🔥 Hot Chips
Ranked by importance, recency, and FP16/BF16 performance (sparse figures, consistent with the comparison table).
- MI455XAMDCDNA 5432 GB HBM4
23.3 TB/s10,100UndisclosedTraining - Rubin R200NVIDIARubin288 GB HBM4
22 TB/s~9,000*1800–2300WTraining - B200NVIDIABlackwell192 GB HBM3e
8 TB/s5,0001000WTraining - B300 UltraNVIDIABlackwell Ultra288 GB HBM3e
8 TB/s4,0001400WTraining - Ascend 960DTHuaweiDa Vinci v6288 GB HBM
9.6 TB/sUndisclosed†700WTraining - MI355XAMDCDNA 3.5288 GB HBM3e
8 TB/s2,5001400WTraining - TPU IronwoodGoogleTPU v7192 GB HBM
7.4 TB/s2,307600WTraining - Ascend 920HuaweiDa Vinci v496 GB HBM3
4 TB/s1,800400WTraining - H200 SXMNVIDIAHopper141 GB HBM3e
4.8 TB/s1,979700WTraining - H100 SXMNVIDIAHopper80 GB HBM3
3.35 TB/s1,979700WTraining - Ascend 950DTHuaweiDa Vinci v5144 GB HiZQ 2.0
4 TB/s~500~500WTraining - MI300XAMDCDNA 3192 GB HBM3
5.3 TB/s1,307750WTraining - Ascend 950PRHuaweiDa Vinci v5128 GB HiBL 1.0
1.6 TB/s~500~400WInference - MLU690CambriconMLUarch 04196 GB HBM3
3.35 TB/s700+~500WTraining - Ascend 910CHuaweiDa Vinci 3.0128 GB HBM
3.2 TB/s800600WTraining - MTT S5000Moore ThreadsMUSA 4.080 GB
1.6 TB/s500300WTraining
FP16 in TFLOPS, sparse figures (consistent with the comparison table); * estimated; † FP16 undisclosed (FP4: 4 PFLOPS).
📰 Latest News
Industry news, product launches, technical deep-dives, and selection guides.View all articles →