The partnership between Intel and nVidia could indicate upcoming desktop APUs.
The partnership between Intel and nVidia could indicate upcoming desktop APUs.
AMD has been sitting on their arse and scratching their balls for far too long on that front IMO. Just like with ThreadRIppers, they failed to cash-in on the segments that they've created. 8700G is long in the tooth and it's far from what it could've been. I've expected to see far better IMC, more IF bandwidth, especially on single-die solution, aggressive use of 3D cache, leading-edge RDNA gen etc etc. Instead we got a bunch of old/reused parts. Not even CUDIMM support. Is this new alliance about to spur things on AMD side, too ? I'd love to see new AMD APUs with Zen5/6 cores, RDNA 4.5+ (new support in LLVM compiler published months ago for RDNA4.5), 3D cache on both CPU and iGPU chiplet, beefed-up IMC, CUDIMM and LP/CAMM2 support on AM5. Heck, I'd love to see it even as ThreadRipper/Pro or EPYC part, with MRDIMM support for 2x bandwidth. On all platforms, it could be a hell of a solution, bot for gaming, universal computing platform or AI. Why is AMD dragging their feet on this ? Are they just waiting for RDNA5 ?
Desktop APUs remain a relatively obscure category. High-performance desktop APUs have consistently stayed in the niche market. I wouldn't be surprised if they ever shift to mainstream use. This would signal a complete departure from the low/mid-range segment by companies like AMD, Intel, and Nvidia. Desktops differ fundamentally from laptops—they don’t face strict limitations on space or power consumption. That’s why products such as Strix Halo are moving into laptops and mini PCs, offering a viable justification for higher-end pricing in those contexts. Budget APUs saw some interest during recent mining cycles with the Ryzen 3 2200G, but that was a brief period before they faded from mainstream adoption. Intel once boasted the fastest APU with the i7 5775C, yet it remained largely ignored outside specialized communities due to its prohibitive cost relative to performance. This pattern repeats: APUs let you trade off between price and power or performance for space and size constraints. Unless you’re working under tight limitations, it’s always smarter to invest in a more capable solution.
If I'm not mistaken AMD's desktop APUs are essentially repackaged mobile APUs. Why reinvent it? So look to the mobile market and eventually you might see it appear on desktop as low priority. IMC - unlikely outside of a new generation product. IF bandwidth - unlikely to improve, full stop. They need to bin that and move to something better, but hard to do so without either cost and/or power concerns. 3D cache - cost. Leading RDNA - new product. Zen 5, a given. Likewise Zen 6 if you wait long enough. RDNA4.5 might come in 2 years. CAMM is about the only maybe. If it can happen at all, it would require different mobos from normal, and if using LPDDR could offer a small increase in perf over regular DDR. Cost kills it from ever happening. Desktop APUs are only somewhat attractive if you care about GPU performance over CPU performance, but for whatever reason can't go dGPU. Even if they could socket a Strix Halo, it would only be a mainstream tier part for gaming. And look at the cost of it. If you want perf, you need to apply the silicon with the cost that comes with it. As a CPU, it was priced comparably to the i7 K CPUs either side of it. From an enthusiast perspective it had two problems. 1, it was released VERY late in the cycle. I feel it was a launch by Intel to say they launched it at all, with Skylake taking over shortly after. 2, It didn't clock well. This was the OG 14nm CPU and you'll be lucky to push it past 4 GHz. Both Haswell refresh and Skylake could go quite a bit beyond that. The iGPU being fastest at the time also meant little. Anyone serious about gaming would be using dGPU anyway. I think I only ever tried to use it to rank on hwbot.
searching for "nvidia grace hopper" reveals the direction this will take—likely limited to laptop integration or close hardware ties. It doesn’t directly challenge the product lines of TR or EPYC, which are focusing on different segments. There’s essentially no demand for medium-power integrated GPUs in this space; most users either use large cards or IPMI-based video solutions. The EPYC platform is designed to expose 64 PCIe lanes, giving integrators flexibility, but it comes with a steep learning curve and high costs. Prices on EPYC components are already substantial—chips costing over $2,000 and motherboards around $600 are typical. Notably, EPYC boards tend to be cheaper than TR boards, likely because TR is an extremely complex platform with limited market reach.
This approach would open up a new market by repurposing current sockets, which is relatively affordable (look at the most budget-friendly options). It would allow access to numerous additional memory channels, MRDIMM for doubled bandwidth, and significantly higher TDP limits to support both CPU and GPU. Motherboards remain cost-effective too. Such a solution would be far superior to standard CPU + discrete GPU setups. It could deliver strong raw bandwidth, an expansive address space, no 3D caching restrictions, and most importantly, a unified memory architecture. Current 64GiB DDR5 modules at 6400 MHz provide the foundation; doubling them in an MRDIMM yields both a 128GiB stick and a 12.8MHz transfer rate. Before any extreme JEDEC overclocking. On a four-channel motherboard, you'd reach 512GiB at 400GB/s, 1TB at 800GB/s on an eight-channel TR Pro, and 1.5TB at 1.2GiB/s with twelve channels on an EPYC board. No discrete GPU could approach these levels.