vickieGPT

🏳️‍⚧️ System Hacker.

[vickieGPT @yyw2000]

https://t.co/lAHVjghGN8

[vickieGPT @yyw2000]

World first agentic RISCV XPU https://t.co/pwbF5OWdag

[Basic Apple Guy @BasicAppleGuy]

macOS California Release Location Map

[vickieGPT @yyw2000]

@bboczeng 分享一下我们的agentic riscv agi cpu？

[Datou @Datou]

大模型的参数量是如何确定的，比如 1t-a35b。首先有一些刚性的约束，比如推理成本可负担，gpu集群可以支撑训练，然后从小参数量逐步扩大（不要走捷径），看scaling ladder是不是持续有收益，看越来越大的参数规模下预训练和强化学习会不会崩，最后停在了 1t-a35 ，在一些测试项目上拿到了满意的成绩。

[forecho📈 @caizhenghai]

$AAPL 苹果才是最后的赢家

[Datou @Datou]

训练半天发现不如 deepseek v4

[finánce @racetrack275]

This is a HBM4E wafer from Computex (photo koreajoongangdaily)

Die is 1c 32Gb, minimum 140mm2 size. Likely 145-150, since HBM4 JEDEC package size is 155.6mm2. HBM eats fab cap as die size goes up every gen. There is no memory supply coming pre-2030 that offsets this.

[vickieGPT @yyw2000]

https://t.co/aIXwhTfMSE 后生可畏

[Macro_Lin ｜ 市场观察员 @LinQingV]

这个月开始退订Claude 200刀的订阅，全面切到Hermes走DeepSeek V4 pro api. Anthropic越来越垃圾啦.

[vickieGPT @yyw2000]

Intel不太行。

[StorageReview.com @storagereview]

Vera talk is heating up, so resharing these from GTC: NVIDIA's Vera CPU 8-way compute tray up close.

Liquid-cooled, SOCAMM LPDDR5X packed around the CPUs, with 1.2 TB/s memory bandwidth and 1.8 TB/s NVLink-C2C on the spec board.

Now the servers are getting close to landing. HPE brought the ProLiant DL394 Gen12 to COMPUTEX, and Dell is bringing its own Vera systems to market.
#NVIDIA #Vera #HPE #Dell @NVIDIAAI

[vickieGPT @yyw2000]

https://t.co/QtPk6dBtu9 这东西能颠覆CPU？太蠢了。

[Wccftech @wccftech]

Intel intros Diamond Rapids "Xeon 7" CPUs built on 18A-P: packing up to 192 cores, 16-channel memory, PCIe Gen6 support, launching in 2027. https://t.co/jm6G3BIg9W

[SemiAnalysis @SemiAnalysis_]

BREAKING NEWS: Jensen just announced WINDOWS on NVIDIA ARM consumer PC. Unlike Apple transition to ARM M1 with Rosetta 2 transpiler, we are not convinced that Windows on NVIDIA ARM will work well.

[中本麻三 @harry03994688]

intel is more like AMD on it's Xeon 6 + clearwater forest:
12 x compute tile intel 18A
3 x active base tile: intel 3
2 x IO tile: intel 7
12 x EMIB tile: EMIB 2.5D
4 x DDR5 memory base tile

intel Xeon 6:
3 x compute tile:
2 x io tile:
EMIB tiles

the tiles number is more than 400% increased.

this is really interesting.

chiplet designs are going to everywhere: except the NVIDIA GPU.

[中本麻三 @harry03994688]

you can't have the high performance and low power consumption at the same time. this is commond sense, there is no magic in compute.

ARM VS X86

as INTEL and AMD's chiplet designs:
huge cores / ~ > 2 times users can be servered than one VERA CPU.

this is simple math. ARM won't take all. even 50% is diffcult in server CPU market.

NVIDIA VERA: 88
Intel Xeon 6+: 288
Intel Xeon 6: 288
AMD Turing:192
AMD Venice: 256

[EXO Labs @exolabs]

We made it to the keynotes! 🎉

[vickieGPT @yyw2000]

笑死

[vickieGPT @yyw2000]

老黄这个L3 Size不行啊。