in advance this month, AMD unveiled its first twin GPU based
on its highest-cease Fiji GPU. The Radeon pro is virtually a gap card with a
$1500 rate and a specific recognition on growing for VR — but AMD clearly has
designs on a couple of market. nowadays the organisation introduced its FirePro
S9300 x2 — the primary and likely handiest Fiji-derived GPU the business
enterprise will release for the HPC space.
The S3200 x2’s patron roots are seen in its specs. each GPU
is restricted to 4GB of reminiscence in a market where 8GB RAM buffers are
considered low-cease and top-give up cards can p.c. as tons as 32GB of RAM. in
contrast to the Hawaii-based totally S9170, the S9300 x2 lacks aid for ECC
memory and its double-precision floating factor performance is constrained to
at least one/16 of FP32 (0.8 TFLOPS). The S9170, in comparison, gives up to
two.6TFLOPS of double-precision performance.
The S9300 X2’s primary electricity is its
unmarried-precision floating point performance. At thirteen.9TFLOPS, it’s the
maximum effective unmarried-precision card AMD has ever released and
theoretically faster than whatever in Nvidia’s cutting-edge solid of hardware
as well.
The S9300 x2 is designed for passive cooling and a 300W
electricity envelope, which means that clock speeds are considerably lower than
what we usually see in a card of this kind. AMD hasn’t revealed the predicted
clock speeds at the Radeon seasoned Duo, but the GPUs it demoed at GDC had 3
8-pin PCI specific energy connectors. even as agencies usually overprovision
their GPUs to make certain they can handle sharp temporary spikes, the Radeon
seasoned Duo is absolutely designed to pull greater than 300W.
then again, this matches what we noticed with the Radeon
Nano. Fiji’s power consumption rises sharply as its clock pace will increase;
retaining the GPU at a rather low 850MHz guarantees it may fit within its
predicted electricity envelope.
The HBM bottleneck
whilst AMD decided to apply excessive bandwidth reminiscence
(HBM) for its Fiji GPU, it made a conscious tradeoff between memory ability and
uncooked bandwidth. The benefit of HBM is that it offers Fiji a 512GB/s
reminiscence bus in a much smaller footprint even as the usage of lots less
electricity. The downside of first-technology HBM is that it’s confined to a
4GB memory buffer.
Ever in view that AMD introduced Fiji, there’s been
hypothesis that the enterprise would possibly use a few kind of workaround to
construct a GPU with more than 4GB RAM. If this kind of technology changed into
ever going to make an look, the S9300 x2 is in which it might be deployed — the
card’s $6000 charge tag gives it room to bear the value of development.
The turn aspect to this, but, is that the HPC area has a
tendency to be ruled via industry professionals who are privy to the memory
necessities of their very own software program. AMD has already introduced it
will provide an HPC cluster for processing records collected by the Chime
telescope, as shown above, and it also believes it may make inroads inside the
oil and gas commercial enterprise. those are all workloads in which
unmarried-precision floating factor overall performance is king, and the S9300
x2 ought to excel in those regions.
a spot product, but an important step
The S9300 x2 is a niche product even in the confines of the
HPC space, but it’s nevertheless proper to peer AMD stepping up to the plate.
One motive Nvidia has enjoyed tons more healthy margins compared to AMD is that
it holds between 75-85% of the professional GPU marketplace (the percentages
shift rather depending on the zone) and almost all of the HPC area. Of the top
500 structures that use some type of co-processor, 66 use an Nvidia GPU, 29
depend on Xeon Phi, and just three use AMD Radeon.
given that AMD GPUs now have the ability to run as a minimum
a few CUDA code, thanks to the business enterprise’s Boltzmann initiative, the
argument for the use of AMD may additionally have just gotten stronger than it
has ever been. historically AMD’s excessive-give up workstation and expert GPUs
are less expensive than Nvidia’s (and sure, $6000 is reasonably-priced on this
marketplace). We’ll see if those modifications yield fruit within the quarters
to come back.
No comments:
Post a Comment