Maintaining graphics processors fed, whether or not for giant language fashions or the most recent AAA video games, takes a whole lot of reminiscence bandwidth. That is precisely what Samsung’s GDDR7 reminiscence module, unveiled Wednesday, claims to ship.
As we speak most GPUs and accelerators utilized in AI/ML work use speedy albeit costly high-bandwidth reminiscence that is carefully if circuitously coupled to the graphics processor silicon. Within the client and workstation markets, graphics double knowledge charge (GDDR) DRAM chips reign in all however the lowest price range playing cards.
With the introduction of 16-gigabit GDDR7 modules, Samsung is making some daring efficiency and effectivity claims. The manufacturing titan says the modules are able to delivering as much as 1.5 TB/sec of reminiscence bandwidth – about 1.4x that of 24-gigabit GDDR6, which topped out at round 1.1 TB/sec. The chaebol says it is also boosted its per-pin velocity to 32 Gb/sec.
Based on Samsung, these enhancements in efficiency effectivity are right down to a shift from non-return-to-zero (NRZ) signaling to pulse amplitude modulation – extra particularly PAM3. “PAM3 permits 50 p.c extra knowledge to be transmitted than NRZ throughout the identical signaling cycle,” the biz claimed.
On high of the efficiency enhancements, Samsung says it is also managed to drop energy consumption by about 20 p.c in comparison with GDDR6, and plans to supply a low voltage possibility for energy constrained merchandise – comparable to laptops and tablets.
All of this will likely sound spectacular – but when our math is correct, that 1.5TB/sec bandwidth declare depends on a reasonably fats 384-bit reminiscence interface. Sadly many chipmakers, together with Nvidia and AMD, have taken to shrinking the reminiscence bus over the previous few generations. Nvidia halved the reminiscence bus on its just lately announced 4060 TI GPU from 256 bits and 448GB/sec of bandwidth to 128 bits and 288GB/sec.
Assuming reminiscence buses on next-gen playing cards do not shrink additional, the supply of GDDR7 ought to assist to spice up bandwidth and enhance total efficiency. However not as a lot because it’s promising, except the buses begin getting greater once more.
Nevertheless, it’s going to be some time earlier than you’ll be able to count on to see Samsung’s GDDR7 modules within the wild. The Korean large says the chips will begin making their solution to key prospects later this yr. Due to this, we do not count on to see GDDR7-toting GPUs till not less than CES 2024 in January on the earliest – and certain solely in high-end playing cards.
We should always be aware that on the finish of June Micron teased [PDF] its upcoming GDDR7 RAM too – because of arrive within the first half of 2024.
The launch comes as Samsung slogs via a weak reminiscence market, which has battered the mega-corp’s revenue margins and compelled artistic inter-division cash infusions.
Earlier this month, Samsung warned that its second-quarter earnings would probably fall 96 percent yr over yr. And whereas modern chip households, like GDDR7, do command greater margins, the business watchers at Trendforce stay doubtful that GPU demand from AI will probably be sufficient to reinvigorate the business.
On the intense aspect, rumor has it Samsung has managed to work out the kinks in its 3nm and 4nm course of nodes to realize yields on par with or higher than rival TSMC. ®