I’ve created texture for normal RT in Deferred Renderer:

D3D11_TEXTURE2D_DESC dtd {width, height, 1, 1,              
DXGI_FORMAT_R11G11B10_FLOAT,
1, 0, D3D11_USAGE_DEFAULT, D3D11_BIND_RENDER_TARGET | D3D11_BIND_SHADER_RESOURCE, 0, 0};

_check_hr(device->CreateTexture2D(&dtd, nullptr, &m_rtNormal));

It seems that normal have some artifacts:

[attachment=30093:Normals.png]

Other render targets for now have 32-bit format DXGI_FORMAT_R8G8B8A8_UNORM.

I only can imagine 3 choices how to improve normals:
1) Use 1.5 RT for Normals: DXGI_FORMAT_R32G32_FLOAT or DXGI_FORMAT_R16G16_FLOAT
X and Y will be in one RT, and Z – in the other.

2) Use 64-bit Render targets. For Normals use 16-bit floats: DXGI_FORMAT_R16G16B16A16_FLOAT

3) Use some math to calculate Z-value on first X and Y. But I want to avoid this approach.

So here are the questions:
Are there other better solutions for floating point normal with better that 11/10-bit per float?

In cases 1 and 2 in second RT there will be unused 16 or 32-bit float value.
Is there a way to represent it as UINT for my own use? How it can be done?

Thanks in advance.

First, there is no sign bit. So I suppose negative values become positive or get clamped to 0. You definitely don't want that.
Second, normals are in the [-1; 1] range. You will get much better precision by using DXGI_FORMAT_R10G10B10A2_UNORM which gets you 9 bits for the value and 1 bit for the sign; vs this float format which uses 5 bits for mantissa and 5 for the exponent.

Looks like you made a poor choice of format.

Thank you, Matias!

DXGI_FORMAT_R10G10B10A2_UNORM removed all the artifacts.