This paper presents parallel versions of a wavelet radiosity algorithm. Wavelet radiosity is based on a general framework of projection methods and wavelet theory. The resulting algorithm has a cost proportional to O(n) versus the O(n^2) complexity of the classical radiosity algorithms. However, designing a parallel wavelet radiosity is challenging because of its irregular and dynamic nature. Since explicit message passing approaches fail to deal with such applications, we have experimented various parallel implementations on a hardware ccNUMA architecture, the SGI Origin2000. Our experiments show that load balancing is a crucial performance issue to handle the dynamic distribution of work and communication, while we do make all reasonable efforts to exploit data locality efficiently. Our best results yield a speed-up of 24 with 36 processors, even when dealing with extremely complex models.