No it’s happening right now
What tsk said.
Yes it is.
What no 16/32bit tensor core? Where’s muh raytracing?
Vega FE has 13.1 tflops FP32 for reference.
Presumably 7nm means significantly less space-heater effect… but that’s a little underwhelming of a bump generation to generation.
did VegaFE including packed int4? (i.e. int4 performance at 8x 32bit?)
Nope, I don’t think it did.
If the leaks are true, Vega 7nm benches about the same as a GTX1080, which is of course not an improvement over the old Vega64. This implies that they used the die shrink to reduce power usage and die size, rather than increasing the number of cores and clocking higher. And that implies a much lower cost to manufacture and thus a lower price for consumers.
The RTX 2070 offers basically that same performance and starts around $500. If AMD can substantially undercut that and price at $349 or less, they would really put the hurt on Nvidia.
I mean the 7nm Vega won’t be in that market so it won’t really matter.
But that performance makes sense, a die shrink in itself doesn’t add much.
You can do a number of things with a die shrink.
- It’s more power efficient, so you can either clock it higher or reduce its power usage.
- Each core consumes less space, so you can either make a smaller chip at the same performance or pack more cores in the same size to make it faster.
If AMD chose to reduce power usage and produce a smaller chip that would cut their costs considerably, all at the same performance. Less cooling and each chip takes up a smaller portion of the die.
Indeed, and here it looks like the shrunk the same die and added some deep learning goodies.
Ahh. Well yeah, you could also use that extra space with stuff that doesn’t help games. Nvidia did that too. Basically.
Get ze glue!
This seems really smart and will probably cut manufacturing cost.
Also will be able to transition to new nodes much quicker due to no yuuuuge chips(like that skylake 28 core)
Zen 2 dies only have 8 cores. No mention of CCXes, maybe they’ve done away with the split, making the (core chiplet) dies just straight-up 8 cores+cache.
I think the upscaling of cores comes from just 8 x 8-core chiplets on the highest-end EPYCs. Also makes it much easier and cost-effective to create regular desktop CPUs with 8 or 6 cores, and any variants in between, because they can vary the number of chiplets instead of disabling cores on non-faulty dies.
Conversation is already happening at my office about switching to (a) variants in AWS.