Radeon Instinct mi60 - Double float is usable on GPU Projects Y/N

QuantumHelos
QuantumHelos
Joined: 5 Nov 17
Posts: 190
Credit: 64239858
RAC: 0
Topic 218730

Radeon Instinct mi60 - Double float is usable on GPU Projects Y/N

 

Usable in the best Servers & MiPower Blades (TM)

https://www.amd.com/system/files/documents/radeon-instinct-mi60-datasheet.pdf

PERFORMANCE Compute Units Stream Processors Peak INT8 Peak FP16 Peak FP32 Peak FP64 Bus Interface
64 4,096 Up to 59.0 TOPS Up to 29.5 TFLOPS Up to 14.7 TFLOPS Up to 7.4 TFLOPS PCIe®Gen 3 and Gen 4 Capable

https://www.amd.com/en/products/professional-graphics/instinct-mi60

https://www.amd.com/system/files/documents/lawrence-livermore-national-laboratory-case-study.pdf

https://www.amd.com/en/case-studies/lawrence-livermore-national-laboratory

https://www.amd.com/en/processors/server-tech-docs

Regards RS
https://science.n-helix.com

Peter van Kalleveen
Peter van Kalleveen
Joined: 15 Jan 19
Posts: 45
Credit: 250329645
RAC: 0

yes, its the datacenter

yes, its the datacenter radeon VII.

Works very well, most jobs only have partial or small amounts of fp64, so don't expect astronomical benefits here.

But probably the compute/datacenter drivers are a lot better for low overhead, so I would expect you to blow the VII away, and that one already blows almost all other away for compute.

QuantumHelos
QuantumHelos
Joined: 5 Nov 17
Posts: 190
Credit: 64239858
RAC: 0

So what are the key

So what are the key differences ? peter

Mad_Max
Mad_Max
Joined: 2 Jan 10
Posts: 153
Credit: 2134772970
RAC: 461437

Key differences between 

Key differences between  radeon VII and Radeon Instinct MI 50

- even more FP64 speed (twice as radeon VII)
- full multi-GPU support to combine bunch of GPUs into single computing cluster
- drivers optimized more for work/computations not games

Radeon Instinct MI 60 is slightly faster in all aspects as it has a little more compute units (64 in MI 60 vs 60 in MI 50 and radeon VII) and twice as much VRAM volume. radeon VII is a consumer/gamer version of MI 50.

But actually all these differences mean not too much in terms of BOINC calculations. As most of software here developed to use FP64 at minimum level (or do not use it all if it possible) and intended to use with standard ("gamers") drivers capabilities. Also i did not hear about any multi-GPU support - each GPU work independently.

So you can expect from radeon VII almost same results as from Radeon Instinct. May be 5-20% slower depended of app/project.
The only exception will be projects that fully rely on double-precision calculations. At the moment I only know one such - MilkyWay@Home.

Peter van Kalleveen
Peter van Kalleveen
Joined: 15 Jan 19
Posts: 45
Credit: 250329645
RAC: 0

What the previous poster told

What the previous poster told was a good explanationWink

Lastly just to add, the data center versions are the Top binned chips, they are the best working chips that need the least amount of voltage to operate on as high as possible frequency.

The mi50 has the lesser working chips that have small defects that result in inoperable shades, need more voltage and achieves a less high frequency.

All the chips that remain can go in the trash,...… or in consumer cards where voltage is less important because we have big coolers and nothing like datacenter density in racks. Frequency is achieved by just giving even more voltage. And of course inoperable shaders and disabled double precision and so the VII is born.

Same goes for all brands, cards and other silicon like cpu's.

1 Datacenter 1500-15000 per gpu

2 professional workstation 300-6000 per gpu

3 gamers and other people, 75-1200 per gpu.

In the end its the same gpu die, only in more crippled and worse condition.

QuantumHelos
QuantumHelos
Joined: 5 Nov 17
Posts: 190
Credit: 64239858
RAC: 0

So lots of room for decent

So lots of room for decent models to sift down to the "consumer market" afterall are we gamers? do we qualify with only 5 years of Quake N7 or elite playtime :p

 

But yes einstein@home is a most exciting way to explore the stars.

 

Binning being the only way datacenters can actually affort to play with complex tools like the Epyc

Peter van Kalleveen
Peter van Kalleveen
Joined: 15 Jan 19
Posts: 45
Credit: 250329645
RAC: 0

Its not as much that Epyc is

Its not as much that Epyc is complex, just 4 standard ryzen dies in a package, nothing more.

But the environment of the datacenter is so much more punishing, rack density is going true the roof with al the gpu accelerated nodes, often the cooling is problematic because of the low height of the rack enclosure and again density. It needs to be able to run 24/7 for years without failure and is much more mission critical than consumer orientated hardware.

But we@home can contribute a lot in this way, and at hearth we are all nerds and geeks because,.... well we contribute and love these kind of things Cool 

 

I am very curious how much improved performance compute orientated drivers bring to the table.

At present I am very far away from reaching my VII max theoretical flops in the real world

QuantumHelos
QuantumHelos
Joined: 5 Nov 17
Posts: 190
Credit: 64239858
RAC: 0

https://wccftech.com/amd-ryze

https://wccftech.com/amd-ryzen-3000-cpus-7nm-zen-2-first-epyc-rome-q3-2019-launch/

https://wccftech.com/playstation-5-special-sauce/

Recent rumor revealed that the console’s devkit is running a nearly 13 teraflops GPU. What we know for sure is that the console will be powered by a CPU based on the third generation of AMD’s Ryzen line and by a GPU that will support ray tracing.

https://wccftech.com/amd-confirms-navi-gpu-launching-in-q3-2019/

QuantumHelos
QuantumHelos
Joined: 5 Nov 17
Posts: 190
Credit: 64239858
RAC: 0

Open Source Driver for Vulkan

Open Source Driver for Vulkan : Debian/Ubuntu/Linux
https://github.com/GPUOpen-Drivers/AMDVLK

run this after downloading file: https://is.gd/Install_gpl_amd_drivers_sh

sudo chmod 774 Install-gpl-amd-drivers.sh
sudo ./Install-gpl-amd-drivers.sh

****
GL to Vulkan : gfx-portability : Prototype library implementing Vulkan Portability Initiative using gfx-hal. See gfx-rs meta issue for backend limitations and further details.

https://github.com/gfx-rs/portability

****
OpenGLES/OpenCL/OpenGL/Vulkan API : Mac:Windows:Linux:Android

https://developer.arm.com/tools-and-software/graphics-and-gaming/graphics-development-tools/opengl-es-emulator

Mac Vulkan Open CL

https://github.com/KhronosGroup/MoltenVK/releases/tag/v1.0.35

 

Usable for coding:

https://github.com/KhronosGroup/OpenCL-ICD-Loader

https://github.com/KhronosGroup/OpenCL-CLHPP

****
Texture & polygon optimiser & compressor : may be useful for Einstein development

https://github.com/GPUOpen-Tools/Compressonator/releases

https://github.com/KhronosGroup/glTF-Compressonator

 

Useful for phone development:

WebCLGL : Libraries & JS
https://github.com/stormcolor/webclgl
https://github.com/stormcolor/webclgl/blob/master/dist/webclgl/WebCLGL.min.js

WebCLGL use WebGL2 specification to interpret code.
WebGL is used like OpenCL for GPGPU calculus using the traditional Render To Texture technique.

****
WebGL Compute
https://www.khronos.org/registry/webgl/specs/latest/2.0-compute/#diff-with-gles31

https://www.khronos.org/assets/uploads/developers/library/2017-webgl-webinar/Khronos-Webinar-WebGL-20-is-here_What-you-need-to-know_Apr17.pdf

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.