Phytium showed off working 64-core FT-2000 ARM server SoCs and 4-core FT-1500 laptops at Hot Chips. SemiAccurate didn’t think they would have silicon so fast but it was up, running, and they let you play with the systems.
First of all congratulations to Phytium, they have two of their own ARM cores, both 4-issue OoO designs running at up to 2GHz. If you think that is impressive then the 64-core version at 2GHz in a claimed 100W envelope will really floor you. From the look of the demo system heat sink, that claim is quite reasonable too, and all the systems we played with were just fine in responsiveness but we couldn’t run benchmarks at the show.
SemiAccurate regularly runs into people who downplay the capabilities of the homegrown Chinese CPU and SoC teams. We don’t share their pessimism and have trouble changing their point of view, partially because that market is so opaque to most western enthusiasts. With luck these two chips will start to reverse that trend, starting with the big one, the FT-2000.
Phytium’s FT-2000 and FT-1500/4 CPUs
It all starts with the FTC661 which is said to be more efficient and tuned for higher IPC than the consumer oriented FTC660 core in the FT-1500 SoC. It runs between 1.5-2.0GHz, has two PCIe3 16x interfaces, and sports 32MB of L2 cache along with 128MB of L3 cache. Then there are eight “proprietary extension interfaces” that go to eight external chips. The system looks like this.
Running Phytium FT-2000 server system
Those eight external chips are on the eight cards surrounding the main SoC and each house 1/8th of the L3 and two DDR3-1600 memory controllers plus memory. This results in 208.4GBps of memory bandwidth or 19.2GBps of effective usable memory bandwidth per external interface. When we first saw this architecture described last year at Hot Chips, we honestly didn’t think we would ever see it running. One year later we were impressed by running silicon and even more impressed by the fact that Phytium tried an architecture that no one else has really attempted. We can’t say it is a better or worse path yet, but we can say it is different.
A Phytium FT-1500 based 4C ARM laptop
What impressed us more is the Phytium FT-1500 line which comprises either 4 or 16 FTC660 cores. Each is a single SoC, no external controllers, running up to 2.0GHz for the 4C and 1.5GHz for the 16C version. For cache the 4C has 2MB of L2 and 8MB of LLC, the 16C has 8MB of L2 and 8MB of LLC. The smaller SoC has two DDR3-1600 controllers and the big one has four but both chips have two PCIe3 16x interfaces. Better yet the FT-1500/4 does all this in a 15W envelope while the FT-1500/16 only pulls 35W.
The part that really caught our attention was the running laptop pictured above, it wasn’t a prototype, it was a finished product running Ubuntu Kylin. If memory serves you can buy these in China now. The CPU was more than fast enough when we played with it, the desktop was responsive, and nothing stood out to us. Part of this was due to an unnamed external GPU but that only makes it more impressive, it just worked. If we can get our hands on one, we will try and see how it works as an everyday laptop, this isn’t a tablet with pretensions, it is a real computer.
So we left Hot Chips deeply impressed by what Phytium has accomplished. They have two homegrown ARM cores, three CPUs/SoCs from 4-64 cores, and they run. Not only that they are not delicate demo systems, they just work. OK the server was a bit early in development and was shipped over from China but it was open and running for three days of the show without any hiccups we saw or heard about. If this kind of showing doesn’t dissuade the critics about the capabilities of the Chinese silicon ecosystem, nothing will. In the meantime we will continue to be impressed by Phytium and others.S|A
Latest posts by Charlie Demerjian (see all)
- Intel shows off 10nm 112Gbps SerDes - Mar 12, 2019
- Intel releases Compute Express Link spec - Mar 11, 2019
- Qualcomm rolls out a second gen 5G modem called X55 - Feb 19, 2019
- What is Intel’s Foveros tech and what isn’t it? - Feb 11, 2019
- Why SemiAccurate called 10nm wrong - Jan 25, 2019