AMD's Zen 2 EPYC Processors will offer up to 64 Physical Cores

I would guess they mean IPC. However, it all depends on how Zen 2 clocks. If it clocks higher (and it should, really) then the difference will be more pronounced.

Now I can see why Intel are bricking it and making 48 core CPUs though. It makes sense now.
 
Only a 2x increase in performance? I'd say that's actually worse than expected considering they double core/thread counts.

Remember that this is a server part and that power and thermals are a major factor. It's impressive that they got a 2x increase from 2x the cores, never mind the 4x increase in floating point.
 
You can't get a 1:1 2x in performance just be doubling core counts (See Amdahls law). Combined with the fact that the number of memory channels has to remain at 8 so bandwidth/core will likely be little more than half, and the fact that you've now got the thermal output of twice the number of cores in the same size area (Which means the benefits of 7nm here have to go towards the same-clocks improved efficiency side of things as opposed to increased clocks), as well as twice the amount of data being thrown around the interconnects, that means that in order to actually attain a x2 increase would require ridiculous amounts of work when it comes to reorganising cache's and interconnect technology to avoid everything bogging down from data transfer bottlenecks, alongside massive increases in efficiency and IPC.
 
Remember that this is a server part and that power and thermals are a major factor. It's impressive that they got a 2x increase from 2x the cores, never mind the 4x increase in floating point.

Same as it was for first gen Epyc. So I don't see that as being a factor considering they will most likely be running low frequencies higher up in the range you go, and considering the switch to 7nm allowing for the same performance at much lower power consumption I bet they will see similar frequencies. Hence the simple 2x increase instead of more as they elected to run same power consumption. Which is probably the only option without wanting to melt the coolers.
FP is a different story considering it has native support whereas the first gen did not. So really anything is impressive on that part
 
Faster memory is the key here I've been told. Throughput has been improved as has latency. The fact that in 1 generation they have improved by a factor of 2 is massive.
AMD's biggest problem will be keeping up with demand...
I have 4 EPYC servers that I can basically drop in this new chip and double my compute (well I'll probably have to swap out the RAM too for faster variants) - fun tomorrow at work roadmap wise.
 
Faster memory is the key here I've been told. Throughput has been improved as has latency. The fact that in 1 generation they have improved by a factor of 2 is massive.
AMD's biggest problem will be keeping up with demand...
I have 4 EPYC servers that I can basically drop in this new chip and double my compute (well I'll probably have to swap out the RAM too for faster variants) - fun tomorrow at work roadmap wise.

I'm fascinated to see how well everything works in practice. The whole memory/PCIe locality stuff will 1st Gen EPYC will go away with Zen 2. I'm very interested in seeing how good memory latency and core-core latency etc is when it has to move through one die to get to another.

TBH, the I/O die is massive, which makes me think that AMD has added some kind of L4 level cache to the chip. It will be interesting to see if the same design moves to mainstream desktop parts.

This chip design is a real slap in the face for Intel, especially after their 48-core announcement, which is basically just two Cascade Lake-SP processors on a single package with the same UPI interconnect as a dual-socket Cascade Lake-SP server.
 
Back
Top