TECH TALK Where Are the PCIe 4.0 Graphics Cards & Mobos?
It’s been over a decade since PCI-SIG announced the preliminary specs for what would become PCIe 3.0, and we’ve been able to use Gen3 graphics cards since 2010. The PCIe 4.0 specs, targeting roughly double the bandwidth per pin of Gen3, started the next y
Going a step further, PCI-SIG announced the initial plans for PCIe 5.0 alongside the Gen4 final specs, targeting 32GT/s per pin. Yet we still don’t have any PCIe 4.0 graphics cards or motherboards.
It’s not that Gen4 solutions don’t exist—IBM is shipping servers with PCIe 4.0 slots—but on the consumer side, not even Nvidia’s Turing architecture supports Gen4. That’s largely because most consumer workloads don’t need the extra bandwidth. Gaming on a Gen3 x8 slot causes almost no loss in performance, and with SLI and CrossFire seeing less support and use, we’ve been resting on the 8GT/s plateau for years.
I said “most” consumer workloads, as there are areas where we’d benefit from Gen4’s added bandwidth. Intel’s mainstream consumer CPUs have 20 Gen3 PCIe lanes, with four used for the DMI 3.0 link to the PCH chipset. AMD’s AM4 Ryzens have 24 Gen3 lanes, again with four for the chipset—four more are dedicated to the first M.2 NVMe slot. While our graphics cards generally don’t need a faster PCIe link, especially with x16 connections, chipsets can run into bottlenecks.
Consider Intel’s Z370 chipset, which can support up to 24 PCIe Gen3 links for M.2 slots, USB ports, network connections, audio, and additional PCIe slots. All that data still goes through a tiny straw. Most of the time, we don’t hit everything at once, so it’s not a problem, but if you try doing RAID 0 with two M.2 SSDs, the chipset link is a limiting factor, and performance isn’t much better than the fastest single M.2 drives.
M.2 SSDs in general are another area where Gen3 connections are a bottleneck. Many NVMe drives can saturate the interface, topping out at 3.5–3.8GB/s for sequential transfers. With the size of the M.2 connector, going to a wider x8 PCIe link isn’t an option—PCIe 4.0 would solve this. USB 3.1 Gen2 is another relatively recent tech, but with 10Gb/s per port, it would only take four ports to saturate the chipset to CPU link.
Graphics cards could also benefit in some workloads—like machine learning and GPGPU. And even though SLI and CrossFire see less use these days, doubling the PCIe speed would mean consumer systems like LGA1151 and AM4 could provide two x8 Gen4 connections that would have the same bandwidth as x16 Gen3.
I wonder if we might skip Gen4 support. There’s cost involved in bringing a product through validation and testing, so why test and validate against Gen4, then repeat the process a year later for Gen5? Then again, if we’ve been OK with Gen3 bandwidth for seven years, doubling it should carry us well into the next decade.
We don’t necessarily need the extra bandwidth now, but there are scenarios where it could be useful in the future. Leaving it out of current-generation motherboards and chipsets means they’re more likely to need an update. It also means we’re less likely to see SSDs and graphics cards adopt Gen4.
It’s the classic chicken and egg scenario, and for now, PCIe 4.0 is solely used in supercomputers. But I’m hopeful that 2019 will change things, and increase our chipset link bandwidth.
While graphics cards generally don’t need a faster PCIe link, chipsets run into bottlenecks.