In this episode of the Tech Poutine podcast, hosts Ian Cutress and George from Chips and Cheese discuss a range of recent developments in the semiconductor industry. The conversation covers the long-delayed tape-out of SiPearl's Rhea 1 processor, a European HPC chip whose specifications (TSMC N6, ARM Neoverse V1, HBM2e) are critiqued as being uncompetitive against existing hardware by its projected 2026 launch. They also analyze GlobalFoundries' strategic acquisition of the RISC-V design house MIPS, positioning it as a move to create more vertically integrated solutions for "essential node" markets like automotive and embedded systems.
The discussion includes a critique of a clause in NVIDIA's Enterprise AI software license that restricts the publication of benchmark data, leading to a broader conversation on corporate control over information. Other key topics include the announcement of IBM's Power11 processor, highlighting its under-marketed innovation in advanced packaging with integrated capacitors; the finalization of the JEDEC LPDDR6 standard introducing non-binary memory channels; the persistent issue of AI model fallibility, demonstrated with a real-world example of Grok providing incorrect information; and the debate on AI's energy consumption, concluding that power delivery infrastructure, not generation capacity, is the primary bottleneck.
The Tech Poutine: SiPearl's Delays, MIPS's New Home, and NVIDIA's Rules
0:04:31 Weird Hardware Access: George reveals that his publication, Chips and Cheese, has access to a Huawei DPU (Data Processing Unit) which features a Kunpeng 920 compute chiplet with 24 ARM Taishan v110 cores.
0:08:05 A Switch to Intel: Ian discusses his recent laptop upgrade, moving from an AMD-based system to a Lenovo IdeaPad Pro 5 with an Intel Core Ultra 9 285H, which he has equipped with 10 terabytes of M.2 storage.
0:20:05 Europe's Delayed Supercomputer Chip: The hosts discuss the tape-out of SiPearl's Rhea 1 CPU, intended for Europe's Jupiter exascale system. The chip, built on TSMC N6 with 80 ARM Neoverse V1 cores and 64GB of HBM2e, is now scheduled for sampling in 2026.
0:22:44 A "Functionally Redundant" Chip: A major critique of the SiPearl Rhea 1 is its competitiveness. By the time it launches, it will be significantly outperformed by existing products like AMD's HBM-equipped Epic CPUs, which have more cores and faster HBM3 memory, leading the hosts to label the project a "sunk cost fallacy."
0:47:39 A Foundry Buys a CPU Designer: The discussion moves to GlobalFoundries' acquisition of MIPS, which has pivoted from its legacy ISA to become a RISC-V design house.
0:54:46 The MIPS Acquisition Strategy: The hosts analyze the deal's logic: MIPS will operate as a standalone IP provider but will optimize its RISC-V designs for GlobalFoundries' "essential nodes" (e.g., 22FDX), strengthening GF's offerings for growing markets like automotive and embedded systems.
1:10:00 NVIDIA's Benchmark Restrictions: George initiates a rant about a clause in NVIDIA's Enterprise AI software license agreement that prohibits users from publishing performance benchmarks or competitive analysis without explicit written permission from NVIDIA.
1:38:20 IBM Power11's Hidden Innovation: IBM's new Power11 processor is analyzed as an evolutionary update to Power10 ("Power10+"). The hosts note that its most significant innovation—the use of an interposer with integrated stacked capacitors for advanced power delivery—was poorly highlighted in its business-focused marketing.
2:33:23 JEDEC Finalizes LPDDR6: The new LPDDR6 memory standard has been released. Its key architectural change is the move to non-binary channels, featuring a 24-bit channel composed of two 12-bit sub-channels, offering greater flexibility for system designers.
2:46:10 AI's Fact-Checking Problem: Using an example where the AI chatbot Grok provided completely incorrect details about an astronomical image, the hosts discuss the ongoing issue of AI model fallibility and "hallucinations."
3:00:14 The AI Energy Debate: The podcast concludes with a discussion on AI's power consumption. The hosts argue that while the power demand from new AI data centers is significant (est. 6 GW in the last year), it's dwarfed by new renewable energy generation (e.g., China adding 60 GW of solar). The real bottleneck for AI expansion is the power delivery infrastructure, such as transformers and substations.