[loeg]

According to wikipedia, the fastest GDDR5 can do 256 Gbit per chip[0]. I don't know how many chips are typically used. MCDRAM in the article does 400 GB/s, or 3,200 Gb/s. That would require 12.5 of those GDDR5 chips, assuming they scale linearly.

[0]: https://en.wikipedia.org/wiki/GDDR5_SDRAM#Commercial_impleme...

[throwawayish]

Speed and size are somewhat independent. Speed is set by the bus frequency and how many chips you use - every chip has a fixed data bus width, typically 32 bits, while overall size is then how many chips you have times how big each chip is.

Unlike typical computer memory architectures, where the memory bus connects multiple chips or modules to one controller, GDDR doesn't do that; every slice of the memory controller only speaks with a single chip, strictly point-to-point. (Reducing bus load and layout issues and thus allowing higher clock rates).

That's why, with GPUs, it's usually sufficient to say how wide the bus is (often 64 - 128 - 256 - 384 - 512 bits) to get a rough idea of it's performance, since memory clock frequencies occupy a rather narrow range. (However, narrow-bus, lower-end GPUs often don't use the same technology as higher-end GPUs, eg. DDR3 instead of GDDR5)

[shaklee3]

GDDR5 can typically do 240GB/s access time on a typical GPU, and there are multiple chips on many cards (Tesla K80). The newer cards use HBM2 and can do 732GB/s (http://www.nvidia.com/object/tesla-p100.html).

[AlphaSite]

HBM2 is 1024GB/s (256 per stack).

[shaklee3]

Kind of. Nvidia lowered the voltage on their P100 so it does not hit those rates. Theoretically it can go that high, but the power draw was too large. Next gen we'll likely see that.

[astrodust]

1TB/s is pretty nuts by today's standards but I bet it'll elicit a yawn in ten years time. Amazing indeed.