“The datacenter space is still a work in progress, no doubt about it. But getting Sapphire Rapids out, which in certain workloads is in a leadership position. And so I think we are on a good path although really until we get to Sierra Forest and Granite Rapids, we won’t be all the way to where we want to be. But Pat was spending so much time in Washington. I think this area was probably an area, given that he was so focused on process technology that wasn’t getting as much attention. He has really doubled down on that since CHIPS Act was passed and I think we are in a really good place there.” Dave Zinsner - Chief Financial Officer Intel
Seems like further enforcement that AMD is in solid shape in server space. By the time INTC launches Granite Rapids to try to seek parity with Genoa, AMD would be moving on to their next release - possibly utilizing Xilinx tech among. other things
The expectation is granite rapids is more than Genoa, but we will need to wait and see. Genoa is on TSMC N5, while Granite Rapids is on the INTC version of TSMC N3. Intel is still stuck in the past where they were taking much too large of steps, and failing as a result. This risk looms large with granite rapids. The rumor is Granite Rapids is 128 cores, to the 96 in Genoa. The Sierra Forest product will exceed 256 cores. While Intel claims functional silicon booting several operating systems already, most don’t expect Granite Rapids to ship until sometime in 2024.
From all the talks I have been tracking from INTC executives, they are not even expecting a leadership position in servers at this point. In a few years from now, maybe (a big maybe). The headstart AMD has in terms of chiplets and infinity fabric is not something that just INTC can skip over. That modular architecture will now lend itself to incorporating the XLNX IP. So, as investors, AMD looks like a safe bet for the near term at least
Also in the Credit Suisse talk, you finally hear INTC singing the AMD tune – that customers want Total Cost of Ownership as the key metric. A stunning concession which flew under the radar
That is reasonably consistent with the current expectations for the Granite Rapids schedule…That said, the zen 5 servers should be out about the same time as Granite Rapids.
I heard this from Intel decades ago, so it is not really new… it is just an area AMD has been leading lately, so AMD pushes and INTC is generally silent. Intel mentioning it now either means they expect the gap to close, or perhaps they need to gain credibility with their customers. There may also be several application areas where Sapphire Rapids TCO bests Genoa, and Intel would like to highlight those.
Sure it is. Intel has used UPI for a while now (QPI before that) which are very similar to infinity fabric. Both are used to build cache coherent NUMA clusters using PCIe signaling as the physical transport layer. Intel does it at the circuit board level, while AMD integrated it into a single chip. Intel’s chiplet implementation in the now shipping Sapphire Rapids using a much higher bandwidth and lower latency connection to build a chiplet based system with uniform memory access.
CXL is here now with both INTC Sapphire Rapids, and AMD Genoa. CXL is what will be used for future connectivity in heterogenous compute systems, and for accelerators like those from XLNX.
I certainly agree with this. They have excellent server products and are growing revenue fairly quickly. Even after 5 years of great products, AMD only has around 20% of the server market. This market tends to have a lot of inertia, which is about at the point where it will shift from holding AMD back, to pushing them forward.
On a related note, AWS released their Milan instances about a year ago, but has not yet released Genoa instances. They recently announced they have released new Intel Ice Lake based instances (I have no idea why), but also Intel Sapphire Rapids based instances.
At this point any rumors or future promises about Intel’s next nodes need to be taken with a huge amount of salt. Intel has yet to hit any deadlines they have set. Further delays in Granit Rapids or Sierra Forest will just keep them one step behind AMD.
Intel is focused on catching up to an AMD that is focused on innovation and advancement. “Together We Advance” is more than just a slogan, it’s AMD’s promise to the market. I think it is going to be a tough for Intel for a few more years yet.
I don’t know… I see AMD releasing an FPGA, or multiples of them, in a Genoa-compatible socket. When you think about it for a few minutes, you realize that the right way to do this, which reduces time to market considerably, is to add one or more FPGA chiplets to a Genoa chip with one CPU chiplet and one I/O chiplet. I expect AMD to offer as many as three families: standard FPGA, lots of neurons, and probably a GPU. Huh? Why the last? When doing linear algebra (such as an FFT) on a GPU versus on CPU cores for small matrices, the cost of the traffic back and forth becomes significant. Not so much the bandwidth as the latency. Having everything on one socket should be a huge win. (Yes, in a two-socket system, there will be longer latency, but less than a nanosecond.)
I am not sure what you are disagreeing with here. Both Genoa and the Versal FPGA’s have a CXL interface. I am not sure what the value is of moving the FPGA to the Genoa socket. It seems likely at some point AMD will integrate some, if not all, of the FPGA functions into the Genoa socket, along with some Genoa X86 cores. I suspect this would be driven by a specific customer request for a particular feature set.
I totally agree with this. However, even if Intel introduced Intel_4 right now, it would be about 2 years late. While it is easy to pick on Intel right now, TSMC slipped their N3 over a year, and just slipped N2 almost 2 years.
Intel’s client execution has been pretty solid, with a new product release on an annual cadence. We will see if they can get GPU’s and server products back into a reliable release schedule.
Interesting. Released (in preview). Introducing Amazon EC2 R7iz instances Could be several months before GA of those in volume, for a host of reasons. Intel probably begged and arm-twisted to get this announcement.
Starting today, memory-optimized, high-frequency Amazon EC2 R7iz instances are available in preview. R7iz instances are the first EC2 instances powered by 4th generation Intel Xeon Scalable processors (code named Sapphire Rapids) with an all core turbo frequency up to 3.9 GHz. These instances have the highest performance per vCPU among x86-based EC2 instances, and they deliver up to 20% higher performance than z1d instances. The instances are built on the AWS Nitro System, a combination of dedicated hardware and lightweight hypervisor that delivers practically all of the compute and memory resources of the host hardware to your instances for better overall performance and security. R7iz instances are ideal for front-end Electronic Design Automation (EDA), relational database with high per-core licensing fees, financial, actuarial, data analytics simulations, and other workloads requiring a combination of high compute performance and high memory footprint.
For increased memory and scalability, R7iz instances are available in various sizes, including two bare metal sizes, with up to 128 vCPUs and up to 1,024GiB of memory. R7iz instances are the first x86-based EC2 instances to use DDR5 memory and deliver up to 2.4x higher memory bandwidth than comparable high-frequency instances. They also deliver up to 50 Gbps of networking speed and 40 Gbps of Amazon Elastic Block Store (EBS) bandwidth.
It is a little more than that. Intel agreed to move all their EDA to AWS, if amazon would deploy SR instances. With all the new foundry designs it makes sense for Intel to move to something that will be easy for their customers to access, as well as getting away from their proprietary solutions.