Page 23 - EE Times Europe Magazine – November 2023
P. 23

EE|Times EUROPE   23

                                                           The Mind-Boggling Cost of Generative AI Ownership


                                                                                The latest algorithms pose
         Attributes               Google Search      GPT-3         GPT-4
                                                                                a challenge to the current
         Cost Per Query               ¢0.2            ~¢3           ~¢10
                                                                                state-of-the-art processing
        Table 2: Comparing the cost per query of GPT-3 and GPT-4 against Google search shows   hardware, and GenAI
        the leap in cost associated with GPT-4. (Source: Vsora)
                                                                                accelerators are not
                                                                                keeping up.
          •  Theoretical throughput of one     (25 W / 0.007 queries per second)
           leading-edge GenAI system processing   •  Total energy cost: US$0.11 per kWh
           ChatGPT-4: ~0.055 queries/second   •  Energy cost per query: US$1.2 –4  processing hardware, and GenAI accelerators
          •  Number of systems needed to meet a   •  Total power consumption for    are not keeping up. In fact, no hardware on
           processing capability of            100,000 queries/second: ~363.7 MW  the market today is capable of running the
           100,000 queries/second:            The energy cost amounts to about    full GPT-4.
           ~1,800,000 (100,000 / 0.055)     US$1 million per day (power consumption for   Current LLM development efforts that focus
          •  Total acquisition cost:        the chips × 24 hours × 0.11).       on creating smaller but more specialized LLMs
           ~US$900,000,000,000                Clearly, the cost is dominated by hardware   that can run on existing hardware are a diver-
           (1,800,000 × 500,000), approaching    acquisition.                   sion. The GenAI industry needs semiconductor
           US$1 trillion                                                        innovations in computing methods and archi-
          The daily depreciation amounts to about   The best-guess total daily cost is in the ball-  tectures capable of delivering performance of
        ~US$820 million (900,000,000,000 / 1,095).  park of US$820 million.     multiple petaFLOPS with efficiency greater than
                                            The above leads to a GPT-4 cost per query for   50%, reducing latency to less than 2 seconds per
        Estimated energy costs to execute the   a system running 100,000 queries/second of   query, constraining energy consumption and
        hardware                            $US0.095 (820,000,000 / (100,000 × 24 × 60 × 60)   shrinking cost to $US0.002 per query.
        Assumptions:                        [(cost per day) / (# of queries × # of hours × # of   Once this is in place—and it is only a
          •  Average power consumption per chip:    seconds)]. Table 2 compares costs per query.  matter of time—the promise of transformers’
           25 W, based on nominal power, efficiency                             deployment on edge devices will be fully
           and memory bandwidth             FULFILLING THE PROMISE OF GenAI AT   exploited. ■
          •  Throughput per chip:           THE EDGE
           ~0.007 queries/second (0.055 / 8)  The latest algorithms, such as GPT-4, pose   Lauro Rizzatti is a business adviser for
          •  Energy consumption per query: 3,637 J   a challenge to the current state-of-the-art   Vsora.







                                                                  GaN & SiC: Design,


                                                                     Devices and the

                                                                     Evolving Market




                                                                         Explore the future of
                                                                      power electronics with our

                                                                      comprehensive report on
                                                                  wide-bandgap semiconductors:
                                                                              SiC and GaN.


                                                                          DOWNLOAD THE EBOOK:

                                                                  https://www.powerelectronicsnews.com/
                                                                      special-report-gan-sic-worldwide/
   18   19   20   21   22   23   24   25   26   27   28