Intel Xeon processors have dominated the
market for over a decade and so the emergence of a new server processor
from AMD might be considered an event but not an advent. AMD recently
published a theoretical study indicating that EYPC would achieve the
same or higher level of performance and productivity than Xeon for the
same or less price in terms of Total Cost of Ownership (TCO). Though it
was a marketing piece, it did contain enough arguments to support its
claims. Perhaps it is a wake-up call to the market and so this forum
wakes up. Ref:
https://www.nextplatform.com/2018/09/27/virtualization-is-the-real-opportunity-for-epyc/
EPYC
is a multi-chip module processor meaning that the processor is not made
of one die but is a package of multiple dies. This approach allows
more processor cores to be packaged without incurring the high cost of
fabricating high core density dies. The concept may be simple, but can
it be implemented well in practice?
This is what the Compucon
team has done recently by hands-on benchmark testing (in September and
October 2018) to find out the productivity of the implementation. The
test sample is the EYPC 7351P- single socket, 16 cores, 32 threads, 64MB L3,
and 2.4GHz base frequency supporting 8 channels of DDR4 memory transfer
and 128 PCIe links with a TDP of 155W. It is made from a 14nm manufacturing process.
This specification is very impressive.
Compucon tested the EPYC
with a popular open source scientific and mathematical application
called Cholesky Factorization. The runtime using 16 cores achieved a
speed up of 15.8 times over 1 core. This is a very high score; though
not a perfect score, perfection is not achievable in real life. In
fact, this application requires data to be transferred between cores
before computing can proceed during the entire computing process, and
data transfer took up some clock time. If the application does not need
inter-core data transfers, the speed up may reach 15.99 or something along those. The test result confirms that the multi-chip packaging
arrangement has not introduced extra latency to computation and this is
good.
EPYC has 2 other merits over Xeon. One is the higher
memory bandwidth from 8 channels. Xeon has 6 channels only. The other
one is the support of 128 PCIe links by a single socket processor. This
quantity is higher than that supplied by 2 Xeon sockets. A GPU card needs 16 of them, so a single socket motherboard has more than enough bandwidth to support 4 GPU cards and more, subject to the real estate of the motherboard form factor and the intended design criteria decided by the motherboard maker.
All in all, EPYC is impressive.
Benefits