2
\$\begingroup\$

I'm designing a custom FPGA board, for something low-cost like a Xilinx Spartan 6. I want to perform research about solving memory-intensive algorithms in an FPGA.

As we all know, memory bandwidth is often a bottleneck, especially in low-cost solutions like Spartan 6. However a middle-end GPU has 150+ GB/s of memory bandwidth.

Is there any way to increase bandwidth in a low-cost FPGA to near-GPU levels?

I see only few ways:

  • Connecting high-bandwidth memory like DDR4 to GPU chip and connecting all to the FPGA (kinda strange solution and I don't know if it's feasible and, if it is, won't the bandwidth between GPU and FPGA become a bottleneck?)
  • Using multiple wide and fast memory interfaces to connect off-chip memory to FPGA
  • Using custom controllers, connections or something else, optimized especially for this task to improve bandwidth

I care for at least 100 GB/s. On a low-cost Spartan 6 FPGA, this bandwidth would be success. Or it's impossible with this piece of hardware at all?

\$\endgroup\$
7
  • \$\begingroup\$ Achieving 100GigaBytes on anything "low cost" just isn't going to happen without an ASIC, but, achieving 100Gigabits is doable, QDRIV SRAM chips have a pair of bi-directional 36bit busses, affordable FPGAs top out at ~1.4gbits per pin (slowst QDRIV chip I've seen is 1.333Gbit), this gives ~96Gbits of (low latency, low overhead) bidirectional bandwidth without the complexity of a very wide DDR3 bus \$\endgroup\$
    – Sam
    Commented Feb 19, 2018 at 20:08
  • \$\begingroup\$ Spartan 6 came out it 2009. Current Virtex Ultrascale+ parts can be bought with High Bandwidth Memory (HBM) which gets you 460GB/s or Hybrid Memory Cubes (HMC) for 160GB/s. xilinx.com/support/documentation/white_papers/wp485-hbm.pdf \$\endgroup\$
    – ks0ze
    Commented Feb 19, 2018 at 20:50
  • \$\begingroup\$ @Sam, but the OP clearly said GB not Gb so...? \$\endgroup\$
    – TonyM
    Commented Feb 19, 2018 at 22:40
  • \$\begingroup\$ Are you making a cryptocurrency miner? (I have no reason for asking this beyond wondering if my guess is correct) \$\endgroup\$
    – user20574
    Commented Feb 20, 2018 at 4:19
  • \$\begingroup\$ @TonyM True, but the number of times I've seen GB instead of Gb... the question was asking about hitting 100GB on a spartan 6 which is far in excess of what any spartan 6 could realistically achieve so I thought I'd offer some advice just in case it was supposed to be 100Gb. \$\endgroup\$
    – Sam
    Commented Feb 21, 2018 at 7:55

1 Answer 1

5
\$\begingroup\$

I care for at least 100GB/s. On low-cost Spartan 6 FPGA this bandwidth would be success. Or it's impossible with this piece of hardware at all?

Impossible. Let's assume you're using a 512-bit memory interface, consisting of 8 memory modules in parallel. This would be very close to the maximum user I/Os on the largest available Spartan-6 part (540 user I/Os on the XC6SLX150T), and even then you might run over the limit with control signals. Even assuming this memory interface was possible, 100 GB/sec would require an I/O clock rate of ~1.5 GHz (assuming 100% efficient memory access!), which is unlikely to be attainable on a Spartan-6.

For memory-heavy applications, a GPU is often a surprisingly good solution.

\$\endgroup\$
3
  • \$\begingroup\$ I thought so. So I need a GPU high memory bandwidth and FPGA speed in order to create extreme hardware for extreme memory tasks? Is this combination feasible? \$\endgroup\$ Commented Feb 19, 2018 at 19:43
  • 3
    \$\begingroup\$ No, you just need to remove the "low-cost" constraint. FPGAs with HBM (high-bandwidth memory) exist; they're just orders of magnitude more expensive than the devices you're currently looking at. \$\endgroup\$
    – user39382
    Commented Feb 19, 2018 at 19:48
  • 4
    \$\begingroup\$ There is one more caveat for FPGA regarding wide bus switching: FPGAs usually have a limit on how many pins can be toggled in parallel, due to limits in internal power distribution and resulting current starvation of I/O ring. \$\endgroup\$ Commented Feb 19, 2018 at 20:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.