I've noticed one such workunit on one of my hosts.
It takes 11200 seconds (3 hours and 7 minutes) (~44%) more than a GERARD_FXCXCL12_LIG to process, while it gaves 4125 less credits.
It has a very low GPU (GTX 980) usage (~86%) even on my WinXPx64 host.

It takes 11200 seconds (3 hours and 7 minutes) (~44%) more than a GERARD_FXCXCL12_LIG to process, while it gaves 4125 less credits.

You're GIANNI time per step is 2.937ms. 45.2% faster than my 5.355ms.
For both of us: the GIANNI runtimes are nearly 1.5 GERALD's WU's.
On a 4 lane slot 970 at 1519MHz - I completed a GIANNI in 18.5hr which is 6hr more than my (1506Mhz) GERALD.
My GIANNI 48% slower than you're 980. (Is the 980 on x16 lane and >1450Mhz?)

It has a very low GPU (GTX 980) usage (~86%) even on my WinXPx64 host

W8.1 only managed 73%. GPU was around 117W (7400 WU total BTU > 400BTU/h > 6.6BTU/min.)

(Is the 980 on x16 lane and >1450Mhz?)

My GTX980s are on PCIe3.0 x8 lanes, GPU clock is 1420MHz.
At the moment my hosts are on vacation, because we have daytime outside temperatures around 37°C (99°F), and now I have 29.5°C (85°F) in my room (no AC).