StorageReview.com

11.1K posts

StorageReview.com banner
StorageReview.com

StorageReview.com

@storagereview

https://t.co/Sd7aWQerOs offers news and in-depth reviews of the entire enterprise IT stack from AI to end user computing and everything in between.

Cincinnati, OH Katılım Kasım 2011
247 Takip Edilen13.9K Takipçiler
StorageReview.com
StorageReview.com@storagereview·
Sizing math: 128GB SRAM per rack across 256 chips. GPT-OSS 120B? One rack. DeepSeek R1? ~5 racks. GLM-5 in BF16? 10+ racks. The front-panel C2C ports chain racks for models that exceed single-rack capacity. Shipping H2 2026. Full deep-dive with diagrams and topology analysis: storagereview.com/news/nvidia-gr… @NVIDIAAI
English
0
0
1
159
StorageReview.com
StorageReview.com@storagereview·
This also solved Groq's pre-acquisition bottleneck. Standalone Groq struggled with long-context models because KV cache ate into LPU capacity. Split the workload: attention on Rubin's 288GB HBM4, FFN on Groq's 150 TB/s SRAM. The LPU never touches the KV cache.
English
1
0
0
145
StorageReview.com
StorageReview.com@storagereview·
🧵 GTC 2026 is in the books. Here's everything you need to know about the most interesting piece of the Vera Rubin platform: NVIDIA Groq 3 LPX.
English
1
0
4
252
StorageReview.com
StorageReview.com@storagereview·
NVIDIA Groq 3 LPX at GTC 2026: each LPU hits 150 TB/s memory bandwidth via SRAM vs ~22 TB/s on Rubin's HBM4. That gap is why LPUs offload FFN layers from Rubin GPUs to boost token throughput. 8 LPUs per 1U unit, 315 PFLOPS across the full 256-chip rack. Available 2H 2026. @NVIDIA #AI #Groq #GTC2026
StorageReview.com tweet mediaStorageReview.com tweet media
English
0
11
65
4K
StorageReview.com
StorageReview.com@storagereview·
NVIDIA pulled two cold plates at GTC 2026 to show what's underneath the Vera CPU board. Each board carries 2 Vera CPUs. Per CPU: 88 custom Olympus cores with SMT, 1.2 TB/s memory bandwidth, 1.5TB SOCAMM LPDDR5X, and 1.8 TB/s NVLink-C2C. The full Vera CPU Rack is a liquid-cooled MGX system scaling to 256 Vera CPUs, 400TB of LPDDR5X memory, 64 BlueField-4 DPUs, Spectrum-X Ethernet, and 2x performance per watt versus x86. Purpose-built for reinforcement learning workloads. @NVIDIAAI #AI #VeraCPU #datacenter #GTC2026
StorageReview.com tweet mediaStorageReview.com tweet mediaStorageReview.com tweet mediaStorageReview.com tweet media
English
2
19
170
14.2K