Did some pixel counting: Based on the known dimension for the LPDDR5X packages, the Snapdragon X2 Elite Extreme is about 13.9 × 20.7 mm.

Qualcomm uses the N3 node by TSMC for the Snapdragon X2 Elite Extreme. But it's utilizing TSMCs FinFlex, so primarily N3P and for some parts N3X – for the Prime Cores for example. So Qualcomm gets the trade off between efficiency and performance for their monolithic design.

Wafer based on #imec 300 mm spin qubit platform used by #Diraq.

Together with Corintis, Microsoft is working on a liquid cooling based on microfluidics methods. Microfluidics could get datacenter chips to run at a higher efficiency and makes per package TDPs of 2+ kW even possible. www.hardwareluxx.de/index.php/ne...

MediaTek announced the tape out of its first SoC in N2 by TSMC. The chip is set to be available at the end of 2026. corp.mediatek.tw/news-events/...

According to DigiTimes, due to the high costs incurred by external suppliers and in this context the challenges with EUV, TSMC has decided to manufacture the so called pellicles, an extremely thin, transparent protective film for photomasks by itself. www.digitimes.com.tw/tech/dt/n/sh...

Some lovely details on that fibre attachment!

This is Omega, a quantum chip by PsiQuantum with single photon sources, superconducting single photon detectors, and a next-generation optical switch. I don't know much about quantum computing, but the wafer looks super nice.

That might be the #Rubin based datacenter product that is the closest to a consumer product. A GDDR7 memory interface, video de- and encoders. But that's to much thinking into it ... #RubinCPX

There is now a GPU with 128 GB GDDR7 officially announced by NVIDIA.

By the end of 2026 NVIDIA will release the Rubin CPX – Purpose-built for context performance and efficiency. - 30 PF NVFP4 performance - 128 GB GDDR7 - four NVENC and four NVDEC for video processing www.hardwareluxx.de/index.php/ne...

Recogni becomes #Tensordyne and give some more details on how it will try to make inference way more efficient. Logarithmic Math is not a now approach, bit #Tensordyne might have solved the problem how to get to linear data again. www.hardwareluxx.de/index.php/ne...

"Qualcomm CEO Says Intel's Chip Production Not Good Enough Yet" via @bloomberglp.bsky.social www.bloomberg.com/news/article...

For those looking for a entry-level workstation graphics card the Intel Arc Pro B50 is an interesting and attractively priced option. www.hardwareluxx.de/index.php/ar...

I did a review of the G.Skill Trident Z Neo RGB 2x 64 GB (DDR5-6000, CL34-44-44-96). It's the first time for me to get in touch with the new Samsung 32 Gbit M-Die. With 64 GB it's 8 chips on both sides of the PCB. www.hardwareluxx.de/index.php/ar...

I have already seen a Threadripper fall out of the orange frame and then fall into the socket. Always a tense moment ...

I finally had the chance to reviews the "HEDT" SKUs of the AMD Ryzen Threadripper 9000 series. There is simply no alternative to the processors and I can't see how Intel is going to come up to par here in the near future. www.hardwareluxx.de/index.php/ar...

The main differentiator for Rapidus will be faster turnaround times be using a single wafer processing instead of batches. Standard TAT will be around 50 days, while a (super) hot lot can be run in 15 days.

More details on the GB10 Grace Blackwell Superchip by NVIDIA as shown at Hot Chips - S-Dielet/S-Die (CPU, Memory Subsystem, etc.) - G-Dielet/G-Die (Blackwell-GPU) - TSMC 3nm (both) - Advanced 2.5D Packaging - 256b LPDDR5X-9400 (301 GB/s) www.hardwareluxx.de/index.php/ne...

I'm currently digging in some details about Athos Silicon, DreamBig Semi and their mSoC designs Polaris and Northstar. While they chose Samsung as a foundry partner, Silicon Box is doing the packaging.

Detailed analysis by Kurnal of the Huawei #KunPeng930 - two compute dies (252,3 mm², TSMC N5?) - two I/O dies (172,3 mm², SMIC 14?) www.hardwareluxx.de/index.php/ne... Video: youtu.be/ALXcH1uXHgY?...

Intel gives some more details on #ClearwaterForest at @hotchipsorg #hc2025. - wider instruction decoders - better branch prediction - massive parallel OoO engine - deeper execution engine - optimized memory subsystem hardwareluxx.de/index.php/ne...

NVIDIA did release some further details on the Blackwell Ultra GPU. - higher FP4 Tensor Core (Dense) performance - much lower INT8 Tensor Core (Dense/Sparse) performance - even lower FP64 (Tensor Core) performance - new 1,400 TDP - PCIe 6.0 to host CPU www.hardwareluxx.de/index.php/ne...

Hot Chips 2025 started yesterday with a datacenter racks tutorial. AMD, NVIDIA, Meta, Microsoft and Google did talk about challenges regarding power, cooling, interconnect complexity and more. www.hardwareluxx.de/index.php/ne...

SK Hynix 321-Layer QLC NAND will start mass production now with first products available in H1 2026. Thats 2 Tbit (256 GB) per chip with the option to do a 32DP3 stacking for 32 chips in one package, that allows 8 TB per package. news.skhynix.com/sk-hynix-beg...

For those who are not aware of these details with NVLink Fusion: For custom XPU configurations, the interface to NVLink utilizes integration of Universal Chiplet Interconnect Express (UCIe) IP and interface. NVIDIA provides the bridge / NVLink chiplet for UCIe to NVLink.

With Spectrum-XGS @NVIDIAAI wants to connect several AI data centers to run as a giga-scale AI super factory. Scale-up, scale-out and now scale-across. Auto-adjusted distance congestion control, precision latency management and end-to-end telemetry.

Kioxia High-Bandwidth Flash Memory Module - 5 TB - 64 GB/s - 40 W - daisy chained flash/controller - serial 128 Gbps PAM4 interface - prototype memory module with PCIe 6.0 (64 Gbps, 8 lanes) www.kioxia.com/en-jp/about/...

I like that carbon finish of the INNO3D GeForce RTX 50 Frostbite Pro!

That SMD arrangement ...

"AI startup Perplexity made a formal offer to acquire Google’s Chrome browser for $34.5 billion, an audacious bid to get ahead of a potential requirement for the search giant to sell the web browser in US antitrust proceedings." www.bloomberg.com/news/article...

Testing sTR5 platform / Ryzen Threadripper is no fun. Memory training takes ages - easily 4-5 minutes for the first boot.

Letter from Lip-Bu Tan to Intel employees:

Microsoft incorporates OpenAI’s GPT-5 into consumer, developer and enterprise offerings. GPT-5 in Copilot for free. GitHub is bringing GPT-5 to all paid GitHub Copilot plans. news.microsoft.com/source/featu...

PCI-SIG is doubling up raw data rate again with PCI-Express 8.0. www.hardwareluxx.de/index.php/ne...

AMD Ryzen Threadripper PRO (9000WX series) vs Ryzen Threadripper (9000X series)

Still on vacation. But those two monsters are waiting at home and will be the first things that will be fired up. Can’t wait to see if there still is a niche for HEDT. #RyzenThreadripper

The discussion about the use of #Palantir software in Germany is frightening. Talking about 100% sovereign data usage here, both on the software and hardware side, simply cannot be upheld. It's frightening how naively politicians talk about things they have no idea about.

IBM did add a note (note a) in the newly released P11 Redbook. It indicates a change in how the transistors are counted. So there is no significant increase in the transistor count for the Power11 relative to the Power10. Thanks to Tom!

Huawei Ascend 910C die shots and annotations by Kurnal - SMIC N+2 7nm - two die connected through substrate/RDL - 4x HBM / 4x dummy per GPU youtu.be/0feji7gMrh8?...

While I'm on vacation, Intel shakes the foundations of the company and turns everything upside down. - 15% headcount reduction to 75,000 - fabs in Germany and Poland cancelled - Intel 14A will decide the future of the Foundry division - further slowing construction in Ohio

Oracle and OpenAI have entered an agreement to develop 4.5 GE of additional Stargate data center capacity. The dimensions are just crazy. openai.com/index/starga...

As a German it's kind of funny that Rapidus calls it's "Rapid and Unified Manufacturing Service" in short RUMs.

Wafer with 2 nm GAA Transistors as shown by #Rapidus. The PDK will be released to advance customers by Q1 2026. Rapidus plans the begin of mass-production in 2027.

Nikon presents the Lithography System DSP-100 that is able to produce square substrate with up to 600 x 600 mm. It can do 50 panels/hour with "standard" 510 x 515 mm substrate. - Resolution: 1.0 μm L/S - Overlay Accuracy: ≦±0.3 μm - Release: 2026 www.hardwareluxx.de/index.php/ne...

But I have a question: Where is the Ryzen Threadripper Pro 9945WX with 12 cores?

I did run some benchmarks on @cyberpunk.net with a M3 Pro. 1.800 x 1.125 Pixel (30 FPS Target): 30 FPS 1.800 x 1.125 Pixel (60 FPS Target): 49.5 FPS (38.2 Min/96.4 Max) 1.800 x 1.125 Pixel + Raytracing (60 FPS Target): 25.5 FPS (21.2 Min/30.6 Max) www.hardwareluxx.de/index.php/ne...

AMD did reveal the prices for the Ryzen Threadripper PRO 9000 WX-Series processors. www.amd.com/en/blogs/202...

TSMC reports its 2nd quarter numbers and here is my quarterly graphical representation of revenue by technology.

Intel significantly reduces the price of the Core Ultra 5 225 und 225F in Germany. The Core Ultra 5 225 is down from about 220 to 250 Euros to now 185 Euros. The Core Ultra 5 225F costs about 175 Euros – compared to 225 Euros in the last week. www.hardwareluxx.de/index.php/ne...