AMD Radeon VII - Double Precision

[jlopz]

Does anyone know if the AMD Radeon Instinct MI50 is supported in FAH? The GPU listing shows several possible GPU references, but it's not clear. There's the MI25 and MI100, but I don't see anything like MI50.

[muziqaz]

It is Vega 20, double precision or not, it is supported by FAH. Output ranges from 1m to 2.6m PPD in OpenCL. Probably double that with HIP

[arisu]

I don't think GPU vendors "hobble" their FP64 precision. It's true that each double-precision unit will have a 1:2 speed ratio compared with each single-precision unit, but these cards include more single-precision units for FP32 than double-precision units for FP64 because single precision is very useful in games. You can't just combine two SP units to become one DP unit without a lot of extra silicon (it's not as simple as an integer ALU where you can just double their width snapping two together and wiring a trace between them for the carry). I think you can use an DP unit as two SP units, but it comes with a lot of caveats that make it only useful for some kind of vector arithmetic, but don't quote me on that.

A single Kepler GK110 SMX (Streaming Multiprocessor) has 192 SP units, 64 DP units, 32 special-function units, and 32 load/store units. I think that is the the Nvidia card with the best single-to-double ratio of 1:3 (192:64). But that's also a datacenter GPU. Consumer GPUs have closer to one double-precision unit for every 24 to 64 single-precision units.

This is fine for FAH because each DP unit is still operating at full speed (1:2 the speed of an SP unit) and because there are fewer calculations that need DP, you won't be saturating all the DP units until long after the SP units are all in use.

[muziqaz]

Yes, implementing pure double precision units in hardware requires a lot of extra transistors and space on the die. I'm not sure what AMD was smoking when they released Radeon 7, as I'm sure they have never got their money back from those cards.