Infinisearch
Veteran
@Andrew Lauritzen .. Am I right about IMT (interleaved multi-threading) being fixed temporal multiplexed threads in hardware? If so how many threads?
I'm not sure actually. I'd have to go digging... that's not something that is relevant to writing code on the machine very often
pixelio
Newcomer
Summer must be over since, as of yesterday, the showstopper bugs I was seeing in Broadwell's OpenCL 2.x driver are fixed in Windows 10 (driver 4256) and there is a new Code Builder. Huzzah!
I think I'm seeing the result of Broadwell Gen8's doubled integer throughput too.
FWIW, GPU-Z reports a 950 MHz HD 6000 using 5.5-6.0 Watts when continuously sorting 131K 64-bit keys and only 4 Watts for 262K 32-bit keys.
We're really at the dawn of the golden age of GPUs where even power-sensitive GPUs have significant FLOPS.
Interesting. Haswell turns out even more limited in hardware than I imagined. Intel just upgraded Broadwell and Skylake to OpenGL 4.4., but Haswell is staying on 4.3.
Now I'm not saying this is a bad thing, as there certainly were far more important features to include/improve when Haswell was being developed. I had just expected that as Fermi got OpenGL 4.5, that the OpenGL 4.4 extensions do not require any actual hardware support.
The missing extension seems to be GL_ARB_bindless_texture – any connection to DX12 resource binding?
Now I'm not saying this is a bad thing, as there certainly were far more important features to include/improve when Haswell was being developed. I had just expected that as Fermi got OpenGL 4.5, that the OpenGL 4.4 extensions do not require any actual hardware support.
The missing extension seems to be GL_ARB_bindless_texture – any connection to DX12 resource binding?
GL_ARB_indirect_parameters is also problematic. However DX12 ExecuteIndirect is a superset of it, and Haswell can emulate it just fine.
The only real hardware functionality differences between Haswell and Broadwell are likely related to ARB_sparse_texture. Haswell's limited VA makes a useful implementation of sparse textures difficult (i.e. the virtual texture size limitations are typically too severe to be very useful). ARB_bindless_texture is likely not supported on either of them, but I don't think bindless textures are required for GL 4.4 (Fermi wouldn't support 4.4 then) or even 4.5? For that matter, I don't think sparse textures are required by any current GL either.
The difference here may not even be related to hardware architecture.
Infinisearch
Veteran
Does anyone know if Intel GPU's have a dedicated frame and depth buffer cache? Or if the depth or frame buffer reads/writes go through the standard cache?
Both - they each have small dedicated caches which are themselves backed by the GPU's L3, then the SoC's LLC/eLLC.
Infinisearch
Veteran
Do you know there sizes off-hand or know a document where I can find such info?
edit - well I found one data point in the docs in the OP.
Last edited:
Similar threads
