irisparadox (@irisparadox) header
irisparadox (@irisparadox) profile picture

Software & Hardware Engineer. 3D Modeler.

Spain
Joined September 2025
Photo by irisparadox (@irisparadox)Photo by irisparadox (@irisparadox)Photo by irisparadox (@irisparadox)Photo by irisparadox (@irisparadox)Photo by irisparadox (@irisparadox)Photo by irisparadox (@irisparadox)Photo by irisparadox (@irisparadox)Photo by irisparadox (@irisparadox)Photo by irisparadox (@irisparadox)Photo by irisparadox (@irisparadox)
The 8192 core CPU (@uscompco) is coming along well. I've been doing some major changes the past few weeks. This is a core layout from yesterday. The core got a lot bigger with the redesigns (mostly some naive approaches to patching gaps and bugs). Some new OpenROAD jobs running tonight with a projected 20% logic reduction. This area plot also has some obvious low density spots and is very disorganized, that should get better too. Routing should be better than the old core though, much less wiring. Raw performance should be a lot higher with the redesign though, as well as fixing a lot of interconnect deadlock bugs. At peak the current design can issue 15 instructions per cycle per core, not counting SIMD! (Yes, we do have cases where the (magic-mininal) compiler actually makes use of all this! Most of the time it's around 3-6, but that would be impressive sustained for an out of order CPU and this is in-order!) Inner loops and math workloads tend to be on the higher end. This is also done in a way (using some inspiration from 1940s/50s wisdom) where we can pay only a very tiny fraction of the regfile/crossbar logic cost you'd usually need to pull that off. Nonzero chance markets make upcoming fundraising a bit tougher, but we're still aiming for tapeout next year!
8
5
77
2,034
512 address RAM blocked into cache lines of 32. Here we see a flurry of activity kicked off by a cache miss.
13
41
4
492
11,955
Neat, you can just have the LLM build a regfile compiler to get much higher logic density.
16
23
2
397
24,214
irisparadox retweeted
ICsprout 55nm Open Source PDK now has initial LibreLane support. SPM went from RTL to GDS in one command: 617 standard cells, die area 2996.9 μm², utilization 100%. KLayout DRC passed. Magic DRC, Netgen LVS, and Magic PEX still fail, with hold violations at the TT corner.
3
14
102
3,044
irisparadox retweeted
the Linux kernel wraps almost every multi-line macro in do { } while (0) it looks like a loop it never actually loops the pattern makes multiple statements behave like a single statement so macros work correctly inside if blocks and other control flow that's why you'll find it throughout the kernel source
20
26
6
375
30,778
I've been working on my KiCad autorouter, and these are single layers of copper.
24
16
3
390
33,061
A new video on computer architecture from @CoreDumpped youtu.be/eNtADpJ2EC4
System Design and First Principles are not from a parallel universe; they should not live in isolation. - Computer architecture provides the "missing taste" - Critical "hot paths" get identified early in the game - Performance engineering is never an afterthought - The show-and-tell of performance is always measure-first
24
245
7,620
intel 4004 is beautiful
5
49
1
531
21,385
When I post plots of the CPU I'm working on, people are surprised at how organic it looks. If you look through @arenamag's Silicon book, especially at chips from the past 10-15 years, most also look very organic. The organic-looking parts are logic. The grids are memory.
7
33
5
353
21,013
As an Embedded System Engineer, Please learn: • C & Modern C++ or Rust • Data Structures & Algorithms • Computer Architecture • Digital Logic & Boolean Algebra • Microcontrollers (STM32, ESP32, AVR, RP2040) • ARM Cortex-M/A & RISC-V • Embedded Linux • Linux Kernel & Device Drivers • RTOS (FreeRTOS, Zephyr) • Bootloaders (U-Boot) • Build Systems (CMake, Yocto, Buildroot) • Memory-Mapped I/O • Interrupts, DMA & Timers • UART, SPI, I²C, CAN, USB & Ethernet • TCP/IP, BLE, Wi-Fi & MQTT • Sensors & Actuators • ADC, DAC & PWM • Power Management & Low-Power Design • Firmware Architecture • Linker Scripts & Memory Layout • Debugging (GDB, JTAG, SWD) • Oscilloscopes, Logic Analyzers & Multimeters • PCB Basics & Hardware Bring-up • FPGA Fundamentals & HLS • Embedded AI & TinyML • CUDA & Edge AI (for embedded AI systems) • Python & Bash for Automation • Git, CI/CD & Testing • Secure Boot & Firmware Security • Functional Safety Basics (IEC 61508, ISO 26262) • Robotics & ROS 2 (optional but valuable)
24
117
6
926
37,025
irisparadox retweeted
I am pretty impressed what open source EDA tool can achieve already. This is an programmable instrumentation amplifier. And yes. Its design and the IHP PDK used is open source :-)
7
39
311
13,148
Learning computer architecture by building a full GPU from scratch. ▪️ 16-bit ISA & assembler ▪️ High-level programming language (J++) & compiler ▪️ Branch divergence & warp scheduling ▪️ 2 compute cores Running AI model inference at 81 MHz on a 30$ FPGA 🧵
16
87
8
877
66,738
irisparadox retweeted
week 127, streaming kernel dev (c++) topic: boot bump allocator instead of using global objects for the kernel (and its large, static buffers), I finally implemented a basic bump allocator for boot time and used that instead link + recap ↓
1
3
116
5,312
This diagram is art, a symphony. It's from WikiChip, a site now probably gone into oblivion; Chips and Cheese carries the mantle now.
7
2
126
17,853
irisparadox retweeted
If you want to understand Snapdragon X Elite CPU architecture, read chips and cheese substack articles, they're really good. You will understand ROBs, Cache hierarchy, OOO, n-way decoding, branch predictions etc
1
11
112
3,505
irisparadox retweeted
One research paper completely changed how architects think about DRAM, even though it's not implemented in today's commodity DDR memory. For decades, copying memory followed the same path: DRAM → Memory Controller → CPU Cache → CPU → Cache → Memory Controller → DRAM Even when the CPU isn't doing any computation, it still burns cycles just moving data. Every "memcpy()", "fork()", Copy-on-Write, page migration, or zero-page initialization consumes memory bandwidth, pollutes caches, and wastes energy. The RowClone paper asked a simple question: «Why move the data to the CPU if all you want to do is copy it?» The key insight was to exploit existing DRAM circuitry. When a DRAM row is activated, its contents are latched into the row buffer (sense amplifiers). By issuing carefully timed back-to-back ACTIVATE commands within the same subarray, an entire DRAM row can be copied directly to another row, without the data ever leaving the DRAM chip. No CPU. No cache traffic. No external memory bus. For copies across different banks or subarrays, RowClone proposes another mechanism that still avoids involving the CPU in the data movement. The projected benefits were impressive: • Up to 11× lower copy latency • Up to 74× lower energy consumption • Reduced memory bandwidth usage • Faster OS operations like "fork()", Copy-on-Write, VM cloning, page migration, and bulk memory initialization. Here's the interesting part: RowClone isn't implemented in today's mainstream DDR4/DDR5 systems. It requires changes to the DRAM chips, memory controller, and software stack, making commercial adoption difficult. But its impact was far bigger than one feature. RowClone fundamentally changed how researchers viewed memory, not as passive storage, but as hardware capable of performing useful operations itself. That single idea inspired a decade of Processing-in-Memory (PIM) research, including works like Ambit, Samsung's HBM-PIM, and many modern near-memory computing architectures. Sometimes, the biggest breakthrough isn't making the CPU faster. It's realizing the CPU didn't need to be involved in the first place.
4
18
1
127
5,294
irisparadox retweeted
CVE-2018-8611 is one of my favorite Windows kernel race-condition exploits because it demonstrates that winning the race is only half the battle. The vulnerability lives in the Windows Kernel Transaction Manager (KTM). During transaction recovery, a race condition creates a use-after-free on a "_KENLISTMENT" object. But a dangling pointer alone isn't enough, you still need to turn it into something useful. NCC Group's exploit is a masterclass in exploit development: • Win the race and reclaim the freed "_KENLISTMENT" with controlled data. • Build an increment primitive by abusing "KeReleaseMutex()" on a fake object. • Convert that into an arbitrary kernel read primitive. • Finally achieve arbitrary kernel read/write, steal the SYSTEM token, and elevate privileges. What's fascinating is that the exploit doesn't magically jump from UAF → SYSTEM. It carefully chains together increasingly powerful primitives until full kernel control becomes possible. This series is also an excellent lesson in modern Windows internals, reference counting, object lifetime, KTM recovery, pool manipulation, synchronization bugs, and why race-condition exploitation is often far more difficult than traditional memory corruption. If you want to learn Windows kernel exploitation beyond toy examples, this five-part KTM series is essential reading.
1
13
62
2,881
irisparadox retweeted
180/365 of GPU Programming One of my favorite resources recently has been Fundamentals of GPU Architecture which is a 9-part video series which goes through the book General-Purpose Graphics Processor Architectures. The lectures are from 2019 but still useful today for getting a more foundational understanding of how the hardware architecture connects to GPU programming primitives.
179/365 of GPU Programming. Studying the top 3 QR kernels in a bit more depth today. For context, all three landed in the 1200μs range whereas the next closest was above 1400μs and my submission ended up in the 1700μs. Interesting to compare my approach to the top ones and to see where the speed difference ended up coming from (getting lower precision formats to work reliably, EMULATED_16BFX9, no reconstruction/gates, et cetera). So much I still have (or rather, get) to learn, exciting times!
2
34
1
407
21,037
irisparadox retweeted
Open source EDA + open source IP + open source PDK! The world's highest-level open-source chip (55nm) has been successfully lit up! Open for trial: buff.ly/BynJT1r
9
49
1
310
11,841
irisparadox retweeted
In 2018, researchers challenged one of the biggest assumptions in storage security: Hardware full-disk encryption isn't automatically more secure than software encryption. Their paper, Self-Encrypting Deception, reverse engineered the firmware of several self-encrypting SSDs (SEDs) from major vendors to see whether the security claims matched reality. The result was surprising. The cryptographic algorithms themselves weren't the primary problem. The weaknesses were in the design and implementation. Some drives generated predictable encryption keys. Others failed to properly authenticate users before unlocking encrypted data. In several cases, the researchers were able to recover all user data without knowing the disk password by exploiting flaws in the drive's firmware and key management. The encryption wasn't "broken", the implementation around it was. The impact extended beyond the SSDs themselves. At the time, Microsoft BitLocker would automatically use hardware encryption if a drive advertised support for it. If that hardware implementation was vulnerable, BitLocker inherited the SSD's weaknesses instead of using its own software-based encryption. This wasn't an attack against AES. It wasn't a break of the TCG Opal specification either. It was a reminder that secure cryptography can still be undermined by insecure firmware, poor key management, weak authentication, or implementation mistakes. The paper fundamentally changed how many people viewed self-encrypting drives. After its publication, Microsoft updated BitLocker guidance so that new Windows installations default to software encryption unless administrators explicitly choose to trust hardware encryption. SSD vendors also released firmware updates for affected products. The biggest lesson wasn't "don't trust hardware encryption." It was this: Security is a property of the entire system not just the cryptographic algorithm. A mathematically strong cipher can't compensate for weak firmware, flawed key management, or broken authentication. That's why cryptography is often the easiest part of building a secure system.
Most people think "hardware-encrypted SSD" and "TCG Opal" mean the same thing. They don't. A Self-Encrypting Drive (SED) encrypts data in hardware. TCG Opal is an open industry standard that defines how software securely provisions, authenticates, locks, unlocks, and manages that encrypted storage. The encryption itself happens inside the SSD controller. As data is written to NAND flash, the controller encrypts it transparently using media encryption keys that remain within the drive. Reads are decrypted just as transparently after the drive has been successfully unlocked. The host doesn't encrypt every sector when Opal hardware encryption is being used. Instead, the operating system or management software communicates with the drive using the Opal protocol to configure security policies, authenticate users, and control access to protected data. One of Opal's most useful features is Locking Ranges. Instead of protecting the entire drive as a single encrypted volume, Opal allows independent ranges of logical block addresses (LBAs) to be locked or unlocked separately. This enables different regions of the same drive to have different access policies. Authentication determines whether the drive will make protected data accessible. The Opal specification defines the authentication and authorization framework but intentionally does not mandate exactly how vendors implement their internal cryptographic key hierarchy. Those implementation details are left to SSD manufacturers. Another powerful capability is cryptographic erase. Because data is already encrypted, securely sanitizing a drive doesn't necessarily require overwriting every NAND cell. By replacing or destroying the encryption keys protecting the stored data, previously written ciphertext becomes computationally infeasible to recover. Opal also supports Pre-Boot Authentication (PBA). When deployed, the drive remains locked before the operating system starts, helping protect against offline attacks where an attacker removes the SSD and attempts to access it from another machine. Since encryption and decryption are performed by dedicated hardware in the SSD controller, the CPU doesn't perform the bulk cryptographic work during normal I/O. The host primarily handles authentication, policy management, and storage commands. One important caveat: The security of an Opal-based system depends on both the specification and the vendor's implementation. In 2018, researchers showed that several commercially available self-encrypting SSDs contained implementation flaws that could undermine their security even though they supported hardware encryption. A good specification is only half the story. A secure implementation is what ultimately protects your data.
2
9
2
95
4,715