On a processor-to-processor comparison, Apple (NASDAQ:AAPL)'s silicon chips might already match or even surpass NVDA’s H-series and possibly the B-series class of GPUs, Lynx Equity Strategy analysts said in a new note.
An important observation from their checks indicates that Apple Silicon designed for data center servers might utilize the same packaging as that used for client devices, eliminating the need for liquid cooling.
Along with power consumption considerations, Apple's silicon could prove to be “more than competitive,” Lynx noted, highlighting four technical reasons to support this claim.
1) ‘Apple Silicon’s within-module memory – higher than NVDA’s:’ Lynx analysts highlight that Apple Silicon features superior within-module memory compared to NVDA's offerings, significantly impacting processor performance.
NVDA’s H100 SGX contains 80GB of unified memory, while its B100 model approximately doubles this capacity to 175GB per module, achieving 1.4TB per server with 8 GPUs. In contrast, Apple’s M2 Max offers 96GB, the M2 Ultra provides 192GB, and the M3 Max contains 128 GB. An anticipated M3 Ultra is expected to offer 256GB. Furthermore, speculative figures suggest the M4 Ultra could reach as high as 500GB, according to a recent Bloomberg report.
“There is no reason an Apple server would not place multiple M4 Ultras on a high-speed bus. Just four M4 Ultra gets you 2TB, higher than NVDA’s yet-to-be launched 8GPU B100 server.”
2) ‘Apple Silicon – server packaging is likely identical to client packaging:’ Analysts suggest that Apple’s server module packaging might mirror its client module packaging, eliminating the need for specialized designs for liquid cooling and the complex CoWoS packaging used by NVDA’s GPUs.
“We think Apple Silicon could avoid expensive 2.5D packaging such as TSM’s CoWoS process, and with it, avoid CoWoS’ high cost, low supply capacity and low yield. We think packaging of Apple Silicon could be a key differentiator vs. the GPU solution from NVDA and AMD (NASDAQ:AMD),” analysts noted.
Comparatively, the M4 Ultra’s compute power could match NVDA’s H100, which delivers 34 trillion FLOPS at 64-bit precision. Although Apple stated the M4/iPad achieves 38 trillion FLOPS, the precision level was not disclosed.
If the M4 client processor rivals the FLOPS of an NVDA H100 GPU, and the M4 Ultra surpasses the DRAM of NVDA’s B100 GPU, and given that Apple Silicon for client devices uses standard packaging, analysts believe the server version could utilize the same straightforward packaging. This standardization could provide a “substantial advantage in cost and in ease of supply,” Lynx highlighted.
3) ‘Apple Silicon – cuts out a major memory copy operation vs. GPU servers:’ In their note, analysts explain that within the traditional Windows/Intel architecture where NVDA operates, GPUs are treated as peripheral devices, a holdover from the WinTel era.
This configuration requires data transfer between discrete GPU modules to pass through the server's main memory, managed by the CPU, involving multiple steps: data is read from the GPU memory, passed across the PCIe bus, written onto the main memory, then read by the CPU to be sent to the receiving GPU.
In contrast, Apple's integrated approach with Apple Silicon, where the CPU module serves as the main processing unit, simplifies this process. Data transfers directly between CPU modules via the system bus, bypassing the need to write to a main memory, potentially reducing power consumption and decreasing system latency.
4) ‘Apple server racks - standard air-cooled CPU racks vs. water-cooled GPU racks:’ Should Apple opt to utilize its client processors in servers, which are air-cooled (with Apple MacBooks not requiring fans), Apple servers could forgo liquid cooling in favor of standard air cooling.
“And this could deliver a huge advantage in terms of capex and opex to Apple data centers,” Lynx analysts continued.
“As investors mull over the potential for Apple Silicon in client devices, quarterbacked by Apple Silicon in the cloud, for the delivery of Gen AI Services, we expect a slow run-up in the stock into the WWDC event next month. We reiterate our $220PT,” they concluded.