[閒聊] 一些關於 intel 和 AMD的傳聞 from reddit

作者: flylee (牛轉乾坤)   2020-07-15 12:01:18
https://bit.ly/3h1IFTm
Charlie Demerjian 在 Susquehanna 新爆的一些料,他以前爆些有關intel的料都還蠻準的
uzzi38整理如下
.Facebook like bfloat16 and Cooper Lake <lost phone line, reconnect 10 mins later>
.Alder Lake is 14nm GPU and 10nm CPU
.Several projects - including major ones, not PCHs etc were going to be TSMC.
No divulging of codenames for now.
.Jim Keller was the one pushing for projects at TSMC?
.Jim and Murphy were butting heads as of last year on technical decisions
.No chance at 7nm volume in '21. Maybe a launch for certain projects, but volume is '22.
Was planning for H1 21 not too long ago, but no chance.
.SPR taping out in more steps than normal. Intel are taping out regardless of any bugs
they know will exist, fixing and retaping immedietly. Phone line was choppy here,
but I believe something was said along the lines of this being a very expensive yet
faster way of getting the chip ready for production,
.Apple will definitely make a full transition to ARM in 2 years.
.No specific numbers for Milan over Rome today, but 20%+ should be a minimum expectation.
Milan is on track for later this year.
.Genoa and Turin are same platform. 10 channel DDR5, CXL, PCIe Gen 5, all the good stuff.
Was already known but a re-iteration for those that want a refresher.
.COVID is having an impact on AMD pickup in servers. Projects with new server hardware were
put on hold. No boots on the ground, nobody ready to break new systems in. This will blow
over by EoY, especially with Milan coming.
.ICL-SP availability Nov -> Mar
.No chance of GNR in '22. Expect it in '23.
.Apple moving to ARM is pushing MS to get better support with ARM solutions...
but so far progress is lacklustre. Will take a while for things to change.
.Xe loses competitiveness very fast. Starts of adequate, but won't remain that way.
.Nvidia: Lacking 7nm capacity hard because of shenanigans with TSMC. TSMC just did not care
about them trying to play games with them.
.Rome security features are quite something. Only 1-3% perf hit for real security.
.Best case is SPR for security as good as Rome
.When Milan releases Intel will not have a ST lead in server
xceryx的整理如下
1) Icelake paper launched this year due to security bugs but no OEM volume until March 2021
2) 2 of 3 fabs running test line 10nm. Still issues with 10nm process.
3) Tigerlake is supposedly a better 10nm+ process but yield is still low.
4) Cooperlake is bad vs Icelake but Facebook is the primary customer for Bfloat 16 AI workload
and 4P setup
5) 4C Tigerlake with big GPU area due to low yield
6) Alderlake 10nm GPU and 14nm CPU, due to low margin
7) Sapphire Rapid is taping it out as new way of taping/steps
8) Kellar pushed hard to outsource TSMC, and that's why he was gone.
9) Apple will transition to ARM completely but MacPro will probably be around for a long time,
similar to last time they moved to x86.
10) AMD has a lot of wafer allocations due to Huawei
11) AMD proven track record of roadmap
12) AMD server share was trending below what Charlie expected, COVID being an example of customer
expanding current capacity.
13) Sapphire Rapid is likely 2022H1 as Icelake is really deploying 2021H1.
14) Intel 7nm Granite Rapid is unlikely to show up in 2022, at least 2023 due to 13).
作者: BangSaint556 (Bang!!)   2020-07-15 13:06:00
推樓上
作者: c52chungyuny (PiPiDa)   2020-07-15 13:30:00
反正處理器極限看起來也就那樣了 intel要不要玩ARM搞不好有搞頭現在intel已經是印度人的天下了 搞Fab這種事還是東亞民族性比較適合
作者: leung3740250 (jenius921)   2020-07-15 15:34:00
zen3 ipc能追平icl就偷笑了,milan st怎麼比icl-sp強icl-sp實際上是snc和wlc的混合體然後拿spr和milan比是不是搞錯了什麼?按照目前spr多了amx的情況下spr很大機率是gdc,amx目前只有adl和spr才有然後你翻錯了一句,‘Best case is SPR for "security" as good as Rome’,你翻漏了關鍵詞security,正確的意思是「最好的情況下spr的『安全性』跟milan差不多」,而不是spr跟milan差不多adl用14nm不知道是哪個平行時空gdc那個規模是要怎麼用14nm做?8c rkl都比10900k還要肥了迫不得已還要砍L3
作者: appleonatree (蘋果在樹上)   2020-07-15 17:23:00
Nvidia前陣子不是一直放話GG太貴要下三星 新聞都有好幾篇他們放話轉下三星想要GG降價 GG哪鳥你那麼多
作者: leung3740250 (jenius921)   2020-07-15 17:53:00
icl-sp是10+的話那我更加想不到st會輸milan的理由頻率和ipc都不可能會輸你先看看tgl-u的頻率再說icl-sp跟tgl一樣是10+tgl怎麼和zen 2 apu比ipc差了30%以上規模都不是一個等級的zen2 apu砍了cache自然功耗會下降,然後ipc跟著一起降你漏電率擺在那啊10+能上4.8g就已經明確告訴你漏電率沒問題你說的問題是功耗限制而不是10+能不能上高頻的問題不過icp-sp能耗比肯定不會樂觀ipc多18%可不是電晶體多18%那麼簡單u系gpu沒辦法砍,然後據說超過150mm良率不能看4c+96eu 146mm,8c+32eu也是14x mm6c+96eu大概要160mm

Links booklink

Contact Us: admin [ a t ] ucptt.com