Anne-Vir­ginie Sal­sac, Direc­tor of Research at the CNRS and Flo­ri­an De Vuyst, Full Pro­fes­sor at UTC, researchers at the Bio­me­chan­ics and Bio­engi­neer­ing Lab­o­ra­to­ry (BMBI), are host­ing two super­com­put­ers ded­i­cat­ed in par­tic­u­lar to appli­ca­tions in bio­med­i­cine and bioengineering.

Sup­plied by Nvidia and fund­ed by the Euro­pean Research Coun­cil (ERC), this new-gen­er­a­tion high­per­for­mance com­put­ing (HPC) equip­ment makes UTC a pio­neer among French engi­neer­ing schools. So, what seemed like a real Christ­mas sto­ry came true? Dur­ing the sum­mer of 2023 in Mar­seille, I took part in a sum­mer school on inten­sive com­put­ing and GPUs (Graph­ics Pro­cess­ing Units). It was there that mem­bers of the CEA, with whom I had worked for a long time before, told me about the immi­nent release of these proces­sors, which I was able to pre­view and meet Cristel Saude­mont, Direc­tor of Nvidia France. The com­pa­ny was prepar­ing the world­wide release of these ‘Super­Chips’, whose name ‘GH200 Grace Hop­per* ‘ was a trib­ute to the woman who invent­ed the Cobol machinelan­guage», Flo­ri­an De Vuyst recounts.

The oppor­tu­ni­ty to have machines that are a con­cen­trate of inno­v­a­tive tech­nolo­gies was obvi­ous, but the price was an obsta­cle. In the end, the solu­tion came from Nvidia. «We learned that the group was offer­ing research insti­tu­tions the pos­si­bil­i­ty of acquir­ing a max­i­mum of two super­com­put­ers at a much low­er price than that ini­tial­ly quot­ed. The icing on the cake was that the ERC agreed to finance the cost of the two machines,» explains Anne-Vir­ginie Salsac.

What are the fea­tures of the GH200 Grace Hop­per? «This «Super­Chip» com­pris­es two types of proces­sors placed side by side. The first, the CPU, is a con­ven­tion­al proces­sor with 72 cores or ARM­type log­ic pro­cess­ing units; the sec­ond is a GPU (graph­ic proces­sor) with 20 000 cores, com­pared with a few thou­sand in recent con­ven­tion­al GPUs. In our algo­rithms, it’s not so much the com­put­ing pow­er that slows us down as the com­mu­ni­ca­tions, i.e., the back and forth exchanges between the ele­ments. The Grace Hop­per Super­Chip is the only one to offer this kind of joint archi­tec­ture, which allows us to opti­mise and improve com­mu­ni­ca­tion between ele­ments with­out using inter­me­di­ary ele­ments such as an exter­nal com­mu­ni­ca­tion bus, for exam­ple, as the bus is inte­grat­ed into the machine. As a result, pro­cess­ing pow­er is mul­ti­plied by a fac­tor of around six­ty Ter­aflops (Tera Float­ing-Point oper­a­tions per sec­ond), equiv­a­lent to the pow­er avail­able to the major com­put­ing cen­tres in the years 2000–2005. Grace Hopper’s oth­er inno­va­tion is a 500 Giga­byte mem­o­ry, where pre­vi­ous­ly we were lim­it­ed to 32 Giga­bytes. This will enable us to do real 3D com­put­ing. Anoth­er advan­tage, which echoes one of the UTC’s major areas of research, is ener­gy con­sump­tion, which, for com­pa­ra­ble per­for­mance, is divid­ed by 1 000, drop­ping from one megawatt for a clus­ter of machines to two kilo­watts», describes Flo­ri­an de Vuyst.

These super-machines will ben­e­fit the research activ­i­ties of UTC in gen­er­al and the UTC-BMBI lab­o­ra­to­ry in par­tic­u­lar, which spe­cialis­es in under­stand­ing the bio­me­chan­ics of the human body and its repair, whether relat­ed to the flow of flu­ids such as blood or lymph, the mus­cu­loskele­tal sys­tem or tis­sue engineering.

Can you quote some projects? «In par­tic­u­lar, we are work­ing on seed­ing bio­ma­te­ri­als with cells to cre­ate faith­ful tis­sue mod­els, mul­ti-scale and mul­ti-physics char­ac­ter­i­sa­tion of tis­sues and the design of med­ical devices. Among the lat­ter, a new implant, designed in col­lab­o­ra­tion with Hos­pi­tal Hen­ri Mon­dor and the CNRS, was patent­ed in 2018 with the aim of repair­ing the mitral valve by pass­ing through the blood ves­sels with­out open­ing the heart. Anoth­er project involves micro­cap­sules pro­tect­ing an active sub­stance, such as a drug, which can be inject­ed to bind to a spe­cif­ic tar­get area,» explains Anne-Vir­ginie Salsac.

Study­ing the dynam­ics of these devices requires com­plex numer­i­cal sim­u­la­tions, because of the strong inter­ac­tions between their movement/ defor­ma­tion and blood flow. These are areas where sim­u­la­tion needs are con­sid­er­able. «The Grace Hop­pers are going to change the game. Their com­put­ing pow­er will make us much more effi­cient. We should be mov­ing towards com­put­ing times that are com­pat­i­ble with clin­i­cal prac­tice», she concludes.