Our plan to to design both. The LPDRR5 version cost will be definitely lower than the LPCAMM version. But the LPCAMM version can save the LPCAMM modules for future usage. We assume overall the LPCAMM version would be 20%~30% higher cost than the LPDDR5 version.
The SG2380 chip is not taped out yet. We havenāt got an official date from Sophgo when it will be taped out.
This is fantastic news!! Iām really glad to hear this
the C910 GhostWriter and the C908 halt bugs.
is the SG2380?
As stated by the research team who found the issues, only RISC-V cores specific to T-Head are affected due to their added hardware customizations:
https://ghostwriteattack.com/
It does not apply to the SG2380 from Sophon since A) the SoC uses cores from SiFive and B) SiFive has only implemented official RISC-V ISA (no SiFive cores has ever contained the custom T-Head extensions).
i understand.
thx
I am very excited to see on-board SFP slots, and if these are 25 GBit capable - this will cover alot of network applications. From specs I understand that 4x slot PCI-E can be used as these 4x SFP 25G slots. This is very smart and efficient decision. Also, I guess 4x25G slots can be aggregated to single 100G channel as usual for high-bandwidth applications (for applications with RDMA / clusters e.t.c).
I hope that 16x slot supports bifurcation, as we still would need 4x devices for NVMe.
Regarding LPCAMM availability and price : obvious solution is to have first boards with max memory (128Gb ECC) soldered in. You donāt need to upgrade memory if you have maximum, so soldered version is very much acceptable for max config.
Then later as LPCAMM availability improves and there are multiple suppliers to compete - lower cost boards can sell with 32-64Gb LPCAMM memory. But honestly, I would just go with soldered memory. 128Gb max configuration, and 32Gb cheap configuration. And thatās it.
Looking forward to see it working in my lab!
I am very excited to see on-board SFP slots, and if these are 25 GBit capable - this will cover alot of network applications.
Only two of them do 25 Gbit/s, the other do 10 Gbit/s
Just my opinion, but Iād prefer if the 10Gb ports were RJ45 since 10Gb SFP ports are becoming increasingly uncommon, and 10Gb RJ45 ports are quickly entering the consumer space, and already has been common in the server space.
You donāt need to upgrade memory if you have maximum, so soldered version is very much acceptable for max config.
Unless it fails.
128Gb max configuration, and 32Gb cheap configuration. And thatās it.
Both of these are a bit outside of my comfort zone to commit to right now, although that may change depending on how software support evolves. I definitely donāt want to spend large amounts of money on 32GB of RAM right out of the gate on an untested platform.
Only two of them do 25 Gbit/s, the other do 10 Gbit/s
Oh, interesting! How do you know that?
Iād prefer if the 10Gb ports were RJ45 since 10Gb SFP ports are becoming increasingly uncommo
I like the flexibility of SFP+, especially since 10GBASE-T is power-hungry compared to the power consumption of the machine. However, if the other two SFP ports are indeed SFP28, then I donāt see much point in having additional, non-aggregatable (with SFP28s, not with each other), slower SFP ports.
Oh, interesting! How do you know that?
The product brief lists 2x 25Gb ports and 2x 10Gb ports, so Iām guessing thatās where the SFP ports come from oasis-files/sg2380/SG2380 Product Brief V1.4.pdf at main Ā· milkv-oasis/oasis-files Ā· GitHub
That makes sense, I still wonder what those UDE ports are
Just my opinion, but Iād prefer if the 10Gb ports were RJ45 since 10Gb SFP ports are becoming increasingly uncommon
SFP+ 10GBase-T modules are very cheap and widely available. Those who need copper - can use copper module or direct attach. These who need fiber - can install fiber module. So it offers maximum flexibility. Also, SFP+ slot is much cheaper than 10GBase-T copper port, as SFP+ slot has no electronics (direct connect to CPU), while 10GBase-T requires controller, radiator e.t.c.
These who need fiber - can install fiber module.
Those who need fiber can just, use the other 2 faster SFP ports?
Also Iām not entirely sure where you got āvery cheapā from. My 10Gb NIC costs less than the price of a 10G SFP+ transceiver module, and it even has 2 ports on it!
EDIT: To elaborate on what I was saying in my previous post, even low grade consumer networks are beginning to move past 1Gb/s. Even a lot of cheaper SBCs have begun to include multi-gig RJ45 ports. I definitely donāt think itās a good idea to only have 1G RJ45s on the back.
I am slightly confused, the board with LPCAMM2 slots will be more expensive, even with no LPCAMM2 being sold with the board?
During the vote, it was said that the board itself will be sold at a lower price should one opt for buying LPCAMM2 modules themselves.
There did not seem to be any direct statement or impression given that their product pricing has changed? All they said was āWe assume overall the LPCAMM version would be 20%~30% higher cost than the LPDDR5 version.ā
The keyword here being āassumeā. Meaning if one were to currently price out and compare for example between the Oasis with 16GB soldered memory versus a combined total of the LPCAMM2-based board with a 16GB memory module (contingent on the memory controller they plan to use support stepping down to dual-channel mode). However that does not imply the module is from Milk-V or that they would be only selling LPCAMM2 Oasis bundled with memory. They are simply being straightforward, as they have been all along several times now, to give their own estimate of cost differences between LPDDR5 and LPCAMM2 at this point in time given that they do not control those external prices and Crucial is the only one with public quotes thus far on LPCAMM2.
As to their own products Milk-V clearly promised Oasis LPCAMM2 with no memory has a lower price tag if made. Makes sense when something important like a large memory chip is not there. So with that recent surprise announcement both board versions will be delivered, probably the starting line-up in order of price:
Oasis LPCAMM2 - with no memory
Oasis LPDDR5 4GB onboard
Oasis LPDDR5 8GB onboard
Oasis LPDDR5 16GB onboard
Anyone needing more than 16GB or prefers the flexibility of removable memory can easily buy the LPCAMM2 version. The only question is if the special coupon will be good for an Oasis 4GB or 8GB version.
Any way that line-up would reasonably cover all memory needs from low to high end. With Milk-V having been receptive to user requests and so accommodating by now offering the two different memory versions, there will be even more interest. Most likely then, inventory will sell out very quickly at launch. It may be good to consider putting together a long term strategy (and get Sophgo to commit as well) to reserve manufacturing resources for ongoing production of the board and SoC. Particularly for the Oasis LPCAMM2, put together a forecast to have weekly output at least for the first few months to satisfy demand to keep momentum going. Otherwise having to wait for availability between batches months at a time will cause many who are not die-hard RISC-V fans to quickly loose track and interest for the board from frustration.
Right, I missed the keyword āoverallā in their statement.
In that case 20~30% is not much.
I surely do not want anything less than 16GB!
Have there been any updates on the tapeout?
I asked on the Sophgo forum, in the SG2380(OASIS) category: