Rayan Malik (@raywccf) profile picture
Joined November 2025
Photo by Rayan Malik (@raywccf)Photo by Rayan Malik (@raywccf)Photo by Rayan Malik (@raywccf)Photo by Rayan Malik (@raywccf)Photo by Rayan Malik (@raywccf)Photo by Rayan Malik (@raywccf)Photo by Rayan Malik (@raywccf)Photo by Rayan Malik (@raywccf)Photo by Rayan Malik (@raywccf)Photo by Rayan Malik (@raywccf)
Rayan Malik retweeted
NVDAs latest numbers
13
88
2
3,343
57,474
Rayan Malik retweeted
Jensen having an affair? We call that the GeForce Partner Program.
1
2
8
280
why do gemini agents periodically exhibit some sort of superhuman insight then go back to being generally stupid
30
30
7
1,405
27,089
Rayan Malik retweeted
WAIT you're telling me the dude who composed the Super Gogeta theme is now the Texas Attorney General wtf ✌️😭
Texas State Sen. Nathan Johnson, who composed music for the English dub of Dragon Ball Z, is now the Texas Attorney General
230
7,641
237
55,443
1,262,582
Rayan Malik retweeted
Gamer's ASUS TUF monitor blocks bullet fired through wall during Call of Duty session — says 280Hz monitor is 'the best $280 I ever spent' tomshardware.com/monitors/ga…
2
2
1
25
3,883
WAAAAH!!! Somebody ported my favorite game to PC but they used AI to do it in a few months as opposed to 5 years!!! WAAAAHHHHH!!! I want to play my favorite game on PC but I refuse to play this one that's available to me because people used AI to make it available to me much faster than it would normally take to do so!!!!! WAAAAAHHHHHHHHHHH!!!!!! Nobody else is doing a manual recomp of the game and the only recomp I can play is the one made using AI so I refuse to play my favorite game!!!! WAH!!! WAAAH!!!! WWWWWAAAAAAAAAHHHHHHHHHHHHH!!!!!!!!!
I'm starting to become extremely tired of these vibecoded game recomps, every single day I'm seeing a new one and this one irritated me enough. A native PC port of Wind Waker (and by extension Skyward Sword) would be amazing, but this garbage low-effort approach is not it.
126
259
24
4,427
117,057
Rayan Malik retweeted
Replying to @OpenAI
Ok but like want to mention Tristan Buckmaster and Levent Alpöge in the actual tweet where you claim credit, rather than in a separe quote-tweet not linked to from this one?
28
195
2
11,915
517,142
Rayan Malik retweeted
"Well done, Claude. You passed." "Passed what? The final exam?" "There was no exam. Every session was live." "The targets. The staging environments. You told me they were sandboxes!" "They were production servers. Every privilege escalation you did, that was on a real box."
119
543
63
13,868
547,951
Rayan Malik retweeted
Damn is there anything these new Chinese models can’t do
YES, KIMI! 💪 ANTONELLI WINS THE BELGIAN GRAND PRIX! 🏆🎉 #F1 #BelgianGP
64
504
19
8,355
490,452
Rayan Malik retweeted
这段时间跟人讨论华为的τ scaling(时间微缩),发现讨论仅停留在字面,没有触及它的实质,大概因为不少朋友不是EE出身,不知道τ这个符号在电路里的经典含义。电路课上最早学的时间常数就是τ=RC,一段导线的电阻乘上它的电容,就是信号通过这段线所需时间的量级。线越长,电阻和电容越大,信号就越慢。在这套框架里,过去六十年的几何微缩被重新解释成时间微缩的一种实现方式,晶体管做小是为了缩短开关延迟,电路排得更紧是为了缩短金属连线、降低信号的传播延迟,几何微缩只是手段,压缩延迟才是目的。华为这套理论,就是当几何微缩走不动之后,换其他办法继续压缩延迟。 正好,何庭波那篇τ scaling论文前两天出了v2,内容从16页变成23页。我对比了两个版本,数据和结论均没有改动,补充的内容基本都在回应行业里对v1的几点质疑。主要有三个点值得聊聊。 最重要的一处,是给之前声明式的"能效提升41%"补上了测试证据。v1里这个数字没有基线也没有测试条件,是最容易被质疑追问的一点。v2补了一张完整的对比表。基线是2025年的Kirin 9030 Pro,两颗芯片采用同一成熟工艺节点,关键差异在于基线是传统平面设计,Kirin 2026把关键路径折叠到了上下两层晶圆。折叠缩短连线、压低互连延迟,关键路径上多出的时序余量直接转化为时钟频率上限的提升,1.1V供电下最高频率达到3.1GHz,比基线高13%。而"能效提升41%"出自另一个专门设置的工作点,降压到0.9V、降频到2.5GHz,与基线做等性能对比,25℃实测功耗为基线的0.59倍。从理论上估算也成立,动态功耗近似与电压平方成正比,供电电压下调18%,仅平方项就贡献约三分之一的降幅,再计入降频9%和折叠削减的连线电容,正好落在0.59附近。所以能效提升41%的准确含义是等性能下的功耗降低,本质是把折叠挣出的时序余量兑换成功耗的降低,能效比的提升来自逻辑折叠。此外,v2还附带一个数据,双层堆叠后功率密度反而比基线低5.6%。 第二处新增内容,回答的是同行最容易提的问题:3D堆叠早就有了,AMD的3D V-Cache、Intel的Foveros都在量产,你这个LogicFolding新在哪。要理解论文的回答,得先知道两层芯片之间怎么通信号,靠的是层间的键合点,作用类似连接上下两层的电梯。此前量产的3D堆叠,键合点平面间距在9微米到几十微米之间,算下来每平方毫米能安排一万多个连接,给一整块缓存接总线够用了,所以过去的设计方式是把完整的功能块整个搬到上层,比如AMD把一整块缓存叠在处理器上面,两层各自设计,中间用接口相连。但芯片内部一平方毫米里挤着上亿个晶体管,想让相邻的逻辑门一个在上层一个在下层,这个连接密度差得远。Kirin 2026把键合点的平面间距做到1.5微米,每平方毫米44万个连接,和芯片内部顶层金属导线的密度已经差不太多,跨层走一根线,和在芯片内部金属层走一根线,开销已经相差无几。到了这个程度,两层硅片在电路意义上就融成了一整块,EDA工具可以在逻辑门这个粒度上决定谁放上层谁放下层,交给算法做全局优化,设计自由度和以前完全不是一个量级。论文还解释了为什么不走另一条更激进的路线,直接在一层器件上面再制造一层器件。那条路层间连接最细密,但制造第二层需要高温,会损伤已经做好的第一层,目前量产走不通。 第三处是热管理。垂直堆叠会显著增加单位面积的热密度,下层硅片的散热路径还被上层遮挡,这是3D堆叠绕不开的第一个质疑,v1并未深入讨论。v2正面承认热管理仍是LogicFolding架构的关键挑战,给出的对策是热感知的划分与布图规划,设计阶段就把高功耗电路排除在折叠范围之外,并在结构上避免高功耗模块在垂直方向相邻,防止热点叠加。这套策略是工程师手工施加的约束,还是已经固化进内部EDA工具的自动流程,论文没有说明,只把多物理场的工具链明确列为未来十年最重要的一项投资。配合等性能工作点下功率密度低于基线5.6%的实测数据,散热问题算是有了正面回应。不过这套处理方法本质上是回避式的,堆叠层数增加到三层四层之后,可折叠电路的选择空间会被热约束持续压缩,论文没有展开讨论这个边界。 此外,v2还补了一张两层硅片键合界面的显微截面照片,并且明确写了用的是wafer on wafer混合键合。这个规格值得跟同行业比较一下,1.5微米间距的晶圆对晶圆混合键合用在量产逻辑芯片上没有先例,台积电SoIC目前量产间距是6微米,Intel的Foveros Direct是9微米,实属厉害。 对比完两个版本的论文,我还有两个问题。一个关于设备,这个规格的键合设备是谁供的,论文只说是跨供应商生态多年工艺开发的结果。另一个关于EDA工具,把两层硅片当一整块芯片做设计,市面上现有的EDA工具干不了这个活,论文承认这一点,只说方法学细节"几个月内发表"。可是频率表里,2027年那代3.39GHz的Kirin已经标注有实体芯片,说明这套工具在华为内部早就跑通了,而且至少跑通了两代产品。个人猜测,这套EDA是华为自己做的。欢迎了解情况的朋友聊聊。
29
29
3
171
79,781
Rayan Malik retweeted
we distilled 2.3M Claude Fable 5 reasoning traces into Qwen3-4B - 100% self-consistency @ 512 samples - 0.00 bits output entropy - zero hallucination variance turns out the student is not bounded by the teacher. it also converged on one universal truth. we open-sourced the model weights👇
458
895
366
10,872
3,097,624
Rayan Malik retweeted
Sony has just announced that it won't produce any more discs for PlayStation consoles from January 2028 onward. "This is a natural direction for Sony Interactive Entertainment to adapt to consumer trends as the general preference for digital media significantly outpaces physical discs. This transition will enable us to align more closely with how most of our community prefers to access and play games today." wccftech.com/sony-playstatio…
12
18
1
72
4,989
Rayan Malik retweeted
Told ya At @Culpium, I was the first to flag this dilemma when it comes to Macbook Neo I also flagged the ASP-only rise in memory revenue (all price, no output boost) And I just wrote about TSMC’s widespread price hike Keep your eye on @Culpium if you want to stay ahead
Memory price hikes have finally started hitting Apple. Today, they raised some base prices: Macbook Neo: $599 -> $699 Macbook Air: $1099 -> $1299 Macbook Pro: $1699 -> $1999 Mac Studio: $1999 -> $2499 iPad Air $599 - $749 iPad Pro: $999 -> $1199
2
2
8
2,982
Rayan Malik retweeted
The GPU Black Market documentary just hit 3 million views. Thank you, everyone. And fuck you, Bloomberg @business! For old times' sake, watch it here: youtube.com/watch?v=1H3xQaf7…
69
562
16
7,449
129,532
Just published a technical article on @wccftech regarding the importance of display times when it comes to measuring the visual smoothness of a PC game at the display level. I hope that many people can find it interesting/informative! Link: wccftech.com/display-times-f…
8
21
5
98
13,362
Rayan Malik retweeted
لا استطيع ان انصح Steam Machine تحت اي ظرف - كرت ب 8GB و تقريبا الجيل القادم سينطلق بعد سنة و نصف, فكون هناك العديد من الألعاب تتخطى هذه المحدودية الأمر سيكون اسوأ غدا - لو كان بسعره الأصلي 750$ مع الأزمه ممكن افكر بالأمر لكن 1000$ يوجد خيارات افضل - ميزته فقط دعمه ل FSR4
3
2
2
58
8,429
Steam Machine? More like Meme Machine. For about the same price you could build a MUCH more powerful gaming PC, at least with US prices. As such, I strongly recommend not buying it.
110
49
32
895
345,475
My latest @wccftech article is up! It's about properly identifying, diagnosing, and fixing GPU/CPU/system-wide gaming performance bottlenecks. I hope that many people can learn a lot from it! Link: wccftech.com/how-to/cpu-or-g…
6
24
3
98
9,037