The next major refresh of AMD's Ryzen AI MAX APUs is still far away, but now we are putting together the pieces of the "Medusa Halo" APU puzzle. According to a famous leaker, @Olrak29_ on X, AMD's next-generation "Medusa Halo" APU will be complemented by LPDDR6 memory. This is one of the first LPDDR6 memory SoCs we are learning about, making it unique. Based on previous rumors, the silicon could have a 384-bit bus powering LPDDR6 memory, which would translate into massive bandwidth powering the SoC's new CPU and GPU configuration. This includes up to 24 "Zen 6" CPU cores and 48 RDNA 5/UDNA compute units for the GPU configuration. Paired with the added bandwidth from LPDDR6 memory—which these APUs greatly benefit from—"Medusa Halo" will be one of the best-performing SoCs when it launches.

Interestingly, memory manufacturers like Samsung and Innosilicon are already supplying LPDDR6 modules to customers for validation. Innosilicon's LPDDR6 modules boast an impressive speed of 14.4 Gbps, significantly faster than Samsung's initial modules, which achieve 10.7 Gbps. Innosilicon's modules offer a 1.5x increase in IO speed capability compared to the 9.6 Gbps of LPDDR5X previously available, along with improved efficiency. The latest LPDDR6 also increases the number of bits per byte of IO from 8 to 12. This results in LPDDR6's bandwidth at a single-channel 24-bit I/O speed being double that of LPDDR5X at a 16-bit single-channel. The company is reportedly collaborating with TSMC and Samsung to ensure sufficient production capacity for LPDDR6 IP, while Samsung relies on its own fabs for manufacturing memory.

Innosilicon has reportedly started shipping its LPDDR6 memory modules, which offer an impressive speed of 14.4 Gbps. This is significantly faster than Samsung's initial LPDDR6 modules, which achieve data rates of 10.7 Gbps. Innosilicon's modules provide a 1.5x increase in IO speed capability compared to the 9.6 Gbps of LPDDR5X that the company previously offered, and they significantly enhance efficiency. The latest LPDDR6 also increases the number of bits per byte of IO from 8 bits to 12 bits. This results in LPDDR6's bandwidth at a single-channel 24-bit I/O speed being twice that of LPDDR5X at a 16-bit single-channel. The company has reportedly prepared collaborative manufacturing with TSMC and Samsung to secure sufficient capacity for LPDDR6 IP production.

As one of the leading LPDDR6 implementations, this move directly challenges competitors like Samsung, SK Hynix, and Micron, who are also presenting their LPDDR6 solutions to customers but are not eager to ship much of the capacity early in the LPDDR6 lifecycle. Additionally, the JEDEC specification lists LPDDR6 running at a maximum speed of 14.4 Gbps per pin, which Innosilicon has aimed for from the start. This contrasts with Samsung's approach, which has developed 10.7 Gbps LPDDR6 memory on a 12 nm node, peaking at the same speed as its earlier LPDDR5X. However, Samsung emphasizes power efficiency as the main selling point of these early modules, claiming a 21% improvement in energy efficiency over its predecessor. It remains to be seen which IP customers will choose, and the availability of off-the-shelf modules will certainly influence the final decision.

InnoSilicon has unveiled the Fenghua No.3 GPU marking a significant shift in how the Chinese company develops GPUs. While for previous versions, they used Imagination Technologies' PowerVR IP, now the Fenghua No.3 uses an open-source RISC-V design based on OpenCore Institute's Nanhu V3 project. The Fenghua No.3 comes with over 112 GB of HBM memory putting it right in the AI accelerator market. InnoSilicon says one card can run 32B and 72B parameter language models. They also claim that if you use eight cards together, you can handle huge 671B and 685B parameter models.

The GPU support DirectX 12 Vulkan 1.2, and OpenGL 4.6 APIs for games, and it can handle even ray tracing. InnoSilicon showed the card running Tomb Raider, Delta Force, and Valorant, however it didn't share exact performance numbers or the settings in use. The card can power up to six 8 K screens at 30 Hz and supports YUV444 format to improve color accuracy. This aims at medical, CAD, and scientific computing markets. What's even more interesting is that InnoSilicon says that Fenghua No.3 it's CUDA compatible—a bold claim that, if true, could expand the GPU's software options. However, since CUDA is proprietary, this claim requires verification.

According to DigiTimes, several Chinese GPU makers are aiming to build GPUs using 5 or 7 nm nodes this year, something that might be a challenge for them, considering that TSMC's key customers are already said to have pre-paid TSMC to get preferential access to these nodes. The companies in question are Innosilicon, who works with Imagination Technologies IP, Changsha Jingjia Microelectronics and Biren Technology, all of which are largely unknown players in the GPU market space.

Innosilicon's Fantasy One is said to offer performance similar to GeForce RTX 2070, which means it might even be a competitor for Intel's Arc GPUs, assuming the Fantasy One doesn't end up being a pipe dream. Changsha Jingjia Microelectronics is said to have announced it first GPU back in December last year with the JM9 series, which is said to offer around 80 percent of the performance of a GeForce GTX 950, which puts it in the office PC category these days and almost make you wonder why they bothered. Finally Biren Technology announced the BR100 in October last year and it's apparently already manufactured on TSMC's 7 nm node, although no word on performance is available. The bigger question is if any of these products will have any impact in the GPU market, since at best, they might offload some customers in the PRC from buying GPUs from AMD, Intel and Nvidia, until these companies have proven that they can deliver viable drivers alongside their hardware.