High Performance Computing

Local Memory and/or CacheOK, this subject is just too big. High performance computing covers everything from hot graphics workstations to servers to supercomputers, and each of them has different needs and constraints. Since the memory industry has been almost exclusively focused on the needs of the compute market, most of their needs are being fairly well served. But the high performance fringe keeps bumping against the stops … and that’s where Tezzaron® comes in.

highperformance-diagramSupercomputers have been our first adopters. We offer more memory in less space at lower power with higher bandwidth and lower latency than any other vendor. When you use terabytes of memory in your machine, having a memory that is twice as dense and eight times faster that consumes 1/10th the power … well, those differences add up. (Maybe you don’t need to install a new power plant next door after all.) Some of you will recognize the hub, switch, optical, and cooling blocks on our product pages as elements in the total solution we offer the supercomputer market. If you are building a graphics processor, some of those memory subsystem components may not be of use to you at all. But imagine what happens when the capability that produced them is brought to bear on your problem.

What about servers? What Tezzaron® can do for servers is as varied as server architectures. Designers intent on using conventional architectures can benefit from all the basic goodness of Tezzaron® DiRAM™ architecture (better power, space, latency) but those advantages are muted by the interfaces and packaging that define “conventional architecture” in this space. So … what about the folks who want to reach higher? What can be done? Well, suppose you had 64 processors in one package. Suppose they were all 64 bit processors, and each of them had 1Gb of local RAM in the same package, with 9ns latency (what we used to call Fast SRAM latency). Now, what if you had a whole array of those packages on your board and racks of them for your server farm? Now what can you do?

graphics-processor-diagramIt is instructive to note that, in the commercial products space, graphics processor developers have been pushing the hardest to get 2.5D RAM mated to their processors. Why? We sound like a broken record – it is the same story: power, latency, density, transaction rate, physical space. Go look inside a teenage boy’s PC and you are likely to find the most aggressive technology on the consumer market. And you are likely to find a bunch of fans and may even find a liquid cooling system. Go look in the computers of the folks who make the games they are playing or the movies you have been watching lately and you will begin to see the scale of the challenge faced by graphics hardware developers. They have pushed every corner of the performance envelope to the limit. Now it is our turn to help.