And yes if you're wondering whether or not techpowerup has also made similar changes to other Gen 11+ graphics entries in their database, I'm glad to show you that yes indeed they have as you may see here and here. If you click on the techpowerup link that I put in my original post, you can see now that whoever maintains the GPU database there has completely removed the FP64 (double) performance section of the "Theoretical Performance" numbers for the GPU, which makes sense as fp64 compute emulation is a Linux-exclusive feature (for now) for Intel Gen 11+ GPUs/iGPUs.
#Fp64 graphics drivers
So before I go and contact sites like cpu-monkey and techpowerup and suggest to them if they can change their GFLOPs numbers for FP64 performance of various Gen 11+ Intel GPUs, are my own findings of the extremely abysmal FP64 performance of Intel Gen 11+ GPUs valid, or have I been missing out on some secret proprietary Intel graphics drivers that can perform miracles and pump out hundreds of GFLOPS if not TeraFLOPS of FP64 performance via emulating FP64 support on consumer grade GPUs that have zero hardware FP64 cores in them?Ĭlarification also techpowerup has corrected their FP64 compute numbers So when techpowerup came along and claimed that the Arc A770 offers 4+ TFLOPs of FP64 I thought oh finally a great General Purpose GPU for a reasonable price until I discovered that Arc GPUs don't even have any fp64 cores in them!
#Fp64 graphics driver
Sure, I could go out and by a used Radeon VII for like $550 right now if I was absolutely desperate for significantly more FP64 performance than what my current RX 6800 has to offer, but given its aging Vega architecture who knows how long AMD will support it with their ROCm stack before they pull the plug on it like they did with driver support for GCN 3 and older consumer GPUs. for GPGPU purposes, as not all of us can afford to purchase cloud subscriptions to high-performance data center GPUs let alone own a datacenter GPU like the MI210, the A100, or Intel's own Ponte Vecchio when that comes out. prosumers) relatively accurate compute capability info of various GPUs so that people can make as informed of decisions as possible when shopping for GPUs esp.
#Fp64 graphics software
Is it just me, or has Intel removed fp64 FPUs entirely from their Gen 11+ graphics solutions (including for their ARC GPUs, as confirmed here, their integrated Gen 12 graphics solutions such as Intel UHD 770, as demonstrated by a lack of the cl_khr_fp64 extension here, AND Intel UHD Graphics for 11th Gen Intel Processors, which also includes the integrated graphics for the 11400H in the laptop which I own, as shown by lack of OpenCL "Double-precision Floating-point support" shown in attached terminal output image for the integrated graphics on my 11400H)? If FP64 FPU units simply aren't present on these graphics solutions and support for fp64 has to be software-emulated via these instructions, then via my testing I've done using mixbench and CLBlast the FP64 compute capability in FLOPS is only ~1/160th - ~1/128th of the FP32 compute capability (please see other two attached images for proof), which completely contradicts the 1 to 4 ratio of FP32 to FP64 compute GFLOPS as claimed by sites like cpu-monkey and techpowerup (EDIT: techpowerup JUST updated their compute capability stats for their Intel Gen 11+ graphics solutions to more accurate numbers).Īnd while I completely understand that FP64 use cases amongst average consumers are pretty niche as compared to the vast majority of people who shop graphics cards, given how there's readily available free software out there such as PyTorch that takes advantage of FP64 compute capabilities on various processors (take Double Tensors in PyTorch for example), I believe that sites like cpu-monkey and techpowerup should provide the average consumer (esp.
0 kommentar(er)
