performance Computing) or artificial intelligence. We will present a summary of the most interesting details and news in the following text.‘ tag=’h3′ style=’blockquote modern-quote modern-centered‘ subheading_active=“ size=“ av-medium-font-size-title=“ av-small-font-size-title=“ av-mini-font-size-title=“ subheading_size=“ av-medium-font-size=“ av-small-font-size=“ av-mini-font-size=“ color=’meta-heading‘ custom_font=“ margin=“ margin_sync=’true‘ padding=’10‘ link=“ link_target=“ id=“ custom_class=“ av_uid=’av-2fgpboy‘ admin_preview_bg=“][/av_heading]
From general purpose GPUs to AI supercomputers
[av_three_fifth first av_uid=’av-2q00l76′]
NVIDIA has always been synonymous with high-performance graphics cards (GPU). In 2007, they introduced the CUDA programming environment for general purpose graphic cards (GP GPU), making it a major player in accelerated computing for HPC (High Performance Computing) applications, data analysis, and more recently algorithms for artificial intelligence (AI). Since that NVIDIA has introduced a range of hardware and software technologies to further accelerate computing, and its accelerators have become the main computing power of the world’s most powerful supercomputers – IBM’s #1 Summit and #2 Sierra are built on NVIDIA Tesla V100 chips.
[/av_three_fifth]
[av_two_fifth av_uid=’av-23fg6s2′]
[/av_two_fifth]
[av_one_full first av_uid=’av-1fy4y36′]
GTC 2020
The GPU Technology Conference (GTC) is the world’s largest conference on graphics cards, accelerated computing and, in recent years, artificial intelligence. Regularly, the greatest interest is in the Jensen Huang’s keynote, who in his typical black leather jacket presents news and trends in the areas in which NVIDIA offers its products. It has been three years since the announcement of the breakthrough accelerator NVIDIA Tesla V100, so this year we were eagerly awaited what they had baked in NVIDIA. Jensen’s presentation begins traditionally … I AM AI.
[/av_one_full]
[av_one_half first av_uid=’av-2agazea‘][av_video src=’https://youtu.be/So7TNRhIYJ8′ mobile_image=“ attachment=“ attachment_size=“ format=’16-9′ width=’16‘ height=’9′ conditional_play=“ id=“ custom_class=“ av_uid=’av-1kt23ma‘][/av_one_half][av_one_half av_uid=’av-1dm0eb6′]
[av_video src=’https://youtu.be/bOf2S7OzFEg‘ mobile_image=“ attachment=“ attachment_size=“ format=’16-9′ width=’16‘ height=’9′ conditional_play=“ id=“ custom_class=“ av_uid=’av-rkqfw2′]
[/av_one_half]
[av_one_full first av_uid=’av-4dpzyq‘]
NVIDIA Ampere
After watching the video from Jensen’s kitchen, it is clear that the biggest news of this year’s presentation will be the GPU architecture of NVIDIA Ampere, the NVIDIA A100 accelerator, which is built on this architecture, and the DGX A100 system. DGX A100 is a reference architecture of a tuned system for AI and HPC calculations, equipped with eight NVIDIA A100 accelerators. In his video, Jensen takes the system board with these eight GPU accelerators out of the oven and introduces it as the most powerful accelerator in the world.
[/av_one_full]
[av_three_fifth first av_uid=’av-1hevyaa‘]
NVIDIA A100
The NVIDIA A100 accelerator is manufactured using 7 nm technology in the TSMC factory. It contains over 54 billion transistors on an area of 826 mm2. The main technological innovations of the new chip include:
- Tensor cores (3rd generation) optimized for HPC calculations and AI
- 40 GB fast HBM2 with 1.6 TB/s throughput, 40 MB L2 cache
- Multi-Instance GPU (MIG) for multi-user work with a single card
- faster NVLINK for interconnection of multiple cards (600 GB/s per card)
- PCIe gen4 bus for connection to CPU and network cards (31.5 GB/s)
Improvements of Tensor cores in FP64 (double precission) computations greatly helps computational performance in traditional HPC computations and simulations, while support for TF32 and BF16 accelerates training and subsequent application (inference) of neural networks. The maximum theoretical performance for the individual accuracies of the calculation is showed in the table on the right. The stated values take into account the Boost frequency of the computational cores. NVIDIA also introduced Fine grained structured sparsity for working with deep neural networks. Thanks to it, it is possible to double the stated performance values!
[/av_three_fifth]
[av_two_fifth av_uid=’av-1bvm642′]
[av_promobox button=’no‘ label=‘ icon_select=’no‘ icon=’ue800′ font=’roboto‘ color=’theme-color‘ custom_bg=’#444444′ custom_font=’#ffffff‘ size=’large‘ box_color=’theme-color‘ box_custom_font=’#ffffff‘ box_custom_bg=’#d30007′ box_custom_border=“ label_display=“ link=“ link_target=“ id=“ custom_class=“ av_uid=’av-qse7hu‘ admin_preview_bg=“]
| ACCURANCY | PERFORMANCE |
|---|---|
| FP64 | 9,7 TFLOPS |
| FP64 Tensor Core | 19,5 TFLOPS |
| FP32 | 19,5 TFLOPS |
| FP16 | 78 TFLOPS |
| BF16 | 39 TFLOPS |
| TF32 Tensor Core | 156 TFLOPS |
| FP16 Tensor Core | 312 TFLOPS |
| BF16 Tensor Core | 312 TFLOPS |
| INT8 Tensor Core | 624 TOPS |
| INT4 Tensor Core | 1 248 TOPS |
[/av_promobox]
[/av_two_fifth]
[av_one_full first av_uid=’av-mpmuvm‘]
Comparison of NVIDIA A100 with existing accelerators
In general, the NVIDIA A100 accelerator is the most powerful hardware on which demanding computing tasks can be processed.
| ParametEr | Tesla T4 | Tesla V100 SXM2 | A100 SXM4 | DGX-1 | DGX-2 | DGX A100 |
|---|---|---|---|---|---|---|
| Architecture | Turing | Volta | Ampere | Volta | Volta | Ampere |
| Nr. of CUDA Cores | 2 560 | 5 120 | 6 912 | 40 960 | 81 920 | 55 296 |
| Nr. of Tensor Cores | 320 | 640 | 432 | 5 120 | 10 240 | 3 456 |
| FP64 Performance | 0,25 TFLOPS | 7,8 TFLOPS | 9,7 TFLOPS | 62 TFLOPS | 125 TFLOPS | 77,6 TFLOPS |
| Tensor (FP64) Performance | — | — | 19,5 TFLOPS | — | — | 156 TFLOPS |
| FP32 Performance | 8,1 TFLOPS | 15,7 TFLOPS | 19,5 TFLOPS | 125 TFLOPS | 250 TFLOPS | 156 TFLOPS |
| Tensor (TF32) Performance | — | — | 156 TFLOPS | — | — | 1,2 PFLOPS |
| Tensor (FP16) Performance | 65 TFLOPS | 125 TFLOPS | 312 TFLOPS | 1 PFLOPS | 2 PFLOPS | 2,5 PFLOPS |
| GPU Memory | 16 GB | 32 GB | 40 GB | 256 GB | 512 GB | 320 GB |
| Memory Technology | GDDR6 | HMB2 | HMB2 | HMB2 | HMB2 | HBM2 |
| Memory Throughput | 300 GB/s | 900 GB/s | 1 ,5 TB/s | 900 GB/s | 900 GB/s | 1,5 TB/s |
| GPU card interconnection | none | NVLink | NVLink3 | NVLink, hypercube topologie | NVSwitch, non-blocking | NVSwitch3, non-blocking |
| Maximum power consumption | 70 W | 300 W | 400 W | 3 500 W | 10 kW | 6,6 kW |
| Form Factor | PCIe card | SXM2 card | SXM4 card | rack, 3U | rack, 10U | rack |
| On the Market since | 2018 | 2017 | 2020 | 2017 | 2018 | 2020 |
We reviewed the current CPU platforms (Intel, AMD, POWER and ARM) in the March Comparsion of Server Processors. We are also preparing a comparison of accelerators for HPC and AI.
[/av_one_full]
[av_one_half first av_uid=’av-30xutv6′]
[av_image src=’https://mcomputers.algaweb.site/wp-content/uploads/2020/05/A100_AI_acceleration.png‘ attachment=’12924′ attachment_size=’full‘ copyright=“ caption=“ styling=“ align=’center‘ font_size=“ overlay_opacity=’0.4′ overlay_color=’#000000′ overlay_text_color=’#ffffff‘ animation=’no-animation‘ hover=“ appearance=“ link=’lightbox‘ target=“ id=“ custom_class=“ av_element_hidden_in_editor=’0′ av_uid=’av-2j8t0ky‘ admin_preview_bg=“][/av_image]
[/av_one_half]
[av_one_half av_uid=’av-28b0x2q‘]
[av_image src=’https://mcomputers.algaweb.site/wp-content/uploads/2020/05/A100_HPC_acceleration.png‘ attachment=’12926′ attachment_size=’full‘ copyright=“ caption=“ styling=“ align=’center‘ font_size=“ overlay_opacity=’0.4′ overlay_color=’#000000′ overlay_text_color=’#ffffff‘ animation=’no-animation‘ hover=“ appearance=“ link=’lightbox‘ target=“ id=“ custom_class=“ av_element_hidden_in_editor=’0′ av_uid=’av-1fz5dte‘ admin_preview_bg=“][/av_image]
[/av_one_half]
[av_one_full av_uid=’av-14vh3f6′]
NVIDIA DGX A100
When announcing the previous generation of Tesla accelerators, NVIDIA also introduced reference systems optimized for HPC and AI calculations – NVIDIA DGX. With the introduction of the NVIDIA A100 accelerator, a new DGX system – the NVIDIA DGX A100 – was also introduced.
[av_image src=’https://mcomputers.algaweb.site/wp-content/uploads/2020/05/dgx-a100-exploded-view.png‘ attachment=’12930′ attachment_size=’full‘ copyright=“ caption=“ styling=“ align=’center‘ font_size=“ overlay_opacity=’0.4′ overlay_color=’#000000′ overlay_text_color=’#ffffff‘ animation=’no-animation‘ hover=“ appearance=“ link=’lightbox‘ target=“ id=“ custom_class=“ av_element_hidden_in_editor=’0′ av_uid=’av-vj11gy‘ admin_preview_bg=“][/av_image]
Along with the DGX A100 system, NVIDIA also introduced the HGX A100 system, a system board equipped with four or eight NVIDIA A100 accelerators that OEM and ODM server manufacturers integrate into their systems. HPE, Gigabyte and Supermicro should be among the first manufacturers of these systems.
[/av_one_full]
[av_section min_height=“ min_height_px=’500px‘ padding=’default‘ shadow=’no-border-styling‘ bottom_border=’no-border-styling‘ bottom_border_diagonal_color=’#333333′ bottom_border_diagonal_direction=“ bottom_border_style=“ custom_arrow_bg=“ id=“ color=’main_color‘ custom_bg=’#d71923′ src=’https://mcomputers.algaweb.site/wp-content/uploads/2018/08/house-wire-model-fade-white-1500×1085.png‘ attachment=’7397′ attachment_size=’full‘ attach=’parallax‘ position=’center right‘ repeat=’stretch‘ video=“ video_ratio=’16:9′ overlay_opacity=’0.5′ overlay_color=“ overlay_pattern=“ overlay_custom_pattern=“ av_element_hidden_in_editor=’0′ av_uid=’av-17n60gn‘]
[av_three_fourth first min_height=’av-equal-height-column‘ vertical_alignment=’av-align-middle‘ space=“ margin=’0px‘ margin_sync=’true‘ padding=’0px‘ padding_sync=’true‘ border=“ border_color=“ radius=’0px‘ radius_sync=’true‘ background_color=“ src=“ attachment=“ attachment_size=“ background_position=’top left‘ background_repeat=’no-repeat‘ animation=“ mobile_breaking=“ mobile_display=“ av_uid=’av-15taedj‘]
[av_heading heading=’NVIDIA offers for both Tesla cards and DGX systems special pricing promotions and programs for specific projects, and also supports educational institutions (EDU) or start-ups.‘ tag=’h3′ link_apply=“ link=’manually,http://‘ link_target=“ style=’blockquote modern-quote‘ size=’25‘ subheading_active=“ subheading_size=’15‘ margin=“ margin_sync=’true‘ padding=’0′ color=’custom-color-heading‘ custom_font=’#ffffff‘ av-medium-font-size-title=“ av-small-font-size-title=“ av-mini-font-size-title=“ av-medium-font-size=“ av-small-font-size=“ av-mini-font-size=“ av_uid=’av-js85uejr‘ admin_preview_bg=“][/av_heading]
[/av_three_fourth][av_one_fourth min_height=“ vertical_alignment=“ space=“ custom_margin=“ margin=’0px‘ padding=’0px‘ border=“ border_color=“ radius=’0px‘ background_color=“ src=“ background_position=’top left‘ background_repeat=’no-repeat‘ animation=“ mobile_breaking=“ mobile_display=“ av_uid=’av-3w8lfr‘]
[av_button label=’Get an offer‘ link=’manually,[email protected]‘ link_target=’_blank‘ size=’large‘ position=’right‘ icon_select=’no‘ icon=’ue800′ font=’entypo-fontello‘ color=’light‘ custom_bg=’#444444′ custom_font=’#ffffff‘ admin_preview_bg=“ av_uid=’av-10wh353′]
[/av_one_fourth]
[/av_section]
More news from NVIDIA
[av_one_half first av_uid=’av-1rukh02′]And this is just a brief list of the main innovations that Jensen has introduced in his kitchen. Individual parts of the presentation can be viewed on the NVIDIA youtube channel:
[/av_one_half][av_one_half av_uid=’av-174brde‘][av_video src=’https://youtu.be/bOf2S7OzFEg‘ mobile_image=“ attachment=“ attachment_size=“ format=’16-9′ width=’16‘ height=’9′ conditional_play=“ id=“ custom_class=“ av_uid=’av-qmuvki‘][/av_one_half]
Webinars
We have prepared for you a series of webinars on the topic of accelerated calculations and especially news from the GTC conference and presentation of Jensen Huang.
- May 20, 2020, 3:00 PM: Introduction to Accelerated Computing
- May 21, 2020, 9:00: NVIDIA News, Tesla Ampere Architecture
- May 27, 2020, 3:00 PM: NVIDIA News, Tesla Ampere Architecture (same content as May 21)
- June 3, 2020, 3:00 PM: NVIDIA GPU Cloud (NGC) in practice
The content and detailed description of individual webinars can be found in the Training and webinars.
Courses
[av_one_third first min_height=’av-equal-height-column‘ vertical_alignment=’av-align-middle‘ space=“ margin=’0px‘ margin_sync=’true‘ padding=’0px‘ padding_sync=’true‘ border=“ border_color=“ radius=’0px‘ radius_sync=’true‘ background_color=“ src=“ attachment=“ attachment_size=“ background_position=’top left‘ background_repeat=’no-repeat‘ animation=“ mobile_breaking=“ mobile_display=“ av_uid=’av-qqkagn‘]
[av_image src=’https://mcomputers.algaweb.site/wp-content/uploads/2018/11/deep-learning-education-logo-300×106.png‘ attachment=’8509′ attachment_size=’medium‘ align=’center‘ styling=“ hover=“ link=“ target=“ caption=“ font_size=“ appearance=“ overlay_opacity=’0.4′ overlay_color=’#000000′ overlay_text_color=’#ffffff‘ animation=’no-animation‘ admin_preview_bg=“ av_uid=’av-q92kpj‘][/av_image]
[/av_one_third][av_two_third min_height=“ vertical_alignment=“ space=“ custom_margin=“ margin=’0px‘ padding=’0px‘ border=“ border_color=“ radius=’0px‘ background_color=“ src=“ background_position=’top left‘ background_repeat=’no-repeat‘ animation=“ mobile_breaking=“ mobile_display=“ av_uid=’av-omtybb‘]
[av_textblock size=“ font_color=“ color=“ av-medium-font-size=“ av-small-font-size=“ av-mini-font-size=“ admin_preview_bg=“ av_uid=’av-mwjspz‘]
NVIDIA Deep Learning Institute
NVIDIA Deep Learning Institute (DLI) offers online and hands-on training for developers, data analysts, or researchers who are challenging artificial intelligence or accelerated computing.
[/av_textblock]
[/av_two_third][av_three_fourth first min_height=’av-equal-height-column‘ vertical_alignment=’av-align-middle‘ space=“ margin=’0px‘ margin_sync=’true‘ padding=’0px‘ padding_sync=’true‘ border=“ border_color=“ radius=’0px‘ radius_sync=’true‘ background_color=“ src=“ attachment=“ attachment_size=“ background_position=’top left‘ background_repeat=’no-repeat‘ animation=“ mobile_breaking=“ mobile_display=“ av_uid=’av-l5n4xz‘]
[av_textblock size=“ font_color=“ color=“ av-medium-font-size=“ av-small-font-size=“ av-mini-font-size=“ admin_preview_bg=“ av_uid=’av-idzlpj‘]
Testing
[/av_textblock]
[av_textblock size=“ font_color=“ color=“ id=“ custom_class=“ av-medium-font-size=“ av-small-font-size=“ av-mini-font-size=“ av_uid=’av-i4mgd3′ admin_preview_bg=“]
To test the performance and especially the speed of deployment of ML and AI applications, we have the NVIDIA DGX Station system and within the NVIDIA Tesla Test Drive program also 2 × NVIDIA Tesla V100 and NVIDIA Tesla T4. If you are interested in testing, please fill out this form.
[/av_textblock]
[av_contact email=’[email protected]‘ from_email=“ title=“ heading_tag=“ heading_class=“ button=’Send‘ on_send=“ sent=’Your interest in testing has been submitted.‘ link=’manually,http://‘ subject=’Testing of NVIDIA DGX products‘ autoresponder_email=’[email protected]‘ autorespond=’I am interested in testing NVIDIA DGX products‘ captcha=“ captcha_theme=’light‘ captcha_size=’normal‘ captcha_score=’0.5′ form_align=“ color=“ alb_description=“ id=“ custom_class=“ aria_label=“ av_uid=’av-gccplz‘ admin_preview_bg=“]
[av_contact_field label=’NVIDIA product‘ type=’select‘ options=’NVIDIA DGX Station, NVIDIA Testa V100, NVIDIA Tesla T4, NVIDIA Jetson Nano‘ check=’is_empty‘ width=’element_fourth‘ av_uid=’av-1v1s1j‘][/av_contact_field]
[av_contact_field label=’Name‘ type=’text‘ options=“ check=’is_empty‘ width=’element_fourth‘ multi_select=“ av_contact_preselect=“ av_uid=’av-bv6vvb‘][/av_contact_field]
[av_contact_field label=’E-Mail‘ type=’text‘ options=“ check=’is_email‘ width=’element_fourth‘ multi_select=“ av_contact_preselect=“ av_uid=’av-13grl3′][/av_contact_field]
[/av_contact]
[/av_three_fourth][av_one_fourth min_height=“ vertical_alignment=“ space=“ custom_margin=“ margin=’0px‘ padding=’0px‘ border=“ border_color=“ radius=’0px‘ background_color=“ src=“ background_position=’top left‘ background_repeat=’no-repeat‘ animation=“ mobile_breaking=“ mobile_display=“ av_uid=’av-3w8lfr‘]
[av_team_member name=’Kamila Jeřábková‘ job=’Product manager‘ src=’https://mcomputers.algaweb.site/wp-content/uploads/2019/09/kalca-kulata.png‘ attachment=’10713′ attachment_size=’full‘ image_width=“ description=’M: 734 161 516
[email protected]‚ font_color=“ custom_title=“ custom_content=“ av_uid=’av-k0bh8721′ admin_preview_bg=“]
[av_team_icon title=’Poslat e-mail‘ link=’mailto:[email protected]‘ link_target=’_blank‘ icon=’ue805′ font=’entypo-fontello‘ av_uid=’av-rzqilq‘]
[/av_team_member]
[/av_one_fourth]
M Computers s.r.o. is NVIDIA’s NPN partner with the highest degree of ELITE partnership for the areas of the most powerful TESLA accelerators and complete AI systems – NVIDIA DGX Systems. We supply NVIDIA products in the Czech Republic, Slovakia and other countries in Central and Eastern Europe.