To obtain in vivo human brain high resolution (<1ml) spectroscopic imaging data at ultra-high (7T) field, in clinically acceptable time (<10minutes), with reduced gradient demands.7T promises a 2.3 fold increase in SNR compared to 3T, which is greatly appealing when measuring small concentration metabolites in human brain. However, for fast spectroscopic imaging (SI) techniques using gradient readout waveforms, this also makes it 2.3 times more challenging because, even at the same spatial resolution, for the the same spectral width (SW) in parts-per-million (ppm), the spectral dwell time spDT=1/SW is 2.3 times as small as at 3T. Thus, the same k-space needs to be sampled 2.3 times as fast. The lower demands on the gradient system associated with the Rosette Spectroscopic Imaging1,2 (RSI) technique, due to the smoothly varying gradient waveform is, perhaps, the most important advantage of the RSI acquisition at ultra-high field (UHF), allowing this technique to achieve easier a greater spectral bandwidth than echo-planar SI (EPSI) or spiral SI (SSI). For the same 1cm in-plane resolution (IPR) and SW=1923 Hz (6.47ppm) achieved by flyback EPSI with two temporal interleaves3 (40 to 48 shots total), RSI can do it using 24 shots, no temporal (one) interleaves and a maximum gradient and slew-rate of Gmax=7.1mT/m and Smax=86mT/m, using settings previously described1 (Figure 1). Thus, RSI can be almost twice as fast, while using peak gradient and slew rates approximately half of the ones used by EPSI, which results in decreased eddy currents4. Furthermore, because data is collected and used for the entire signal decay readout, RSI does not suffer from up to a 30% loss in sensitivity as the flyback EPSI acquisition5. A 7T Siemens Magnetom PTX Step 2 system and an 8x2 transceiver array (2 rows, 8coils/row) was used. For transmission, the 16 coils were driven by 8 independent RF channels (amplitude, phase determined by RF shimming). The two rows are paired superiorly-inferiorly, with each two-coil pair driven by a single independent RF channel using an equal amplitude 30⁰ splitter. RF shimming was used to optimize two B1 + distributions; a ring distribution for outer volume suppression targeting the skull/skin for lipid suppression and a homogeneous distribution for excitation6. A 38cm ID shim insert coil (Resonance Research Inc.) was used for higher degree shimming with a non-iterative least squares B0 mapping algorithm7 (Bolero, B0 loop encoded readout). For NAA/Cr/Cho quantification, a SW=1923Hz (6.47ppm) and TE=40ms were used and for the dual echo J-refocused8 acquisitions SW=2778Hz (9.35ppm), with TE=17/34ms. Lipid/water/lipid inversion and repetition time TI1/TIW/TI2/TR=180/420/780/1500 ms. Data readout/repetition time: Tread/TR=320/1500 ms. A water reference with TR=0.39s was also collected for channel phasing/recombination. Data without a ring suppression was also acquired with TI1/TIW/TR=350/1200/3000ms6. Rosette trajectories were designed and data reconstructed as previously described1. For the 3D acquisition, the z-direction encoding was conventional. All data was automatically processed with an LCModel9 (http://www.s-provencher.com/pages/lcmodel.shtml) based pipeline.Using a BRAINO MRS phantom, the Bland-Altman agreement between RSI and phase encoded (PE) CSI for normalized NAA/Sum is within 7% of the mean (0.25) and average (SD) correlation r=RSI*CSI for spectra in all voxels (Schirda 2015) was 0.97(+/-0.03) (identical spectra for r=1) . When using the ring B1 lipid suppression in vivo, an average 72% greater sensitivity was obtained compared to the spatially non-selective (global) lipid inversion pulse. Human data was acquired as: a) Single slice 2D, 20x20 matrix (IPR=1cm) in as short as 0.6 minutes and up to a J-refocused 48x48 (IPR=0.4cm, 0.16mL nominal, 0.44ml effective voxel after spatial filtering, Figure 2) in 9.5 mins and b) 3D acquisition 20x20x12 with 0.4ml nominal resolution (1.25ml effective) in 7.2mins (Figure 3).High resolution (<0.5ml) 2D and 3D RSI data, with robust SNR, was collected at 7T in clinically acceptable time (less than 10mins). The high SNR of these fully encoded k-t space RSI acquisitions at 7T suggests that even higher spatial resolutions and/or a further decrease in scan time through the use of parallel imaging techniques is feasible without significant impact on data quality. The high order B0 shim insert allows for good quality spectral data in difficult to shim regions, such as temporal lobe, granted targeted data collection in those regions is performed. While we used an 8-channel pTx system to perform a B1 ring scalp lipid suppression, the RSI acquisitions described here could be done using a more typical setup, with receive only coil, by implementing a spatially non-selective lipid pulse. However, using the scalp suppression ring provides for shorter acquisition times and almost a two-fold increase in sensitivity.