Expecting consistency and sense from ARM PR slides is sadly a fools errand.But this slide also directly contradicts the prior slide showing much smaller gains in GB6.
Damn those slides are hilariously bad. So incredibly misleading. You really have to take a good look at all of the slides to notice as well.
Most of the gains they're touting for CPU outside of the AI stuff is just clock gains - most of which won't be used in smartphones anyway (just like how nobody clocked X4 at 3.6GHz, clocking X925 at 3.8GHz is even less likely).
View attachment 99914
They're getting those clock gains by comparing X925 at it's best against a relatively gimped X4:
View attachment 99915
(X4 is capable of hitting 3.6GHz, but nobody shipped this because power is too high for smartphones. X925 peak power consumption looks to be worse, so 3.8GHz comparisons are also equally useless).
IPC uplift looks a touch under 15% at best if you go off this slide:
View attachment 99916
But this slide also directly contradicts the prior slide showing much smaller gains in GB6.
Oh and the comparison platform also just so happened to have twice the L3 and SLC available to it:
What. 80 TOPS in one core??The Cortex-X925 is also designed for and optimized for AI-based workloads, with dedicated AI accelerators and software optimizations that enhance AI processing efficiency. With up to 80 TOPS (trillion operations per second), the core can handle complex AI tasks, from natural language processing to computer vision
A5x and A5xx have typically seen very few updates.ARM's Cambridge team continues to drink tea.
They mostly do ISA work.ARM's Cambridge team continues to drink tea.
Oh, the neoverse V3 trick. Classic.Oh and the comparison platform also just so happened to have twice the L3 and SLC available to it
That's a tricky figure to reason out when 1 TOPS can mean INT4 or FP64 with wildly differing implications.What. 80 TOPS in one core??
This is concerning for the future of Windows-on-ARM (except for Qualcomm maybe, because they use their custom Oryon CPU).Cortex X925 seems to be generally underwhelming. It's seemingly not a clean-sheet design, but an evolution of X4. Also
Based on their numbers, X925's GB6 SC should be about ~3000.
Still incapable of of competing with the forerunners- Apple, Intel and AMD.
A lot of SIMD grunt is pretty much overkill in a purely mobile core.Even Apple M4 P has 4x128b
That doesn't seem possible: even for 4-bit FMA that'd mean 3.8G*2*6*32, which is less than 1.5 TOPS.That's a tricky figure to reason out when 1 TOPS can mean INT4 or FP64 with wildly differing implications.
At 3.8 Ghz w/ 6x 128b SIMD ALUs it doesn't sound impossible for lower precision math.
ARM typically incrementally upgrade their CPU every year, I wouldn't expect QC to be so prompt.This is concerning for the future of Windows-on-ARM (except for Qualcomm maybe, because they use their custom Oryon CPU).
Lol I messed up the math big time 🤣That doesn't seem possible: even for 4-bit FMA that'd mean 3.8G*2*6*32, which is less than 1.5 TOPS.
Yeah but this Cortex A925 will have to compete with Zen5, Lion Cove and M4-P.ARM typically incrementally upgrade their CPU every year, I wouldn't expect QC to be so prompt.
By the time QC field Pegasus or whatever Oryon2 is the ARM contender will not be in the same place.
They're getting those clock gains by comparing X925 at it's best against a relatively gimped X4:
View attachment 99915
(X4 is capable of hitting 3.6GHz, but nobody shipped this because power is too high for smartphones. X925 peak power consumption looks to be worse, so 3.8GHz comparisons are also equally useless).
So I'm thinking that the Client CSS stuff we're seeing now is basically what Qualcomm was talking about in their lawsuit a couple of years ago with the whole "you'll have to bundle Cortex with Mali" complaint. Going to be interesting to see what that ends up looking like in practice - I have a hard time believing that Nvidia and Samsung (and, for that matter, Renesas) will just be forced onto Mali for their lineups.
Arm is also, for the first time, packaging these cores as a GDSII to be provided to OEMs. This is essentially a file that can be provided to a fabricator like TSMC or Samsung for immediate production, and those GDSII files take into account any specific quirks or features of the fabricator. Arm says that a big perk of this is that it can improve time to market for companies that are using these cores, but that those companies can simply license the designs and do all of the leg work themselves if they wish, just like with previous Arm cores.
[SEALED] Memorandum Order granting-in-part and denying-in-part Defendants' motion to amend answer and counterclaims (DI 260, 272). Signed by Judge Laura D. Hatcher on 3/6/24.This order has been emailed to local counsel. (kjk) (Entered: 03/06/2024)
X925 must be having the widest SIMD in an ARM client core.
Even Apple M4 P has 4x128b
You should post updates about that court case more often, if possible.On that topic of litigation, from my recent check at Court Listener, Qualcomm added counter-claims against Arm that Arm did not actually destroy / quarantine all of the NUVIA IP that Arm had access to. It seems to be related to fabrics, as it mentions Arm's NCI and CMN, but also the MMU; Qualcomm says they found out about the claimed breaches during the discovery & deposition phase.
From this PDF and this PDF.
Hard to know the real details as so much is sealed & redacted, but Qualcomm's amendments were partly allowed (Qualcomm are the Defendants below) and QC seems likely to pursue this reverse-infringement-by-Arm claim:
Trial is still set for December 2024, so a long ways to go (for us, but I imagine Arm & QC are under pressure to either win handily or settle early).