{"id":381602,"date":"2026-10-21T01:52:00","date_gmt":"2026-10-21T08:52:00","guid":{"rendered":"https:\/\/cms-articles.softonic.io\/en\/?p=381602"},"modified":"2026-10-21T05:24:27","modified_gmt":"2026-10-21T12:24:27","slug":"groqs-next-lpu-will-use-3d-stacked-dram-bigger-ai-models-ahead","status":"publish","type":"post","link":"https:\/\/cms-articles.softonic.io\/en\/groqs-next-lpu-will-use-3d-stacked-dram-bigger-ai-models-ahead\/","title":{"rendered":"Groq\u2019s next LPU will use 3D stacked DRAM: bigger AI models ahead"},"content":{"rendered":"

Groq is getting its next LPU ready, and the big architectural change comes down to memory<\/strong>. The company is moving from SRAM to 3D-stacked DRAM so it can fit larger AI models. Groq also hasn\u2019t been acquired. In late 2025, it signed a non-exclusive licensing deal with NVIDIA, reportedly worth about $20 billion, and remained independent. For anyone watching the hardware closely, the trade-off is pretty direct: less of the extreme bandwidth Groq\u2019s current design is known for, a lot more memory capacity from DRAM.<\/p>\n

\r\n
\r\n
\r\n
\r\n \"GPU-Z\"\r\n <\/div>\r\n
\r\n GPU-Z<\/span>\r\n Download<\/a>\r\n <\/div>\r\n
\r\n <\/path><\/path><\/text><\/svg>\r\n <\/div>\r\n <\/div>\r\n
\r\n <\/span>\r\n <\/div>\r\n
\r\n \r\n <\/div>\r\n <\/a>\r\n <\/div>\r\n<\/div>\n\n\n\n

On the current chip, Groq uses 500MB of SRAM<\/strong>, roughly 150TB\/s of bandwidth, and 1.20 PFLOPs of FP8 per chip. Scale that to 256 chips and the total lands around 128GB and 640TB\/s. DRAM changes the picture by giving Groq much more capacity to work with. And to clear up a point that\u2019s easy to get wrong, NVIDIA did not buy Groq. It signed that late-2025 non-exclusive licensing deal, reportedly worth about $20 billion, and hired several key Groq people, including founder Jonathan Ross, while Groq continued operating independently.<\/p>\n\n

What seems most likely is TSV-connected, HBM-style<\/strong> stacked DRAM, which would raise density and should also help with power efficiency. Beyond that, a lot is still unknown. Timing, implementation details, and what any of this means for customers haven\u2019t been spelled out yet.<\/p>\n\n

If you follow AI serving hardware, keep an eye on this one. Groq\u2019s LPX racks are already part of NVIDIA\u2019s Vera Rubin ecosystem, reportedly reached full production in August, delivered around 3,400 tokens per second<\/strong> with NVL72 servers, and now sit alongside GroqCloud and Groq\u2019s role as an NVIDIA Cloud Partner.<\/p>\n\n

There\u2019s nothing to download yet. Still, expect the next developments to show up through Groq\u2019s data center and cloud platforms as model sizes keep climbing.<\/p>\n

\r\n
\r\n
\r\n
\r\n \"CPU-Z\"\r\n <\/div>\r\n
\r\n CPU-Z<\/span>\r\n Download<\/a>\r\n <\/div>\r\n
\r\n <\/path><\/path><\/text><\/svg>\r\n <\/div>\r\n <\/div>\r\n
\r\n <\/span>\r\n <\/div>\r\n
\r\n \r\n <\/div>\r\n <\/a>\r\n <\/div>\r\n<\/div>\n\n","protected":false},"excerpt":{"rendered":"

Groq is getting its next LPU ready, and the big architectural change comes down to memory. The company is moving from SRAM to 3D-stacked DRAM so it can fit larger AI models. Groq also hasn\u2019t been acquired. In late 2025, it signed a non-exclusive licensing deal with NVIDIA, reportedly worth about $20 billion, and remained … Continue reading “Groq\u2019s next LPU will use 3D stacked DRAM: bigger AI models ahead”<\/span><\/a><\/p>\n","protected":false},"author":9351,"featured_media":381601,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":"","wpcf-pageviews":0},"categories":[1015],"tags":[],"usertag":[],"vertical":[],"content-category":[6771],"class_list":["post-381602","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news","content-category-ai"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t\n\t\n\t\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t