Most of it comes from MediaTek<\/h3>\n
The MediaTek chiplet (\u201cS-Dielet\u201d or S-Die) contains 20 ARM CPU cores, L3 cache and SLC (system-level cache), a security subsystem, power management block, and all connectivity blocks including PCI Express and USB interfaces. All of this is based on MediaTek\u2019s IP, so they are likely derived from the technologies and blocks the company already used in its other Arm chips.<\/p>\n
The only exceptions to present in the S-dielet that are based on Nvidia-licensed IP are the display output logic, likely including the framebuffer (but excluding the actual blocks and PHY generating DisplayPort and HDMI signals) and the NVLink C2C interface block used to connect with the GPU chiplet.<\/p>\n Interestingly, even the memory controller (for 256-bit LPDDR5X memory) is located on MediaTek\u2019s SoC chiplet and is based on the company’s IP. Until now, it was expected that due to bandwidth requirements, the memory controller would be placed in the GPU chiplet and the CPU would access memory indirectly. This is how AMD handles it in the Strix Halo processors that have comparably powerful GPUs<\/a>.<\/p>\n Nvidia states that it worked closely with MediaTek to ensure that the GPU could achieve sufficient memory bandwidth over the C2C link out of the memory pool managed by the CPU side. Although it is MediaTek\u2019s IP, modifications were made to suit Nvidia\u2019s needs. For example, the CPU and GPU share a coherent memory space, allowing relatively simple data sharing, with the GPU using the same physical address space as the CPU.<\/p>\n Among the advantages of this cooperation, Nvidia highlights that thanks to MediaTek, the CPU\/SoC part including the memory controller is \u201cenergy-efficient.\u201d This could suggests that Nvidia does not feel that confident in the field of mobile and low-power solutions, and a hypothetical solo design developed independently might have struggled to work well in laptops.<\/p>\n The chiplet with the graphics part (\u201cG-Dielet\u201d or G-Die) represents Nvidia\u2019s main contribution. It includes compute units based on the Blackwell architecture. Nvidia is also responsible for the multimedia engine used for video compression and decompression, which is integrated into the GPU chiplet. The NVLink C2C interface for connecting with the MediaTek chiplet is, as expected, also provided by Nvidia (it seems that everything in this piece of silicon is Nvidia\u2019s own design).<\/p>\n It is stated that NVLink C2C provides 300 GB\/s of bandwidth in both directions, enough to carry the full bandwidth made available by the 256-bit RAM used by the SoC. Inevitably, this incurs some additional power consumption and latency compared to the alternative of placing the memory controller in the same silicon, but how significant this overhead is remains unknown.<\/p>\n Both chiplets (dielets) use TSMC\u2019s 3nm process (this is something that has not been confirmed until now), which could theoretically be an advantage over existing competitors (such as AMD\u2019s Ryzen AI Max 300 or Qualcomm\u2019s Snapdragon Elite as an alternative Arm processor for PCs, albeit without a powerful GPU). The chiplets are connected using advanced 2.5D packaging, likely meaning an interposer or silicon bridge are used rather than the less efficient substrate-based connection used in today\u2019s AMD desktop CPUs (something that is supposedly changing only with Zen 6).<\/p>\n The Hot Chips presentation also revealed the power consumption for the first time. The entire SoC officially has a TDP of 140 W, which is not extremely high, but not exactly low either. The figure matches the TDPs of today’s highend laptop discrete GPUs.<\/p>\n As already known, the processor has 20 cores licensed from Arm\u201410 big Cortex-X925 cores and 10 medium or efficient Cortex-A725 cores. These are last year\u2019s architectures, but Arm has not yet unveiled their successors (in the past the company used to unveil new cores mid-year, but not this year). Each of the two 10-core clusters has its own 16 MB L3 cache, and the processor also features 16 MB of SLC. It serves partly as a kind of L4 cache (allowing data sharing between both big and medium core clusters), but also facilitates data sharing between the CPU and other parts of the SoC without having to go all the way to RAM.<\/p>\n The memory controller has a 256-bit width and supports effective speeds up to 9400 MHz, which would provide 301 GB\/s of bandwidth. DGX Spark devices are designed to be equipped with up to 128 GB of memory.<\/p>\n The SoC provides peripheral connectivity through PCI Express 5.0, with at least eight lanes available. Alongside SSDs, DGX Spark devices also use the PCIe interface for a ConnectX 7 Ethernet adapter, which can link two DGX Spark devices.<\/p>\n Nvidia states that the GPU uses the Blackwell architecture with ray tracing support (featuring 4th generation RT cores), so it does not use a non-graphics GPU architecture like the ones Nvidia uses in server AI accelerators\u2014the PC version of GB10 should therefore allow gaming as well. Naturally, 5th generation tensor cores for AI are also included.<\/p>\n The GPU has 24 MB of L2 cache, intended to improve efficiency with the lower memory bandwidth (compared to standalone GPUs) available, and it also manages memory coherence with the CPU. The GPU supports SR-IOV virtualization as well.<\/p>\n The stated performance is 31 TFLOPS in FP32 compute and up to 1000 TFLOPS in FP4\/NVFP4 precision ops running on the tensor cores (likely with sparsity, half that without). The presentation does not directly mention the number of shaders, but leaked data suggest there should be 6144 (48 SM blocks, which would also mean 48 RT cores and 192 tensor cores). This configuration matches that of the GeForce RTX 5070<\/a>.<\/p>\n Based on the figure of 1000 TOPS, the official boost clock should actually be higher (the RTX 5070 is listed at 988 TOPS), around 2.54 GHz. However, due to the limited TDP, real-world clocks under load will certainly be lower than those of the desktop RTX 5070, even with the 3nm process. The lower memory bandwidth (the RTX 5070 has 672 GB\/s) will also reduce performance. As a result, the device will likely deliver significantly lower performance than a desktop GeForce RTX 5070.<\/p>\n Video output support includes HDMI 2.1a with resolutions up to 8K at 120 Hz. Through USB-C in alternate mode, the SoC can drive displays via DisplayPort, but in that case the maximum resolution supported is 4K at 120 Hz. This imbalance is probably due to MediaTek\u2019s focus on the consumer device, mobile, and television markets. The outputs themselves are handled by MediaTek\u2019s IP, as mentioned earlier, though they are connected to the framebuffer and image output processing block from Nvidia.<\/p>\n Nvidia lists the supported software environments as Ubuntu Linux or its own Linux distribution (DGX Base OS), not Windows. Whether Windows support will be officially added later is unknown. According to leaks, the same silicon also underpins the Nvidia N1X PC processor, for which Windows support is expected. It may simply not be fully ready yet.<\/p>\n The release of the GB10 appears somewhat delayed\u2014Nvidia had originally stated that DGX Spark devices (Project Digits) would ship in July, which apparently did not happen, as they are still not available. The N1X PC processors, if based on the same silicon, were also planned for release this year but that too seem delayed. Nvidia, however, describes its collaboration with MediaTek as successful.<\/p>\n Source: ComputerBase<\/a><\/em><\/p>\n English translation and edit by Jozef Dud\u00e1\u0161<\/em><\/p>\n![\"Prezentace](\"https:\/\/www.hwcooling.net\/wp-content\/uploads\/2025\/09\/Prezentace-procesoru-GB10-od-MediaTeku-a-Nvidie-na-Hot-Chips-2025-04.jpg\")
<\/a>GPU from Nvidia<\/h3>\n
![\"Prezentace](\"https:\/\/www.hwcooling.net\/wp-content\/uploads\/2025\/09\/Prezentace-procesoru-GB10-od-MediaTeku-a-Nvidie-na-Hot-Chips-2025-01.jpg\")
<\/a>\n
3nm process and 140W TDP<\/h3>\n
\n
Cortex-X925<\/h3>\n
![\"Prezentace](\"https:\/\/www.hwcooling.net\/wp-content\/uploads\/2025\/09\/Prezentace-procesoru-GB10-od-MediaTeku-a-Nvidie-na-Hot-Chips-2025-02.jpg\")
<\/a>![\"Prezentace](\"https:\/\/www.hwcooling.net\/wp-content\/uploads\/2025\/09\/Prezentace-procesoru-GB10-od-MediaTeku-a-Nvidie-na-Hot-Chips-2025-05.jpg\")
<\/a>Integrated GeForce RTX 5070<\/h3>\n
![\"Prezentace](\"https:\/\/www.hwcooling.net\/wp-content\/uploads\/2025\/09\/Prezentace-procesoru-GB10-od-MediaTeku-a-Nvidie-na-Hot-Chips-2025-03.jpg\")
<\/a>![\"\"](\"https:\/\/www.hwcooling.net\/wp-content\/uploads\/2025\/06\/ARM-procesor-Nvidia-GB10-1600.jpg\")
<\/a>Version for Windows PC?<\/h3>\n