At SC22, QCT had an interesting demo on the show floor. It was showing off its upcoming 4th Generation Intel Xeon Scalable, codenamed “Sapphire Rapids” platform with a new technology. That technology is the Astera Labs CXL memory expansion card that STH is testing.
QCT Demos Astera Labs Leo CXL Memory Expansion on Intel Sapphire Rapids at SC22
As a quick note, we are not going into the QCT QuantaGrid D54Q-2U too much in this piece. We looked at them in both the QCT Liquid Cooled Rack Intel Sapphire Rapids Bake-Off and A Tour of the QCT Intel Sapphire Rapids System Overload Booth a few weeks ago, and the SC22 system is very similar to what we used in Hands-on Benchmarking with Intel Sapphire Rapids Xeon Accelerators. The new trick was really CXL.
Here is the QCT riser with the dual-slot Astera Labs Leo development board.
Here is a look at the board installed with DIMMs.
For those wondering, STH is working on a piece with these, although we are going to be using the new AMD EPYC Genoa unless Intel gives us the OK to show CXL in Sapphire Rapids early. Adding CXL devices like Leo to a system takes about three BIOS settings flipped to enable them.
At SC22, QCT was showing this setup and that the system was running at around 95% the performance of locally attached DDR5 in the 4th Gen Intel Xeon platform.
QCT was also showing that the Leo CXL card shows up as its own NUMA node without CPU cores attached, and has around the same latency of accessing memory on the second socket.
This was really cool to see working almost two months before Intel’s Sapphire Rapids launch at SC22.
Final Words
CXL in this generation is going to be more about development. It is still a huge deal for the industry. Seeing CXL working in both AMD EPYC 9004 “Genoa” and 4th Gen Intel Xeon platforms has been great. Our sense is that this is going to be a technology we will see become more available in Q2/Q3 2023 as more vendors ramp up their product lines. Still, this was a great QCT and Astera Labs demo at SC22.
@Patrick
Would the production product have more DIMM slots?
It seems rather low density on the DIMMs per unit volume metric.
@emerth,
Samsung has launched 512GB DIMMs for this which means 4 DIMMs is 2TB… The interesting thing though is that Genoa supports 12TB without CXL expansion… 16TB is possible with an 8 DIMM expander…
@emerth I suspect that it beats adding another CPU socket just because you need the RAM; but seeing an 8-pin power connector(up to 150w) rather than a 6 pin(75) or just bus power(75 for an x16) along with that heatsink suggests that either they’ve substantially overspecced everything for convenience in development or that you are looking at a pretty substantial power budget for 4 DIMMs as well.
@BaronMatrix, @fuzzyfuzzyfungus I concieved of CXL memory support as being rack mount cases full of RAM addressable over the fabric, not AICs with DIMM slots. You know, many many terabytes. If all the hype comes down to this sort of 25% increment over onboard DIMMs via PCIe5, then I am quite underwhelmed.