3D Stacked DRAM for Inference: d-Matrix's Sudeep Bhoja
Notas do episódio
The fundamental physics of data movement are forcing a shift to new memory architectures for AI inference. Sudeep Bhoja, co-founder and CTO of d-Matrix, joins Vik to detail the company's journey from a non-consensus bet on memory-bound inference to a full-rack 3D-stacked DRAM solution. Sudeep explains the pivot from analog to digital in-memory compute, the system-level disaggregation strategies for their Corsair card, and how the next-generation Raptor chip integrates with Nvidia's ecosystem.
"Essentially, we got 16 times the capacity of SRAM, while we retained SRAM bandwidth."
— Sudeep Bhoja, d-Matrix
Presented by Crusoe: Fine-tune and serve open models without managing a single GPU. When we sat down with Crusoe’s product team, Austin asked: what will you do with my data? Their answer: They don’t touch or see it. Your training data is encrypted so even Crusoe’s own admins can’t see it, and the fine-tuned model is yours. Deploy on Crusoe in one click, or download and run it anywhere. Upload a dataset, pick an open model, and pay per token rather than per GPU hour. Run a few versions, compare them, and download the one that works best. Accounts are free, and get $5 in credit when you sign up. Try Serverless Fine-Tuning: https://semidoped.com/crusoe
Key Takeaways:
- The core problem in AI inference is a 1000x energy gap. Compute costs femtojoules per operation, but moving data from memory costs picojoules per bit.
- d-Matrix's first analog in-memory compute chip failed because it couldn't maintain 32-bit accumulation precision, forcing a pivot to a digital architecture.
- The Corsair card's digital in-memory compute delivers 150 TB/s of on-chip SRAM bandwidth — over 8x more than an Nvidia Blackwell GPU's 18 TB/s HBM.
- The next-gen Raptor chip stacks a logic die on a DRAM die, delivering 16x the capacity (32 GB) of the prior SRAM design while retaining ~100 TB/s bandwidth.
- This 3D stacking is over 10x more energy-efficient than HBM4, consuming just 0.37 pJ/bit versus an estimated 5 pJ/bit for an HBM4-plus-GPU system.
- Choosing the stacking order is a 'pick your poison' trade-off. d-Matrix put logic-on-top for better thermals, accepting a harder power delivery challenge through the DRAM.
- The go-to-market product is a full rack with 144 Raptor packages, aggregating 2.3 TB of high-bandwidth memory designed to drop into Nvidia's VR rack infrastructure.
This episode is supported by Meter. You have the GPUs and the power. Now who builds the network, and who do you call when it breaks? Meter designs, deploys, and runs frontend, storage, and management networks on its own switches and its own network OS. One call reaches the team that designed the hardware and wrote the software. https://semidoped.com/meter
Chapters:
0:00 Intro: The Cost of a Token
3:28 The 2019 Bet on Inference
7:43 Nighthawk: Analog In-Memory Compute
12:08 Why Analog Failed
19:02 Corsair: Digital In-Memory Compute
26:04 System Strategy: Speculative Decoding
37:05 Raptor: 3D Stacked DRAM
39:40 The 'Free Lunch' of 3D Stacking
45:00 Trade-off: Logic-on-Top vs DRAM-on-Top
54:14 Go-to-Market: The Nvidia VR Rack
1:03:17 The Future: Lightning & 4-High Stacks
1:12:45 The End State of Inference
Follow Chipstrat:
Newsletter: https://www.chipstrat.com
X: https://x.com/chipstrat
Follow Vik:
Newsletter: https://www.viksnewsletter.com/
X: https://x.com/vikramskr
Follow Semi Doped:
Get more of Austin and Vik daily, free: https://daily.semidoped.com/