DEV Community: Debarshi (Deb).S

PS3 RSX

Debarshi (Deb).S — Wed, 12 Aug 2026 04:00:15 +0000

Officially my PS3 RSX research journey started 🙂

Started as a normal Cell programming lesson on PS3 turned into a 3D acceleration development for PS3 linux RSX

After 20 years I'm finally working on bringing proper RSX graphics acceleration into PS3 Linux.

DDR3 Memory Engine Architecture

Debarshi (Deb).S — Tue, 11 Aug 2026 15:33:13 +0000

Just discovered that 1GB DDR3 auxiliary pool on PS4 pro can be use on any workload.

Introducing: DDR3 Memory Engine Architecture

The Challenge:
The PS4 Pro isolates its primary 8GB high-bandwidth GDDR5 pool for graphics and system allocation. However, the secondary 1GB DDR3 BGA chip on a completely separate, undocumented bus handled via a custom PCIe Endpoint window on the Belize Southbridge co-processor. While the outbound posted write path bursts past 1 GB/s (and settles at a stable 156 MB/s), the non-posted read path hits an immutable, hardwired serialization bottleneck clamped directly at ~11 MB/s with an average 85ms round-trip latency penalty.

So, instead of accepting the read bottleneck, added custom 8-queue in my DDR3 driver directly into the Linux 6.8.12 kernel, explicitly mapping it to tap into all 8 AMD Jaguar CPU cores simultaneously.

To bypass the silicon-level read throttle and I designed a hybrid memory-tiering architecture inspired by the iconic PS3 Cell Broadband Engine's SPE Local Store layout:

LZ4 Compression Engine (zswap): Implemented an early-boot kernel caching layer to shrink memory blocks at a 4:1 ratio, effectively quadrupling the operational bandwidth over the restricted physical bus width.

Double-Buffered "DMA" Streaming Pipeline: Built an asynchronous thread allocation strategy where background threads pre-fetch contiguous 4MB memory chunks sequentially via the block layer's read-ahead tracking while the core CUs (compute units) crunch live tensor layers directly out of the shared L2 cache—completely hiding the hardware latency loop.

The Real Impact-
By treating this 1GB DDR3 pool as a cold fallback storage vault for background assets & static tensor layers, we can isolate and hand nearly 100% of the ultra-fast 8GB GDDR5 RAM pool directly to active compute tasks, transforming normal Gaming consumer hardware into optimized nodes for Edge AI model inference & distributed training.

Still need to add support of GFX8.0 on PS4 pro that's another part

PS4 Pro Belize DDR3

Debarshi (Deb).S — Tue, 04 Aug 2026 08:27:30 +0000

Finally I did it,

So, I recently workin on reverse-engineering the Sony PlayStation 4 Pro memory subsystem under PS4Linux kernel, specifically targeting the auxiliary 1GB DDR3 DRAM managed by the Belize Southbridge co-processor.

On Orbis OS, this isolated 1GB pool is exclusively for OS asset swapping and background app parking. When running Linux, this memory remains completely unutilized because it is entirely separated from the APU's unified high-speed GDDR5 main memory controller.

So, I build a custom Linux Kernel module to hook onto this unmapped hardware and mount it natively as a zero-latency system swap device.

Engineering Roadblocks

Strict Bus Protocol: My early testing using direct memory mapping triggers immediate Input/output error exceptions. The Belize memory bus controller aggressively drops or faults any read/write transaction that isn't cleanly aligned to a strict 32-bit boundary so standard loops immediately stall the system.

ACPI/Device Tree Isolation: The PS4 have no standard PC like BIOS/ACPI tables, the initialization payload maps out the early E820 hardware boundaries. The memory space spanning 0x80000000 - 0xDFFFFFFF was explicitly flagged as SMAP_TYPE_RESERVED.

Kernel Validation blocks: On Linux kernels like v6.x, the virtual block layer (blk-mq) natively rejects non-PCI block driver allocations via add_disk() if it cannot identify an explicit struct device parent tracking node hierarchy inside sysfs.

So, the best part my 💡Multi-Layered Engineering Solution

Added Payload Carving modified the primary jailbreak kexec() loader payload source (linux_boot.c) to cleanly cut the rigid "Reserved" hardware block and declare the first 1GB slice (0x80000000 with a size of 0x40000000) as safe system MEMORY.

The blk_mq Parent Injection Bypass abandoned standard virtual device structures. Leveraged deep kernel internal abstractions (blk_mq_alloc_disk) to manually inject a custom pseudo-root device tracking parent inline (disk->part0.bd_device.parent).

Strict Word-Pacing IO Loops implemented a dedicated transaction wrapper utilizing ioread32 and iowrite32 streaming chunks to satisfy the co-processor's hardware bus constraints perfectly.

So the Output-
The driver compiles perfectly on my custom 6.8.12 debug kernel source. On boot, the kernel cleanly registers the device node at /dev/ps4pro_ddr3 as a true block storage format structure HELL YEAH

Btw I'm also a Kernel Contributor for PS4linux developer community so yeah u know..

My next Architecture in simple word-
run linux on 1gb DDR3 and use hole GDDR5 pool as a game memory on PS4pro like a true PS4 but on Linux I will see about that later....

DebLink utility console

Debarshi (Deb).S — Sat, 18 Jul 2026 21:27:11 +0000

Finally ported my Payload Laucher app into android now it's called Deblink.

If u don't know it's a simple application to sent payload to your Homebrew devices and it's now finally ported to native android application.

in future I will add SSH controller FTP server and File sharing Integrate with Croc file sharing

Plugin MC

Debarshi (Deb).S — Thu, 16 Jul 2026 03:49:37 +0000

🚀 PluginMC is officially live.

🌐 https://pluginagentmc.com

Every Minecraft server owner has ideas that could make their server unique.
The problem is that many don't know where to start, struggle to find reliable developers, or end up wasting time and money on projects that never get delivered as promised.

That's why we created PluginMC.
We're a Minecraft plugin development agency that connects server owners with experienced developers to build custom systems tailored to their vision.

Whether it's a unique progression system, custom bosses, economy mechanics, enchantments, or completely original gameplay ideas, our goal is simple:

Help server owners turn creative ideas into production-ready Minecraft systems.
Our showcase projects
⚔️ Multi-Phase Boss System
🪙 Mob Token Economy
✨ Custom Enchantment System
🌍 Community Progression Framework (In Development)

What we build

Custom Paper, Spigot, Folia & Bukkit plugins
Progression & RPG systems
Economy mechanics
Boss systems & custom events
Performance-focused server solutions
Long-term development & support
Every great Minecraft server starts with a unique idea.
We're here to help turn that idea into reality.

🌐 https://pluginagentmc.com

We connect server owners with experienced developers and manage the project from idea to delivery.

Minecraft #Java #PaperMC #Spigot #PluginDevelopment #GameDevelopment #JavaDeveloper #Startup #agency #Services

Vulkan Computing with FP16 on PS4 GPU

Debarshi (Deb).S — Sat, 11 Jul 2026 19:06:49 +0000

Finally Vulkan computing possible on PS4 GPU

vulkan compute array running on PS4 GPU(Gradius):
Using physical device index 0 found AMD Radeon Graphics (RADV GLADIUS) Output: { 0 4 12 }

This successful confirms that GPU buffer allocation, compute shader execution, and result readback through the Vulkan compute pipeline.

My Next goals:
• Matrix multiplication benchmarking
• Tiny transformer inference experiments via vulkan
• Vulkan-based neural network research on Gradius

After a lot of kernel, Mesa, and driver work, finally it's possible.
and the best part PS4 Pro shader cores are now can execute FP16 floating-point compute shaders correctly through Vulkan Compute.

My own Heterogeneous AI Super Cluster

Debarshi (Deb).S — Fri, 03 Jul 2026 08:41:44 +0000

Helloo everyone, I'm Deb 14 years old from India West Bengal working on a research project around 1 years workin on that called Heterogeneous AI Super Cluster training big LLM models, it can perform Quantum Simulations around 32-35 Qubits with 10TFLOPS compute Power it power my own Private AI model specially trained for designing rocket engines space rovers designing AI chips CPUs. with my own OpenCL based custom AI distributed computing framework with multipal CPU architecture with custom ai training system under low memory and for Quantum Research & development.

And with my own low-level Programming language PCHL parallel computing hybrid language including Cell it's build for computing on my cluster
it's still in development stage so I will tell about it more stay tune

but one thing I really proud to say I did in last few years that One PS4 pro can perform parallel computations with OpenCL kernels.
It supports floating-point math (including double precision), which is useful in scientific computing and AI.
can run custom compute tasks, neural networks, or image processing on the GPU.
Frameworks that use OpenCL can access this GPU. I did that few years ago.