FAH and Xeon Phi compatibility?

[MeeLee]

Any new info on this article? : https://foldingathome.org/faqs/high-per ... -xeon-phi/

A Xeon Phi 7210 can be found as low as $200 on the second hand market.
Is there support for FAH for these CPUs?

Code: Select all

Essentials
Product Collection
Intel® Xeon Phi™ x200 Product Family

Code Name
Products formerly Knights Landing

Vertical Segment
Server

Processor Number
7210

Status
Discontinued

Launch Date
Q2'16

Lithography
14 nm

Recommended Customer Price
$1881.00
Performance
# of Cores
64

Processor Base Frequency
1.30 GHz

Max Turbo Frequency
1.50 GHz

Cache
32 MB L2

TDP
215 W

VID Voltage Range
0.550-1.125V

Supplemental Information
Embedded Options Available
No

Datasheet
View now

Description
Intel Xeon Phi processors do not currently support Virtualization. Intel is working with interested OEMs, customers, and the virtualization developer community in providing them with Virtualization feature at a later date.

Memory Specifications
Max Memory Size (dependent on memory type)
384 GB

Memory Types
DDR4-2133

Max # of Memory Channels
6

Max Memory Bandwidth
102 GB/s

ECC Memory Supported ‡
Yes

Expansion Options
PCI Express Revision
3.0

PCI Express Configurations ‡
x16 port (Port 2 and 3) may negotiate down to x8, x4, x2, or x1. x4 port (Port1) may negotiate down to x2, or x1

Max # of PCI Express Lanes
36

Package Specifications
Sockets Supported
SVLCLGA3647

Advanced Technologies
Intel® Turbo Boost Technology ‡
2.0

Intel® Virtualization Technology (VT-x) ‡
No

Intel® Virtualization Technology for Directed I/O (VT-d) ‡
No

Intel® VT-x with Extended Page Tables (EPT) ‡
No

Intel® 64 ‡
Yes

Instruction Set
64-bit

Instruction Set Extensions
Intel® AVX-512

Idle States
Yes

Thermal Monitoring Technologies
Yes

Security & Reliability
Intel® AES New Instructions
Yes

Intel® Software Guard Extensions (Intel® SGX)
No

Intel® Memory Protection Extensions (Intel® MPX)
No

Intel® Trusted Execution Technology ‡
No

Execute Disable Bit ‡
Yes

Intel® Boot Guard
No

[bruce]

That's a question for GROMACS.org, not FAH.

See https://redmine.gromacs.org/issues/1944


FAH uses the GROMACS code for native CPUs (in x64 mode plus SSE2 and/or AVX). THe OpenMM code is used for GPUs by-way-of OpenCL. The Phi is neither.

[JimboPalmer]

Knights Landing Xeon Phi is based on an Silvermont Atom CPU with an odd version of AVX512.
https://en.wikipedia.org/wiki/Xeon_Phi#Knights_Landing
https://en.wikipedia.org/wiki/Silvermont
https://en.wikipedia.org/wiki/AVX-512#CPUs_with_AVX-512
https://en.wikipedia.org/wiki/GROMACS

Potentially you could write a Core for Silvermont and SSE2 and ignore the snazzy AVX-512 subset that Knights Landing supports.
Potentially you could try to write a Core for the Subset of AVX-512 that the Knights Landing supports.

Either would depend on GROMACS code that is Xeon Phi aware.

[foldy]

Gromacs supports Xeon Phi but it is not very fast
https://software.intel.com/en-us/articl ... processors
http://manual.gromacs.org/documentation ... l-xeon-phi

[bruce]

How many Donors have a GPU slot? How many Donors have a CPU slot? How many new Phy slots do you expect to be added?

It isn't worth it for FAH to invest development resources into such a negligible increase in total throughput of this massively parallel distributed supercomputer.

[MeeLee]

I was just interested in some of the specs, but it seems Nvidia outdid them.
While Xeon phi is cpu based, and Nvidia is GPU based, the Xeon phi had 1 to 2 tflops of processing power, which is phenomenal for a cpu (even if ran through a socket or PCIE slot).

The cons seem to be the rather exotic hardware requirements. Expensive motherboards, that are 3 years old, with low compatibility (meaning, you'd have to be lucky to be able to get a motherboard that supports such a card), and they need Xeon cpu processors to drive them (or an occasional older i7 would work).

The biggest con isn't performance, as it was made for tasks like distributed computing, but at the promised 1-2Tflops, the card runs at up to 250W.
I was interested because they mentioned 65W and 1+Tflops in one article, however at 65W, it should only be able to get about 250Gflops. Compare that to an rtx2060 doing 8+Tflops at below 150W.

I do have some feeling as if the Xeon phi was the predecessor to the RTX cards today.
Built on the se optimization of massive parrallel clusters of cores.
One being cpu, the other arm.

[Frisa]

it seems Knights Landing supports all SIMD instructions from original SSE all way to AVX512, and its software package supports centos and ubuntu. so in theory it could run FAH just like your ordinary pc does

[Frisa]

How many Donors have a GPU slot? How many Donors have a CPU slot? How many new Phy slots do you expect to be added?

It isn't worth it for FAH to invest development resources into such a negligible increase in total throughput of this massively parallel distributed supercomputer.

in fact, in term of double precision performance it blow away consumer grade graphic card, which deliberately nerfed

the lowest end model of KNL xeon phi (Xeon Phi 7210), which comes with DP performance of 2662 GFLOPS. On the contrary, the DP performance of RTX 2080 Ti is just 421 GFLOPS

[JimboPalmer]

1) the F@H does do Double Precision calculations, but rarely.

2) the AMD Radeon VII consumer card does 2,784 DP GFLOPS at 1.4 Ghz and 3,458.5 DP GFLOPS at 1.75 Ghz. https://en.wikipedia.org/wiki/Radeon_RX ... e_graphics

[MeeLee]

I think the IBM Power 9 processor might be more interesting, if it could be used for folding.
Lower power consumption, built on 14nm, 4Ghz, 12/24 cores...
They're actually going for 'affordable' rates. (about $900 per CPU).
Although I don't think the rest of the system is 'as affordable'

https://en.wikipedia.org/wiki/POWER9

Exotic hardware!

[JimboPalmer]

If I was in charge of software development for F@H, (I am not) I would be focusing on RDNA from AMD. The RX 5700 XT is not currently supported and should be a high volume series of parts for the next decade.

In my mind, a much better return on investment than Phi or POWER.

https://en.wikipedia.org/wiki/RDNA_(microarchitecture)

[bruce]

You're going to get your wish. OpenMM has incorporated support for AMD's RDNA into the latest code. I expect it to be in FAHCore_22 (or it may already be in the beta code). That change has not been back-ported to FAHCore_21 and probably won't be, depending on the progress of the beta testing of the Core_22 code.