AGF @XpeaGPU

GPU Veteran High-NA EUV wafer Joined March 2022

Tweets

1K
Followers

2K
Following

266
Likes

23K

AGF @XpeaGPU

4 months ago

@dnystedt @Mikeair899 Please tell us how the reporter knows. Who in hell at TSMC or Nvidia will leak such information, 2 weeks after GTC and Rubin Ultra roadmap + servers introduced.,,

0 0 0 102 0

View Details

AGF @XpeaGPU

4 months ago

@RehanYawar84469 internal product. they can claim whatever suits their marketing

1 0 0 44 0

View Details

AGF @XpeaGPU

4 months ago

Typical Samsung out of touch propaganda. Number on PowerPoint. Single digit real world yields, even years after introducing a new tech. Cough GAA cough

Samsung to Adopt 'Fork Sheet' for 1nm Process… All-Out Effort to Win Over Big Tech Samsung Electronics' foundry division has set a target to introduce 1nm (nanometer; 1nm = one-billionth of a meter) process technology by 2031. The 1nm node is dubbed a "dream semiconductor"

9 12 119 21K 29

3 1 19 4K 2

View Details

AGF @XpeaGPU

4 months ago

@RehanYawar84469 On what chip? a tiny DRAM die? a simple design like these Chinese mining ASICs with large SRAM arrays, regular structures and few mixed signal/analog blocks? Truth is that nobody trust Samsung marketing. Take everything they say with a truckload of salt

1 0 0 70 1

View Details

AGF @XpeaGPU

4 months ago

@canyoudugit8 @TickerSymbolYOU yeah Nvidia just went from $500B orders in the pipeline to $1T. nothing to be excited about. You are right

2 0 1 109 0

View Details

AGF @XpeaGPU

4 months ago

@HyperTechInvest When I saw the Korean articles ditching Micron, I immediately call the BS. Same story happened with H200. No idea why people is always wrong with Micron

0 0 14 2K 0

View Details

AGF @XpeaGPU

4 months ago

So much for the naysayers, datacenters in space are a reality that can't be ignored anymore. And @elonmusk will be the winner. Exciting times

NIK @ns123abc

4 months ago

🚨 NVIDIA JUST ANNOUNCED DATA CENTERS IN SPACE Jensen Huang: >we’re going into space >we’ve already been out in space >announces “Vera Rubin Space-1” >"in space there's no conduction, no convection, just radiation" >"we have to figure out how to cool these systems in space"

156 380 3K 368K 675

1 0 4 816 0

View Details

AGF @XpeaGPU

4 months ago

As I said last month when I called BS the fake new Korean media, Micron is part of Rubin HBM4 platform.

Austin Lyons @austinsemis

4 months ago

JENSEN TO MICRON CEO: To my friends at Micron. Our partnership changed the world. Industry’s fastest HBM4 delivered to NVIDIA. Seems like Micron HBM4 is a go. $MU $NVDA @MicronTech @NVIDIAGTC

7 31 292 60K 28

0 0 5 866 0

View Details

AGF @XpeaGPU

4 months ago

@canyoudugit8 @TickerSymbolYOU you clearly don't understand the market. LPU is for small models that require ultra low-latency. No rack to rack here. In fact, no inference requires rack to rack. Scale-out is for training. Have a nice day dreaming of 5% market share AMD taking over Nvidia with MI455X NPU lol

3 0 2 149 0

View Details

AGF @XpeaGPU

4 months ago

@canyoudugit8 @TickerSymbolYOU if you cant understand that AMD NPU is a simple old school bunch of VLIW tiles that no bandwidth can compensate for it's inefficiency vs LPU novel software deterministic tensor-streaming engine optimized for inference, then let's agree to disagree and time will tell

0 0 1 90 0

View Details

AGF @XpeaGPU

4 months ago

LOL no, that's not how it works. 1/ KV cache transfer isn't the bottleneck in Rubin + LP30. BlueField-4 STX boost throughput by 5x. Actual inter-LPU bandwidth is 2.5 TB/s bi-di per chip. 2/ AMD 15 TB/s internal? yeah nice for integrated decode, but LPU deterministic arch (static scheduling, not Von Neumann waste) wins in overall latency & speed. 3/ Benchmarks already show LPUs crush NPUs and with much less bandwidth than the Rubin solution.

2 0 2 104 0

View Details

AGF @XpeaGPU

4 months ago

@restricted_ptr @IanCutress Time to market. It's a last minute addition. Attach a NVLink C2C to the LPU and call it a day

0 0 1 35 0

View Details

AGF @XpeaGPU

4 months ago

Groq journey at Nvidia: LP30 is a quick last minute addition. LP35 blends into Nvidia software ecosystem (NVFP4). LP40 adds Nvidia interconnect (NVLink). LP-Next will co-packaged on-die.

𝐷𝑟. 𝐼𝑎𝑛 𝐶𝑢𝑡𝑟𝑒𝑠𝑠 @IanCutress

4 months ago

OK here we go. Same architecture, scaled up. 500 MB SRAM per chip. Jensen saying up to 25% of your datacenter could be groq. 8 way systems for 4 GB SRAM. Use dynamo for attention Decode only. Working over ethernet on special mode to half latency. Samsung LP4X. Ship in 2H/3Q.

12 44 446 67K 129

0 4 20 3K 9

View Details

AGF @XpeaGPU

4 months ago

@canyoudugit8 @TickerSymbolYOU 1/ LPU is different than NPU. Apples vs oranges 2/ LPU is much faster (low latency software deterministic execution with static scheduling vs classic Von Neumann arch that is less efficient for inference) 3/ AMD NPU lacks Nvidia ecosystem, it will go nowhere

1 0 3 88 0

View Details

AGF @XpeaGPU

4 months ago

@paurooteri In VR300 (Vera Rubin Ultra) we have 2x 1TB GPU HBM + 2TB CPU LPDDR5X (=4TB total) as coherent memory pool for KV Cache (thanks to NVLink-C2C) per compute blade. That's an increase of 4.6 times memory capacity in a single generation (vs GB200).

1 0 6 756 2

View Details

AGF @XpeaGPU

4 months ago

That's simply a series of x86 CPUs with NVLink Fusion

ImNotHarsh | 📈💸 @imnotharsh

4 months ago

BREAKING: $INTC Intel Announces they will be at $NVDA Nvidia GTC 2026, next week, March 16 - 19th. Rumors suggest Intel will be releasing their new chips with Nvidia on x86 architecture. For data centers: Intel will build Nvidia-custom x86 CPUs that Nvidia will integrate into

6 14 143 16K 16

0 1 15 2K 4

View Details

AGF @XpeaGPU

4 months ago

N3 capacity was set when Nvidia wasn't targeting top nodes & was smaller TSMC client than Apple. It changes with Feynman: Nvidia is lead customer for A16 node, booking all early capacity like Apple once did.

Ben Pouladian @benitoz

4 months ago

Called it @dylan522p just dropped the numbers: NVIDIA owns 70% of TSMC N3. Google Cloud and DeepMind can’t agree on how many TPUs to give Anthropic. Hard to be generous with chips you can’t guarantee for yourself. One CEO controls delivery. That’s where the power sits. $nvda

12 12 157 41K 63

1 2 20 2K 2

View Details

AGF @XpeaGPU

4 months ago

Honestly, I thought this paper was written by captain obvious. No disrespect but memory bandwidth/capacity were always the bottleneck. But we live in real world, not lalaland, so things like manufacturing capacity and supply chain restrictions made the ASICs/TPU/GPUs what they are today. Unless you believe that all these silicon engineers are stupid...

1 0 13 5K 0

View Details

AGF @XpeaGPU

4 months ago

Corrections: 1. Old news. it was done 2 months ago. 2. Timing adjustments are done until very late in the qualification process. Each cycle is the same and each cycle same outlets publish FUD. Boring. 3. BS. Samsung will love to take more market share with Nvidia. 4. Taken by others (Foxcom and Quanta). 5. no comment. I don't follow DDR4.

0 0 3 486 0

View Details

AGF @XpeaGPU

4 months ago

@jukan05 "Groq, AI Chip Startup Acquired by NVIDIA, to Ramp Up Production at Samsung Foundry" When the title is already wrong, how can you trust this article?