This number -- 14 Mbytes -- comes from the Miercom report. Miercom limited their testing to
packet sizes below 1500. They did not test jumbo frames. Their results seem to indicate that
buffer size is limited by packet count, so a better showing might be expected for 9000 byte
All the switches that Miercom has tested use a buffer particle architecture that chains together fixed length chunks to store large frames. Give or take one buffer particle -- typically 300 bytes -- only the amount of memory needed is used for each packet. If Miercom results are to be believed, none of the manufacturers tested (Broadcom, HP, Cisco, Arista) can successfully use chained buffers to hold packets without wasting memory. Either that, or there is something wrong with the design of Mierom's tests.
Miercom competes with Tolly in the for-hire testing business. Tolly's tests that include microburst capability come to generally a different result. See IBM 8264 review for references.
Other tests performed by the for-hire community follow test plans laid out in standards RFCs. But there is no guidance on how testers should measure microburst