NVIDIA、推論特化チップ「Groq 3 LPU」発表　「Vera Rubin」と組み合わせてAI性能最大35倍に

�@��NVIDIA��3��16��i��n��ԁj�A�N��C�x��g�uGTC 2026�v�ŁA��AI�C��t��v��b�g�t�H�[��uNVIDIA Vera Rubin�v�̖{�i�I�ȗʎY�̐��ɓ��Ɣ��\��B�V��AI��_�ɓ��`�b�v��t�@�~��[�ɓ��A��ڐ��i�̓p�[�g�i�[�e�Ђ��2026�N�㔼�ł̒񋟂�\�肷��B

NVIDIA�̃W�F��X��E�t�A��CEO

�@Vera Rubin�́AGPU�uRubin�v�ACPU�uVera�v�A�uNVLink 6�v�X�C�b�`�ASuperNIC�uConnectX-9�v�ADPU�uBlueField-4�v�AEthernet�X�C�b�`�uSpectrum-6�v�ɁA��V��ɓ��ꂽLPU�iLanguage Processing Unit�A��ꏈ��j�b�g�j�́uGroq 3 LPU�v��7�`�b�v�ō\��B��𓋍ڂ��5��̃��b�N��1��AI�X�[�p�[�R��s��[�^�Ƃ��ċ@�\��݌v��B

�ŐVAI�X�p�R��uVera Rubin�v��7�̃`�b�v�A5�̃��b�N�ō\��

�@��C��́uVera Rubin NVL72�v��b�N�́A72��Rubin GPU��36��Vera CPU��NVLink 6�Őڑ��AConnectX-9 SuperNIC�A��BlueField-4 DPU�𓝍��BMoE�A�[�L�e�N�`��f��̃g��[�j��O�͐��uBlackwell�v��4��1��GPU��őΉ��ł��A��_�X��[�v�b�g�̓��b�g��ő�10�{�A�g�[�N��̃R�X�g��10��1�ɂȂ�Ƃ��Ă��B

�@�G�[�W�F��e�B�b�NAI�̋��w�K��S��uVera CPU��b�N�v��256��Vera�𓋍ڂ��BVera��NVIDIA��А݌v��CPU�ŁAArm�x�[�X�́uOlympus�R�A�v��88��ځB�]��̂��̂Ɣ�ׂď��2�{�A50%��Ă��ق��A1�R�A��2�̃^�X�N��s�ł��B��LPDDR5X�ŁA�ő�1.2TB�̑ш敝��Ƃ��B

�uGroq 3 LPU�v�̈З�

�@Groq 3 LPU�́ANVIDIA��2025�N12��ɒ��𔭕\��AAI�A�N�Z��[�^�[��|��X�^�[�g�A�b�v��Groq�Ƃ̔�Ɛ胉�C�Z��X�ɂ��́B1�`�b�v��̉��Z��\��1.2PFLOPs�iFP8�j�ŁA500MB��SRAM��B��ш��150TB/s��B�Ȃ��AGPU��Rubin�́A288GB�Ƃ��e�ʂ�HBM4��𓋍ڂ��A��ш��22TB/s�A50PFLOPs�̉��Z��\��B

Groq�̃��C�Z��X�ɂ��uGroq 3 LPU�v�`�b�v�i�E�j

�@��V��ɉ��uGroq 3 LPX��b�N�v�́A256��Groq 3 LPU��ڑ��邱�Ƃ�128GB�̃I��`�b�vSRAM�ƁA40PB/s�̃��ш敝��BVera Rubin NVL72�Ƒg�ݍ��킹�邱�ƂŁA��K��b�g��̐��_�X��[�v�b�g�͋��f��Blackwell�Ɣ�ׂčő�35�{�ɏ��Ƃ��Ă��A��p��[�^�K�͂̃��f��ƍő�100��g�[�N��̃R��e�L�X�g��ɍœK��Ă��Ƃ��B

Groq 3 LPU��~��l�߂�ꂽ��b�N�uGroq 3 LPX�v

Groq 3 LPU

��K��b�g��̐��_�X��[�v�b�g

Vera Rubin��Groq 3 LPU��g�ݍ��킹��ƃX��[�v�b�g��ő�35�{��

��ɂ��AI�x��_�[�̎��v�@��ɂ��Ȃ��Ƃ��Ă��

�@AWS�AGoogle Cloud�AMicrosoft Azure�AOracle Cloud Infrastructure�Ȃǂ�Vera Rubin�x�[�X�̐��i��N�㔼�ɒ񋟂��v��BAnthropic�AMeta�AMistral AI�AOpenAI��Vera Rubin�̗̍p��Ă��Ƃ��Ă��BDell Technologies�AHPE�ALenovo�ASupermicro�Ȃǎ�v�V�X�e��[�J�[��T�[�o��i�̒񋟂�\�肷��B

�֘A�L��

NVIDIA�A��e�B��Thinking Machines Lab�ɑ�K�͏o��@Vera Rubin��AI�J��
�~��E��e�B��Thinking Machines Lab�́ANVIDIA�ƕ��N�ɂ킽��헪�I��g��B1GW�K�͂̎��AI�V�X�e��uVera Rubin�v�𓱓��A�Ő�[��f��̊w�K�ƃJ�X�^�}�C�Y�\��AI�v��b�g�t�H�[��\�z�𐄐i��BNVIDIA�͓��Ђ֑�K�͏o��{��B
NVIDIA�A��AI�v��b�g�t�H�[��uRubin�v�ʎY�J�n�@��_�R�X�g��10��1��
NVIDIA�́A��AI�v��b�g�t�H�[��uRubin�v�̗ʎY�J�n�𔭕\��BGPU��CPU�Ȃ�6�̐V�`�b�v��gextreme codesign�h�ň�̉��B��Blackwell��Ő��_�R�X�g��10��1�ɒጸ��B2026�N�㔼�ɒ񋟊J�n�\��ŁAOpenAI��xAI�Ȃǎ�v�e�Ђ��č̗p��\��B
NVIDIA�̃t�A��CEO�A�u��_�̉��҂�Blackwell�v�@��Rubin�ւ̎��M��
NVIDIA��2025�N11��`2026�N1��Z�́A��㍂��O�N��73��̖�681��h��Ɖߋ��ō��X�V��B��v��94��Ƌ}��B�f�[�^�Z��^�[��傪��A��uBlackwell�v�ւ̋��v��p��Ă��B�t�A��CEO�́A��^AI�̕��y�ɔ��A�v�Z�͂��ڎ��v�𐶂ށuAI�t�@�N�g��[�v�ւ̓��͍��Ƌ��B
$NVIDIA�ACoreWeave�ɖ�3100��~�o��@��uRubin�v��Ղ�AI�f�[�^�Z��^�[��5GW�K�͂ō\�z��$ NVIDIA�ACoreWeave�ɖ�3100��~�o��@��uRubin�v��Ղ�AI�f�[�^�Z��^�[��5GW�K�͂ō\�z��
NVIDIA�́AAI�N��E�h��CoreWeave��20��h��o��A��g��B2030�N�܂ł�5GW�K�͂�AI�f�[�^�Z��^�[�\�z��ڎw��A��́uRubin�v�v��b�g�t�H�[��D�擱��B
Google�̌�TPU�J��҂�NVIDIA��@�n�Ƃ��Groq�̐��_�Z�p��C�Z��X�_��
NVIDIA�́ALLM�̏��ɓ��Ǝ��̃v��Z�b�T�uLPU�v��|��Groq�ƁAAI��_�Z�p�̔�Ɛ�I��C�Z��X�_��B��Google��TPU�J��҂Ƃ��Ēm��W��i�T��E��XCEO��v��o�[��NVIDIA��肵�A�Z�p�J��ƕ��y�𐄐i��BNVIDIA��Groq�̐��_�Z�p��荞�݁AAI�C��t��̋��͂��_��B