AMD GPU正式支援深度學習框架tensorflow - 3C

消息來源：
https://gpuopen.com/rocm-tensorflow-1-8-release/
https://rocm.github.io/ROCmInstall.html

對應於Nvidia的cuda SDK的AMD ROCm終於正式支援新版tensorflow (1.8版)

不過目前只支援Ubuntu /CentOS等Linux作業系統

GPU方面也有限制。VEGA應無大礙，至於舊版的架構如Polaris以及VEGA APU似乎還沒。



VEGA的賣點是有高速的HBM2

加上可以拿系統RAM當GPU RAM的HBCC技術

不過AMD的優化...



有空的話，我會拿我的VEGA 56來和1070ti PK看看深度學習的運算效能

看看AMD是到底是優化效能還是優化笑能


以下是官方的支援訊息：

Supported CPUs

Starting with ROCm 1.8 we have relaxed the use of PCIe Atomics and also PCIe
lane choice for Vega10/GFX9 class GPU. So now you can support CPU without
PCIe Atomics and also use Gen2 x1 lanes.

Currently our GFX8 GPU’s (Fiji & Polaris family) still need to use PCIe Gen
3 and PCIe Atomics, but are looking at relaxing this in a future release,
once we have fully tested firmware.

Current CPUs which support PCIe Gen3 + PCIe Atomics are:

AMD Ryzen CPUs;
AMD EPYC CPUs;
Intel Xeon E7 V3 or newer CPUs;
Intel Xeon E5 v3 or newer CPUs;
Intel Xeon E3 v3 or newer CPUs;
Intel Core i7 v4, Core i5 v4, Core i3 v4 or newer CPUs (i.e. Haswell
family or newer).

For Fiji and Polaris GPU’s the ROCm platform leverages PCIe Atomics (Fetch
and Add, Compare and Swap, Unconditional Swap, AtomicsOp Completion). PCIe
Atomics are only supported on PCIe Gen3 enabled CPUs and PCIe Gen3 switches
like Broadcom PLX. When you install your GPUs make sure you install them in a
fully PCIe Gen3 x16 or x8, x4 or x1 slot attached either directly to the CPU’
s Root I/O controller or via a PCIe switch directly attached to the CPU’s
Root I/O controller. In our experience many issues stem from trying to use
consumer motherboards which provide physical x16 connectors that are
electrically connected as e.g. PCIe Gen2 x4 connected via the Southbridge
PCIe I/O controller.

Experimental support for our GFX7 GPUs Radeon R9 290, R9 390, AMD FirePro
S9150, S9170 note they do not support or take advantage of PCIe Atomics.
However, we still recommend that you use a CPU from the list provided above.
Not supported or very limited support under ROCm
Limited support

With ROCm 1.8 and Vega10 it should support PCIe Gen2 enabled CPUs such as
the AMD Opteron, Phenom, Phenom II, Athlon, Athlon X2, Athlon II and older
Intel Xeon and Intel Core Architecture and Pentium CPUs. But we have done
very limited testing. Since our test farm today has been catering to CPU
listed above. This is where we need community support.
Thunderbolt 1,2 and 3 enabled breakout boxes GPU’s should now be able to
work with ROCm. Thunderbolt 1 and 2 are PCIe Gen2 based. But we have done no
testing on this config and would need comunity support do limited access to
this type of equipment

Not supported

We also do not support AMD Carrizo and Kaveri APU as host for compliant
dGPU attachments.
Thunderbolt 1 and 2 enabled GPU’s are not supported by ROCm. Thunderbolt
1 & 2 are PCIe Gen2 based.
AMD Carrizo based APUs have limited support due to OEM & ODM’s choices
when it comes to some key configuration parameters. On point, we have
observed that Carrizo laptops, AIOs and desktop systems showed
inconsistencies in exposing and enabling the System BIOS parameters required
by the ROCm stack. Before purchasing a Carrizo system for ROCm, please verify
that the BIOS provides an option for enabling IOMMUv2. If this is the case,
the final requirement is associated with correct CRAT table support - please
inquire with the OEM about the latter.
AMD Merlin/Falcon Embedded System is also not currently supported by the
public repo.
AMD Raven Ridge APU are currently not supported


--