Anonymous
6/12/2025, 5:56:25 PM No.105572471
How come I can't buy a PCI E to PCI E cable and get 128 GB/s (that's bytes not bits) transfer speeds between two computers?
Replies:
Thread 105572471 - /g/ [Archived: 1190 hours ago]
Anonymous
6/12/2025, 6:03:09 PM No.105572528
>>105572485
It can send data to a graphics card but for some reason can't to another computer even if it's 10 cm away?
It can send data to a graphics card but for some reason can't to another computer even if it's 10 cm away?
Replies:
Anonymous
6/12/2025, 6:16:40 PM No.105572634
Same as with USB you can't just directly connect two PCIe hosts. For USB there is USB OTG to work around that but for PCIe there is nothing.
Anonymous
6/12/2025, 6:18:19 PM No.105572647
>>105572471 (OP)
usecase?
usecase?
Anonymous
6/12/2025, 6:19:24 PM No.105572662
>>105572471 (OP)
That's how early Thunderbolt worked.
It's a massive security risk because PCIe devices have direct access to RAM.
That's how early Thunderbolt worked.
It's a massive security risk because PCIe devices have direct access to RAM.
Anonymous
6/12/2025, 6:24:37 PM No.105572699
Buy? because theres no reason to do such a thing
theres nothing stopping you from making a really long pcie cable and writing drivers to connect two computers, but literally whats the point when you can get a 200g networking card that you'd never even be able to max out anyways?
theres nothing stopping you from making a really long pcie cable and writing drivers to connect two computers, but literally whats the point when you can get a 200g networking card that you'd never even be able to max out anyways?
Replies:
Anonymous
6/12/2025, 8:59:38 PM No.105574550
Replies:
Anonymous
6/12/2025, 9:09:10 PM No.105574665
>>105572528
No practical use for combining two computers.
Now there is a practical use case for combining tens of systems together but this requires changing from PCIe signals to something more robust and but slower like Ethernet for example.
Within some network cards there is a scheme called RDMA which can be used to directly access the memory address space of other systems, you can transfer data about as fast as you can link systems together, to the tune of 200-400 gigabits a second.
No practical use for combining two computers.
Now there is a practical use case for combining tens of systems together but this requires changing from PCIe signals to something more robust and but slower like Ethernet for example.
Within some network cards there is a scheme called RDMA which can be used to directly access the memory address space of other systems, you can transfer data about as fast as you can link systems together, to the tune of 200-400 gigabits a second.
Replies:
Anonymous
6/12/2025, 9:18:13 PM No.105574801
Replies:
Anonymous
6/12/2025, 10:47:16 PM No.105575833
>>105574550
>>105574801
fuck off glowie, you already know full well that there nothing reasonable that needs that kind of bandwidth, and if you were someone on the cutting edge, youd already be smart enough to get investors OR youre already well funded anyways and your 64 core workstation with its fucking nvidia AI node isn't enough for you.
makes no sense, youre not a real person OP
>>105574801
fuck off glowie, you already know full well that there nothing reasonable that needs that kind of bandwidth, and if you were someone on the cutting edge, youd already be smart enough to get investors OR youre already well funded anyways and your 64 core workstation with its fucking nvidia AI node isn't enough for you.
makes no sense, youre not a real person OP
Replies:
Anonymous
6/13/2025, 1:00:57 AM No.105577047
>>105575833
>high performance should be gate-kept for some reason and unavailable to regular consumers despite the technical feasibility
>high performance should be gate-kept for some reason and unavailable to regular consumers despite the technical feasibility
Anonymous
6/13/2025, 4:55:58 AM No.105578575
>>105575833
>you already know full well that there nothing reasonable that needs that kind of bandwidth
https://www.nvidia.com/en-us/data-center/gb200-nvl72/
>NVLink Bandwidth
>130TB/s
>you already know full well that there nothing reasonable that needs that kind of bandwidth
https://www.nvidia.com/en-us/data-center/gb200-nvl72/
>NVLink Bandwidth
>130TB/s
Anonymous
6/13/2025, 8:15:14 AM No.105579512
>>105572471 (OP)
Because they're peripherals.
It's in the name: Peripheral Component Interconnect. There is a pseudo-direct connection-method - GPU-to-GPU - achievable through FPGA (Field-Programmable Gate Array) boards which OpenAI and Microshit use in their Mariner and Gridcoin servers (part of Azure and OpenAI Compute). They GOUs are connected to each other through the FPGAs, with MMF, using either LC or MOO-DAC connectors.
Bus management and responsibility includes frequency matching, impedance matching, access and backoff algorithms, and IO sheduling, among other activities. Traditionally, the GP CPU ingests these requirements as simple opcodes and interrupts along a memory-management route with attendant syscalls in kernel libraries or stacks, but GPUs COULD do all of that themselves.
Things is, at that point, you'd be reinventing the motherboard, simply with the GPU as the underlying ULSI foundation instead of the CPU.
Believe it or not, some AI organizations are already doing that, like the above-mentioned EvilCorps.
Because they're peripherals.
It's in the name: Peripheral Component Interconnect. There is a pseudo-direct connection-method - GPU-to-GPU - achievable through FPGA (Field-Programmable Gate Array) boards which OpenAI and Microshit use in their Mariner and Gridcoin servers (part of Azure and OpenAI Compute). They GOUs are connected to each other through the FPGAs, with MMF, using either LC or MOO-DAC connectors.
Bus management and responsibility includes frequency matching, impedance matching, access and backoff algorithms, and IO sheduling, among other activities. Traditionally, the GP CPU ingests these requirements as simple opcodes and interrupts along a memory-management route with attendant syscalls in kernel libraries or stacks, but GPUs COULD do all of that themselves.
Things is, at that point, you'd be reinventing the motherboard, simply with the GPU as the underlying ULSI foundation instead of the CPU.
Believe it or not, some AI organizations are already doing that, like the above-mentioned EvilCorps.
Anonymous
6/13/2025, 8:21:52 AM No.105579535
>>105572528
Yes. You can hardly send pcie4 over a cable, never mind pcie 6.
Yes. You can hardly send pcie4 over a cable, never mind pcie 6.
Replies:
Anonymous
6/13/2025, 8:32:18 AM No.105579571
>>105579535
Ever heard of a retimer
Ever heard of a retimer
Anonymous
6/13/2025, 8:37:29 AM No.105579588
>>105572471 (OP)
Attenuation.
Attenuation.