PCIe 4.0 Is 2x Bandwidth of 3.0, like 3.0 to 2.0 and 2.0 to 1.0

Subject: General Tech, Motherboards | June 7, 2014 - 05:47 PM |
Tagged: pcie 4.0, pcie, PCI SIG

You know the PCI-SIG might break the pattern with PCIe 5.0, just to mess with us. But for right now, Tom's Hardware seems to have acquired part of the PCIe 4.0 spec and it is expected to get 2 GB/s bandwidth per lane, per direction. This is double the bandwidth of PCIe 3.0, continuing the trend of each major PCIe release doubling bandwidth of the previous major version.

View Full Size

A 16-lane PCIe 4.0-compliant graphics card or storage add-in board (that feels so weird to write...) has a maximum bandwidth of 32 GB/s inbound and 32 GB/s outbound, 64 GB/s total. This is still below GDDR5 bandwidth, but approaching the same order of magnitude. That said, memory bandwidth is the major roadblock for optimizing GPGPU workloads, already. APUs will probably still have an advantage in CPU and GPU tag-teaming tasks, despite their lower compute performance.

According to bit-tech, the spec is expected to arrive with Skylake and its 100-series chipset.

June 7, 2014 | 05:59 PM - Posted by Shahem (not verified)

Speaking of APUs edging out the old PC arch, what are the chances PCs will struggle to play next-gen games because of the lower latency offered by the PS4/Xbox One ?

Aren't there cases where we will be in big trouble, I'm alluding to CPU/GPU working together.

June 7, 2014 | 09:03 PM - Posted by Anonymous (not verified)

Maybe Discrete GPUs will get their own CPUs and use the integrated CPUs to run critical low latency gaming engine code. It would not be too hard to shrink the entire gaming console hardware to fit on a discrete PCI card, hell the Discrete GPUs already have more powerful memory controllers and wide memory buses, so what's more than maybe a few square mm of CPU cores going to going to require, to make a discrete GPU into its own APU, custom tailored for gaming like the gaming APUs are. If they add a few CPU cores, and some on die RAM, to go along with the GDDR5 memory the whole gaming engine could be hosted on the PCI card, and maybe even a custom gaming OS also, the motherboard CPU just there for extra support, and to run the general purpose OS.

June 7, 2014 | 11:03 PM - Posted by Anonymous (not verified)

At that point aren't you just plugging in an entire computing system, into another entire computing system? At which point isn't that a bit redundant?

June 8, 2014 | 05:30 AM - Posted by jcelerier

Actually, not so long ago, in a not so distant galaxy...

http://www.3do.cdinteractive.co.uk/viewtopic.php?t=179

June 8, 2014 | 09:35 AM - Posted by Anonymous (not verified)

Redundant, do you know how much faster a system on die bus is compared to speed of any of the off core data buses are! A CPU placed on Die with the GPU, and sharing a system bus, and the same extra wide GPU style memory/data BUS and Memory controller has the lowest possible latency imaginable. Latency saved by not having to go through any external PCI/whatever protocols to share data between CPU and GPU, and if the CPU/GPU shared die has some on die RAM then the most commonly utilized OS and gaming engine code can be kept staged on the on die RAM. Redundant NO, a discrete gaming APU on a PCI card with say a 512 bit system BUS, to fast GDDR5 memory and a large enough on die ram, would free the motherboard CPU up and bring Gaming to the next level, the motherboard CPU would be relegated to a support role, and not have to be as powerful in a supporting role for gaming, and more.

There are already more CPUs on most computing platforms, the hard drive/SSD drive controller is a CPU, as are other controllers on PC/Laptops/phones. The next level for gaming, is the gaming Cluster, with each discrete GPU/APU being its on computing platform, and hosting its own gaming optimized gaming OS and whatever gaming engine required for running the game. Just imagine a rig with two or more of these complete gaming system/PCI based systems forming a gaming cluster and hosting multiple cooperating gaming engines running different aspects of a 4k or higher game all communicating with each other over PCI 4.0, or their own extra wide backplane BUS, or both. there is no general purpose motherboard CPU that would be able to match the data transfer speeds that a GPU/APU discrete gaming APU would have, by placing the CPU on the same die as the GPU, why do you think the console makers went with a gaming APUs, now imagine that gaming APU made with more/more powerful CPUs and a GPU portion with the GPU core count of high end GPU, with a high end GPU style wide data bus to loads of GDDR5, or faster memory, and Butt loads of On Die RAM for each GPU/CPU(or CPUs) pair.

The motherboard CPU, no matter how powerful, is hamstrung by only having a narrow 64 bit data bus and the higher latency of distance and a few hops(encodes/decodes over any protocol) to even get to the GPU, compared to a CPU that shares the same die and on die, or module, system bus, that a high end discrete gaming APU would have.

For those concerned about a gaming OS to run a cluster, well there are plenty of HPC/Cluster Linux distros out there that could do the job, the HPC/Cluster/ supercomputing world runs on Linux, and all those huge science simulation programs run on CPUs with GPU accelerated vector(graphics) computations, and will be running on future APU type(GPU/CPU) systems, and even most Server SOCs will use APU type systems. AMD will have systems for the HPC/server market, and Nvidia will have its future GPU accelerators on module with IBM Power8(from IBM, or any one that licenses Power8) based systems, and those mezzanine modules buses are fast/wide too.

June 7, 2014 | 06:22 PM - Posted by Anonymous (not verified)

*Note to self: Don't upgrade anything until Skylake.

June 7, 2014 | 07:13 PM - Posted by A Gsync user (not verified)

Totally agree with you.

I'm waiting for at least a 10-12 physical core CPU before I leave X79 3930k and 4930k

June 8, 2014 | 02:11 PM - Posted by Anonymous (not verified)

You can get a 12 Core Xeon as it is on Socket 2011. Xeon E5-2690 v2

June 7, 2014 | 06:34 PM - Posted by Anonymous (not verified)

I'm waiting for PCIe 5.0.

June 7, 2014 | 06:34 PM - Posted by Paul Jones (not verified)

I agree with the previous poster. 4K-3D is my goal, and I want to be a smart shopper and am willing to wait a while.

June 7, 2014 | 07:44 PM - Posted by Scott Michaud

Heh, those tasks are very shader-bound, thankfully for you. You are not really doing any different tasks when you crank up the resolution, just more of them. It's basically all ROPs, shader performance (FLOPs), and memory to hold it. The PCIe bus shouldn't have any meaningful load added to it (unless you move framebuffer(s) to the CPU for some reason).

What this could be useful for is AI, particles, physics, and (of course) PCIe-based SSDs.

June 7, 2014 | 08:33 PM - Posted by Anonymous (not verified)

3D's dead bro. For the 3rd time in as many decades 3D has failed and the industry is once more moving away from it. I do not see 4K changing that (in fact making it more useless) or the industry trying to revive it.

June 8, 2014 | 05:24 AM - Posted by Daniel Nielsen (not verified)

Agree i don't understand why anyone at this point is even considering 3D in their gaming setup.

Post new comment

The content of this field is kept private and will not be shown publicly.
  • Lines and paragraphs break automatically.
  • Allowed HTML tags: <a> <em> <strong> <cite> <code> <ul> <ol> <li> <dl> <dt> <dd> <blockquote><p><br>
  • Web page addresses and e-mail addresses turn into links automatically.

More information about formatting options

By submitting this form, you accept the Mollom privacy policy.